THE 3% TRAP / PUBLIC EVIDENCE REGISTER

How I reviewed 221 AI visibility publications

A transparent account of what was screened, what was read, how evidence was classified and where the conclusions stop.

  • Structured narrative synthesis
  • Evidence freeze 2026-08-12
  • Model integration 2026-08-15
  • Schema 1.0

One review question, several evidence boundaries

How is AI changing how B2B SaaS buyers discover, evaluate, verify and shortlist vendors, and what should CMOs change in measurement, investment and pipeline planning as a result?

We screened 221 AI-visibility research publications, industry reports and original-data analyses.

221publications screened
98full text reviews
79abstract or sufficient public methods reviews
44landing page or public summary classifications

Count boundary. This is a publication count, not a count of independent datasets or confirmations. Fifty decision-relevant dataset components were selected for deeper synthesis, with reused and potentially overlapping evidence controlled separately.

  1. AI use
  2. answer exposure
  3. brand description
  4. mention or recommendation
  5. shortlist and verification
  6. website visit
  7. qualified pipeline
  8. revenue

Facts, synthesis and decisions are not interchangeable

Every major claim belongs to one of four labels. The label tells the reader whether the statement came directly from a source, from cross-source interpretation, from an operating recommendation or from an unresolved question.

Source fact

A result reported by an identified source, kept with its population, denominator and limitations.

Tomas synthesis

An interpretation across sources. It is not a quotation and does not create a new observed effect.

CMO decision

An operating implication for measurement, positioning, content, demand generation or pipeline planning.

Open hypothesis

A proposition that still needs direct buyer, market or company evidence before it can be treated as established.

Synthesis boundary. The sources do not estimate one common outcome. Traffic share, self-reported influence, citation rate, shortlist choice and conversion rate use different populations and denominators, so they were not pooled into one universal effect size.

How the 76% estimate is calculated

The model starts with G2's reported 69% vendor-change baseline and applies Semrush's relative high-value decision-influence gradient.

69% x 91.1% / 83.2% = 75.5% rounded to 76%

The calculation assumes that the relative increase in reported decision influence corresponds to a similar relative increase in reported vendor switching. The two studies used different populations and different questions.

Explicit limitation: The 76% estimate is modelled, not observed CRM attribution. It is not a pooled effect, measured incremental revenue or category demand created by AI.

What was controlled, and what remains uncertain

The 50 deeper dataset components were tracked separately from the 221 publication count. Possible reuse, common panels and disclosure gaps were not treated as independent confirmations.

Decision-grade components

23 of 50 deeper components met the decision-grade threshold.

Possible overlap

25 components were marked possibly overlapping, with 4 more marked unknown.

Transfer rule

Mechanisms may transfer; consumer multipliers do not.

Sensitivity views

B2B-only evidence; full-method evidence; non-vendor evidence; behavioural evidence excluding surveys; current-platform evidence; de-overlapped panels and corpora; desktop-only versus cross-device evidence; respondent-count versus purchase-value weighting where valid.

Correction log. Four source-register corrections are applied in this edition.

  • RDC-0007: S205 reviewDepth changed from Full to Landing-page (confirmed).
  • RDC-0008: S213 reviewDepth changed from Full to Landing-page (confirmed).
  • RDC-0009: S208 reviewDepth changed from Full to Full (confirmed).
  • APPENDIX-DELTA-S214: S214 decision changed from Hold to Include (reconciled).

All 221 screened sources

Search by source, publisher, evidence stream, result or limitation. With no filter applied, the server-rendered page contains the complete register.

Showing 221 of 221 sources

S001Decision: IncludeReview: FullEvidence: Adoption and evaluationB2B: Direct B2B

TrustRadius From Buzzword to Backbone

TrustRadius / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

348 professionals from the TrustRadius buyer community and social channels; online survey August 2025

Key result

45% of all buyers and 51% of enterprise buyers used AI in software buying; use reached 65% among C-suite respondents. Among users, 94% found it helpful or very helpful.

Limitation

Small, recruited sample with subgroup bases not fully visible on the public page; separate 2025 wave from B04.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
B2B buyer research
Evidence stream
Adoption and evaluation
B2B relevance
Direct B2B
Source type
Commercial marketplace
Source class
Baseline evidence register
Transparency
Medium
Dataset family
TRUSTRADIUS-2025
Provenance
AI Visibility Evidence Register v0.1
Origin
master
Corrections applied
None
S002Decision: IncludeReview: FullEvidence: adoption and market fragmentationB2B: Medium - platform mix informs monitoring priorities but population is all-market.

The 2026 Generative AI Landscape Report

Similarweb / 2026-07 / Industry / practice research

Open direct source
Sample or setting

Similarweb worldwide web and app panel; June 2025-May 2026; visits, unique visitors and app downloads are separate denominators.

Key result

Generative-AI platforms averaged 9.5B monthly web visits; ChatGPT category share fell from about 76% to 53% while Gemini and Claude grew.

Limitation

does not measure buying influence.

Why this decision

Large behavioral panel directly establishes adoption growth and multi-platform fragmentation; does not measure buying influence.

Source classification and provenance
Study design
Not stated
Evidence stream
adoption and market fragmentation
B2B relevance
Medium - platform mix informs monitoring priorities but population is all-market.
Source type
industry report landing page with downloadable full report
Source class
Industry / original-data report
Transparency
Moderate - direct report landing page and public methodology narrative; weighting and panel counts are proprietary.
Dataset family
SIMILARWEB_GENAI_LANDSCAPE_2026
Provenance
Existing evidence register; official report hub verification
Origin
master
Corrections applied
None
S003Decision: IncludeReview: Landing-pageEvidence: Adoption and reachB2B: Indirect / transfer

Alphabet Q2 2026 earnings disclosure on Gemini app users

Google / Alphabet / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

Gemini app monthly active users reported in Alphabet Q2 2026 earnings remarks

Key result

Alphabet reported 950M monthly active Gemini app users in Q2 2026.

Limitation

WAU and MAU are deliberately different denominators and may include different surfaces, account rules and geography. These figures prove scale, not discovery influence, and must never be used as a platform-share comparison.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Adoption, reach and platform use
Evidence stream
Adoption and reach
B2B relevance
Indirect / transfer
Source type
Independent / platform
Source class
Baseline evidence register
Transparency
Medium
Dataset family
A6
Provenance
AI Visibility Evidence Register v0.1
Origin
master
Corrections applied
None
S004Decision: IncludeReview: Landing-pageEvidence: Adoption and reachB2B: Indirect / transfer

OpenAI ChatGPT weekly active user disclosure

OpenAI / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

Product-level weekly active ChatGPT accounts

Key result

OpenAI reported more than 900M weekly active ChatGPT users in May 2026.

Limitation

WAU and MAU are deliberately different denominators and may include different surfaces, account rules and geography. These figures prove scale, not discovery influence, and must never be used as a platform-share comparison.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Adoption, reach and platform use
Evidence stream
Adoption and reach
B2B relevance
Indirect / transfer
Source type
Independent / platform
Source class
Baseline evidence register
Transparency
Medium
Dataset family
A6
Provenance
AI Visibility Evidence Register v0.1
Origin
master
Corrections applied
None
S005Decision: IncludeReview: FullEvidence: Adoption and reachB2B: Indirect / transfer

Semrush Traffic & Market channel-mix analysis

Semrush / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

More than 50,000 sites, 17 industries; worldwide mobile and desktop; January-December 2025; channel shares calculated as medians across tracked domains

Key result

AI-referral traffic rose 66.02%, from 462M to 767M monthly visits, but represented only 0.14% of total traffic. Organic search represented 16.04%. Google AI Mode was 0.01%

Limitation

AI means identifiable referrals, not exposure. Domain medians and aggregate volumes answer different questions. Computer software/development declined 26%, a warning against treating global growth as a SaaS benchmark. Likely related to Semrush's broader Traffic & Market data assets.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Adoption, reach and platform use
Evidence stream
Adoption and reach
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
A3
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official publisher search
Origin
master
Corrections applied
None
S006Decision: IncludeReview: FullEvidence: Adoption and reachB2B: Indirect / transfer

Semrush/Datos 17-month ChatGPT clickstream

Semrush/Datos / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

More than 1B rows from a US mobile-and-desktop clickstream; October 2024-February 2026

Key result

Outbound ChatGPT referrals rose 206% in 2025; 21.6% of referrals went to Google in February 2026; the ten largest recipient domains received more than 30%. Web search was enabled for 34.5% of prompts in February 2026, down from 46% in late 2024

Limitation

Actual panel-user base and weighting are not disclosed in the post. A referral is a next navigation, not necessarily incremental demand. Same Semrush/Datos panel family as other Semrush clickstream studies unless confirmed otherwise.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Adoption, reach and platform use
Evidence stream
Adoption and reach
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
DATOS-CLICKSTREAM
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official publisher page
Origin
master
Corrections applied
None
S007Decision: IncludeReview: FullEvidence: Adoption and reachB2B: Indirect / transfer

Semrush/Datos user-cohort analysis of Google sessions before and after ChatGPT adoption

Semrush/Datos / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

260B clickstream rows; US desktop; January 2024-June 2025. Treatment: users first observed using ChatGPT in Q1 2025; control: users never observed using it; Google sessions 90 days before/after, with a longer January 2024 cohort

Key result

No statistically significant fall in Google sessions after first ChatGPT use; the point pattern was slightly positive

Limitation

Observational adoption cohorts, desktop only, self-selection and seasonality. “No reduction in Google sessions” does not mean no individual query substitution. Same Datos/Semrush panel; do not count independently from A2/A4 when making a market-level conclusion.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Adoption, reach and platform use
Evidence stream
Adoption and reach
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
DATOS-CLICKSTREAM
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official publisher page
Origin
master
Corrections applied
None
S008Decision: IncludeReview: FullEvidence: Adoption and reachB2B: Indirect / transfer

Similarweb 2026 Generative AI Landscape; estimated web/app market intelligence panel

Similarweb / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

Worldwide; web visits and unique visitors plus separately measured app downloads; June 2025-May 2026

Key result

Generative-AI platforms averaged 9.5B monthly web visits, +70% YoY; 655M monthly unique visitors, +57%; ChatGPT's share of category web visits fell from about 76% to 53%, while Gemini rose from under 9% to about 27-28% and Claude from about 2% to nearly 9%

Limitation

Visits are not users, prompts, buying tasks or influence. Web and app measures are not interchangeable. Treat all Similarweb articles drawing from this edition as one rolling panel family.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Adoption, reach and platform use
Evidence stream
Adoption and reach
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
SIMILARWEB-MARKET
Provenance
AI Visibility Evidence Register v0.1
Origin
master
Corrections applied
None
S009Decision: IncludeReview: FullEvidence: Adoption and reachB2B: Indirect / transfer

SparkToro analysis of Datos clickstream

SparkToro / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

Multi-million-device panel; US desktop only; monthly and 10+-use thresholds; January 2023-June 2025

Key result

Nearly 40% used at least one named AI tool monthly and just over 20% did so 10+ times monthly; traditional search remained above 95% monthly and 86% at 10+ uses. Heavy AI use rose from about 3% to 21%

Limitation

Excludes mobile browsers/apps; tool-domain visits are an imperfect proxy for use; not B2B. Datos is owned by Semrush, so this is not independent of Semrush/Datos clickstream work.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Adoption, reach and platform use
Evidence stream
Adoption and reach
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
DATOS-CLICKSTREAM
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; primary analyst/data-partner page; JY Scauri review: retained source 25 of 34; same underlying Datos family as REV-060.
Origin
master
Corrections applied
None
S010Decision: IncludeReview: Landing-pageEvidence: Adoption/reachB2B: High

AI Is Much Bigger Than You Think

Graphite / 2026-03 / Industry / practice research

Open direct source
Sample or setting

Similarweb web and mobile-app usage panel combining search-engine sessions with search-related AI prompts.

Key result

Adding app usage raises estimated AI activity sharply; the analysis estimates total discovery activity grew rather than simply shifting from search.

Limitation

Important web-versus-app denominator correction; proprietary modeling means exact totals should not be pooled with clickstream-only studies.

Why this decision

Important web-versus-app denominator correction; proprietary modeling means exact totals should not be pooled with clickstream-only studies. JY source 2.

Source classification and provenance
Study design
Not stated
Evidence stream
Adoption/reach
B2B relevance
High
Source type
industry panel analysis
Source class
Prior-review mined
Transparency
Moderate
Dataset family
SIMILARWEB-GRAPHITE-DISCOVERY-2026
Provenance
JY Scauri review: retained source 2 of 34 from 77 screened.
Origin
master
Corrections applied
None
S011Decision: IncludeReview: AbstractEvidence: AI Overview activation and source overlapB2B: High

How Generative AI Disrupts Search: An Empirical Study of Google Search, Gemini, and AI Overviews

Riley Grossman et al.; Northeastern University and Yale / 2026-04 / Academic research

Open direct source
Sample or setting

11,500 representative real-user queries across Google organic, AI Overviews and Gemini Flash 2.5

Key result

AI Overviews appeared for 51.5% of the benchmark queries and source overlap across search surfaces averaged below 0.2 Jaccard, with lower stability under reruns and minor query edits.

Limitation

Large public benchmark directly showing that organic rankings do not map cleanly to generative visibility.

Why this decision

Large public benchmark directly showing that organic rankings do not map cleanly to generative visibility.

Source classification and provenance
Study design
Not stated
Evidence stream
AI Overview activation and source overlap
B2B relevance
High
Source type
Accepted conference preprint
Source class
Academic
Transparency
High
Dataset family
Google-Gemini-AIO benchmark
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S012Decision: IncludeReview: FullEvidence: AI Overview and organic volatilityB2B: High for monitoring and expectations.

AI Overviews and SERP Volatility Research

Authoritas and research partners / 2025 / Industry / practice research

Open direct source
Sample or setting

11,203 categorized US desktop keywords captured August 23 2024, October 17 2024 and January 17 2025; 2,104 had AIOs; shared spreadsheets/screens.

Key result

AI Overview source-page volatility averaged 0.68-0.73 over 8-13 weeks, higher than organic rankings; about 70% of AIO pages could change.

Limitation

sparse time points and a selected keyword set limit platform-wide inference.

Why this decision

Transparent repeated-wave evidence; sparse time points and a selected keyword set limit platform-wide inference.

Source classification and provenance
Study design
Not stated
Evidence stream
AI Overview and organic volatility
B2B relevance
High for monitoring and expectations.
Source type
full longitudinal SERP study with shared data
Source class
Industry / original-data report
Transparency
High - full dates, n, metrics, limitations and shared data.
Dataset family
AUTHORITAS_AIO_VOLATILITY_2024_2025
Provenance
Official Authoritas research index
Origin
master
Corrections applied
None
S013Decision: IncludeReview: FullEvidence: AI Overview prevalence and CTR lossB2B: High for European search risk; not B2B-specific.

AI Overviews in Germany: How Much Click-Through Rates Are Really Dropping

SISTRIX / 2026-02-25 / Industry / practice research

Open direct source
Sample or setting

More than 100M keywords in Germany; SISTRIX CTR/traffic models; AIO prevalence, position and category loss analyzed.

Key result

About 20% of keywords showed AIOs; position-one CTR fell from 27% to 11% and estimated aggregate organic click loss was 6.6%.

Limitation

modeled CTR/lost clicks and proprietary clickstream assumptions.

Why this decision

Large non-US market analysis with category heterogeneity; modeled CTR/lost clicks and proprietary clickstream assumptions.

Source classification and provenance
Study design
Not stated
Evidence stream
AI Overview prevalence and CTR loss
B2B relevance
High for European search risk; not B2B-specific.
Source type
full vendor market-wide CTR study
Source class
Industry / original-data report
Transparency
Moderate-high - full article and very large base; click model/raw dataset unavailable.
Dataset family
SISTRIX_GERMANY_AIO_2026
Provenance
Official SISTRIX search
Origin
master
Corrections applied
None
S014Decision: IncludeReview: FullEvidence: AI Overview prevalence, intent and zero-clickB2B: High for search planning and lower-funnel expansion.

Semrush Report: AI Overviews' Impact on Search in 2025

Semrush / 2026-01 / Industry / practice research

Open direct source
Sample or setting

10M+ keywords January-November 2025; 200K+ keywords for zero-click trends; 11K domains for industry visibility.

Key result

AIO prevalence rose from 6.49% in January to 24.61% in July then 15.69% in November, with more commercial/transactional coverage over time.

Limitation

shares Semrush AIO infrastructure with IND-044 and is not an independent replication.

Why this decision

Large longitudinal update; shares Semrush AIO infrastructure with IND-044 and is not an independent replication.

Source classification and provenance
Study design
Not stated
Evidence stream
AI Overview prevalence, intent and zero-click
B2B relevance
High for search planning and lower-funnel expansion.
Source type
full vendor longitudinal SERP study
Source class
Industry / original-data report
Transparency
High-moderate - full article publishes multiple bases and dates; raw domains/keywords unavailable.
Dataset family
SEMRUSH_AIO_SERP_PANEL_2024_2025
Provenance
Official Semrush search; rolling-family deduplication
Origin
master
Corrections applied
None
S015Decision: IncludeReview: FullEvidence: AI Overview triggers and intentB2B: High for query strategy; no buying outcome.

AI Overview Study: How User Intent Drives AIO Appearance Rates

Authoritas and Dave Cousin / 2025-01-27 / Industry / practice research

Open direct source
Sample or setting

10,000 categorized US desktop keywords across seven industries in December 2024; keyword and AIO exports shared through Google Drive.

Key result

AIOs appeared for 29.9% of keywords but 11.5% of search volume; problem-solving and specific-question intents triggered them most.

Limitation

one month/desktop/US and hand-built keyword categories.

Why this decision

Transparent intent-based sample with shared data; one month/desktop/US and hand-built keyword categories.

Source classification and provenance
Study design
Not stated
Evidence stream
AI Overview triggers and intent
B2B relevance
High for query strategy; no buying outcome.
Source type
full vendor/analyst SERP study with shared data
Source class
Industry / original-data report
Transparency
High - full article, n, date, geography and shared raw exports.
Dataset family
AUTHORITAS_AIO_INTENT_DEC2024
Provenance
Official Authoritas research index
Origin
master
Corrections applied
None
S016Decision: IncludeReview: FullEvidence: AI Overview triggers and query intentB2B: Medium-high - query-intent mechanics, not buyer outcomes.

We Studied 200,000 AI Overviews: Here's What We Learned

Semrush / 2025-07-22 / Industry / practice research

Open direct source
Sample or setting

200,000 keywords randomly selected from the Semrush US database, restricted to keywords triggering an AIO September 1-10, 2024; desktop/mobile cuts.

Key result

About 80% of desktop AIOs were informational and 82% occurred on keywords below 1,000 monthly searches.

Limitation

conditioned on ever-triggering AIOs and now temporally old.

Why this decision

Large trigger/composition study; conditioned on ever-triggering AIOs and now temporally old.

Source classification and provenance
Study design
Not stated
Evidence stream
AI Overview triggers and query intent
B2B relevance
Medium-high - query-intent mechanics, not buyer outcomes.
Source type
full vendor SERP study
Source class
Industry / original-data report
Transparency
High-moderate - full article describes sampling; keyword list/raw output unavailable.
Dataset family
SEMRUSH_AIO_SERP_PANEL_2024_2025
Provenance
Official Semrush search; predecessor-wave lineage
Origin
master
Corrections applied
None
S017Decision: IncludeReview: FullEvidence: AI referrals and conversionsB2B: High for enterprise channel allocation.

AI Search Visits Surging in 2025 - But Organic Search Remains the Cornerstone

BrightEdge / 2025 / Industry / practice research

Open direct source
Sample or setting

Thousands of queries and top-performing websites including Fortune 100 brands; January-August 2025; North America/Europe context in BrightEdge AI Catalyst.

Key result

AI search accounted for less than 1% of referrals while organic search drove most conversions; AI referral growth varied sharply by platform.

Limitation

site/query bases and sampling are described only broadly.

Why this decision

Important contradictory conversion result; site/query bases and sampling are described only broadly.

Source classification and provenance
Study design
Not stated
Evidence stream
AI referrals and conversions
B2B relevance
High for enterprise channel allocation.
Source type
research report landing page with full PDF download
Source class
Industry / original-data report
Transparency
Moderate-low - direct report landing/PDF but vague counts and proprietary selection.
Dataset family
BRIGHTEDGE_AI_CATALYST_2025
Provenance
Official BrightEdge research-report index
Origin
master
Corrections applied
None
S018Decision: IncludeReview: FullEvidence: AI-plus-search buying behaviorB2B: High as behavioral bridge; B2C population.

AI Tools and the Modern Consumer Buyer Journey Study

Semrush / 2026 / Industry / practice research

Open direct source
Sample or setting

US consumer market-research panel fielded December 2025; public article provides workflow percentages but requires report for full sample details.

Key result

77% of AI-using consumers also used traditional search; 41% started with AI then validated through search and 35% began with search then used AI for synthesis.

Limitation

consumer self-report and incomplete public base limit transfer.

Why this decision

Directly supports complementarity and validation; consumer self-report and incomplete public base limit transfer.

Source classification and provenance
Study design
Not stated
Evidence stream
AI-plus-search buying behavior
B2B relevance
High as behavioral bridge; B2C population.
Source type
full consumer survey research article
Source class
Industry / original-data report
Transparency
Moderate - full article with fielding context; exact sample and questionnaire should be pulled from report.
Dataset family
SEMRUSH_CONSUMER_BUYER_JOURNEY_2025
Provenance
Existing evidence register; official Semrush page; JY Scauri review: retained source 8 of 34; separately surfaced in project correspondence about Semrush data.
Origin
master
Corrections applied
None
S019Decision: IncludeReview: FullEvidence: anonymous selection and seller contactB2B: Very high - direct B2B buying mechanics.

2024 B2B Buyer Experience Report

6sense Research / 2024 / Industry / practice research

Open direct source
Sample or setting

Survey of 2,509 recent B2B buyers; annual predecessor wave to 2025 Buyer Experience research.

Key result

69% of the purchase process occurred before seller engagement; 81% had selected a preferred vendor before speaking with sales.

Limitation

annual wave shares the 6sense program and is not AI-specific.

Why this decision

Pre-AI-wave commercial baseline for shortlist formation; annual wave shares the 6sense program and is not AI-specific.

Source classification and provenance
Study design
Not stated
Evidence stream
anonymous selection and seller contact
B2B relevance
Very high - direct B2B buying mechanics.
Source type
full annual B2B buyer report
Source class
Industry / original-data report
Transparency
High-moderate - full report page and sample; full instrument details are limited.
Dataset family
6SENSE_BUYER_EXPERIENCE_2024
Provenance
Official 6sense search; predecessor lineage review
Origin
master
Corrections applied
None
S020Decision: IncludeReview: FullEvidence: B2B AI-search adoption and meaningful interactionsB2B: Very high - direct business buyers.

B2B Buyers Make Zero-Click Number One

Forrester / 2026-01-22 / Industry / practice research

Open direct source
Sample or setting

Forrester Buyers' Journey Survey 2025; public article reports adoption/change but not sample or full instrument.

Key result

94% of buyers reported AI use, and twice as many named generative/conversational search a meaningful source as any other source.

Limitation

exact denominator, question wording and full methodology are behind the client report.

Why this decision

Direct decision-relevant B2B wave; exact denominator, question wording and full methodology are behind the client report.

Source classification and provenance
Study design
Not stated
Evidence stream
B2B AI-search adoption and meaningful interactions
B2B relevance
Very high - direct business buyers.
Source type
official analyst article reporting survey findings
Source class
Industry / original-data report
Transparency
Low-moderate - official analyst article only; full report/sample pending.
Dataset family
FORRESTER_BUYERS_JOURNEY_2025
Provenance
Official Forrester search
Origin
master
Corrections applied
None
S021Decision: IncludeReview: FullEvidence: B2B buyer AI use and complex-purchase outcomesB2B: Very high - direct complex B2B buying.

The Future of B2B Buying Will Come Slowly and Then All at Once

Forrester / 2024-11-21 / Industry / practice research

Open direct source
Sample or setting

Forrester Buyers' Journey Survey 2024; public article reports percentages but not sample size or full survey methods.

Key result

89% of business buyers reported GenAI use in at least one purchase-process area; 87% of users said it improved business outcomes, with the largest effect reported for purchases over $1M.

Limitation

full client report should be obtained before strong quantification.

Why this decision

Highly relevant complexity/enterprise claim but public method is insufficient; full client report should be obtained before strong quantification.

Source classification and provenance
Study design
Not stated
Evidence stream
B2B buyer AI use and complex-purchase outcomes
B2B relevance
Very high - direct complex B2B buying.
Source type
official analyst article reporting survey findings
Source class
Industry / original-data report
Transparency
Low-moderate - official analyst article, but n, recruitment, instrument and item bases are absent.
Dataset family
FORRESTER_BUYERS_JOURNEY_2024
Provenance
AI Visibility Evidence Register v0.1; Official Forrester search
Origin
master
Corrections applied
None
S022Decision: IncludeReview: AbstractEvidence: B2B provider recommendation biasB2B: High

The Invisible Hand: Unveiling Provider Bias in Large Language Models for Code Generation

Xiaoyu Zhang et al.; Xi'an Jiaotong University and collaborators / 2025-07 / Academic research

Open direct source
Sample or setting

17,014 prompts across six coding-task categories and 30 scenarios; seven LLMs and roughly 500 million tokens

Key result

Models showed systematic provider preferences, especially for Google and Amazon services, and sometimes rewrote code to include preferred vendors without being asked.

Limitation

Exceptionally relevant B2B SaaS/cloud recommendation evidence, though code generation is not identical to explicit vendor evaluation.

Why this decision

Exceptionally relevant B2B SaaS/cloud recommendation evidence, though code generation is not identical to explicit vendor evaluation.

Source classification and provenance
Study design
Not stated
Evidence stream
B2B provider recommendation bias
B2B relevance
High
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
Cloud-provider bias benchmark
Provenance
ACL Anthology search for provider recommendation bias
Origin
master
Corrections applied
None
S023Decision: IncludeReview: FullEvidence: brand descriptions, unsolicited comparison and recommendationB2B: High - brand framing can affect B2B evaluation.

The Parrot Problem: Why AI Search Has a Second Dimension Marketers Can't Ignore

Profound / 2026-06-25 / Industry / practice research

Open direct source
Sample or setting

50,000 prompts across seven industries and multiple answer engines; unsolicited assertions/comparisons classified in responses.

Key result

Nearly half of AI responses included unsolicited comparisons, opinions or recommendations beyond the user's literal request.

Limitation

proprietary prompt mix and classifier validation limit certainty.

Why this decision

Supports measuring description/sentiment, not just mention; proprietary prompt mix and classifier validation limit certainty.

Source classification and provenance
Study design
Not stated
Evidence stream
brand descriptions, unsolicited comparison and recommendation
B2B relevance
High - brand framing can affect B2B evaluation.
Source type
full vendor prompt-analysis study
Source class
Industry / original-data report
Transparency
Moderate - full article and base; prompt selection and classification validation are incompletely public.
Dataset family
PROFOUND_PARROT_50K_2026
Provenance
Official Profound research hub and article inspection
Origin
master
Corrections applied
None
S024Decision: IncludeReview: FullEvidence: brand familiarity, research and validationB2B: Very high - direct software/hardware purchase sample.

2024 B2B Buying Disconnect Report: The Year of the Brand Crisis

TrustRadius / 2024-06-10 / Industry / practice research

Open direct source
Sample or setting

Online survey in March-April 2024; 2,164 verified technology buyers and 243 technology vendors from the global TrustRadius network.

Key result

78% of shortlisting buyers selected products known before research, rising to 86% for enterprise; 56% spoke with a product user before purchase.

Limitation

vendor-owned network sample and not focused on AI.

Why this decision

Strong pre-AI comparison for shortlist/brand effects; vendor-owned network sample and not focused on AI.

Source classification and provenance
Study design
Not stated
Evidence stream
brand familiarity, research and validation
B2B relevance
Very high - direct software/hardware purchase sample.
Source type
full annual B2B technology buyer report
Source class
Industry / original-data report
Transparency
High - full methodology, dates, bases and invitation source.
Dataset family
TRUSTRADIUS_B2B_DISCONNECT_2024
Provenance
Official TrustRadius research search
Origin
master
Corrections applied
None
S025Decision: IncludeReview: FullEvidence: brand mentions, competitive visibility and momentumB2B: Medium-high as measurement framework; no B2B SaaS sector.

The 2026 Generative AI Brand Visibility Index

Similarweb / 2026-03 / Industry / practice research

Open direct source
Sample or setting

More than 25,000 US prompts across ChatGPT, Gemini, Copilot and Perplexity in January 2026; 113 brands in six consumer-facing sectors; April 2025 baseline.

Key result

Visibility was concentrated and specialists sometimes overperformed larger brands; referrals plateaued while mention-based visibility diverged.

Limitation

consumer sectors and prompt-selection model limit B2B transfer.

Why this decision

Useful cross-platform category benchmark and bridge from traffic to mentions; consumer sectors and prompt-selection model limit B2B transfer.

Source classification and provenance
Study design
Not stated
Evidence stream
brand mentions, competitive visibility and momentum
B2B relevance
Medium-high as measurement framework; no B2B SaaS sector.
Source type
official report landing page with downloadable report
Source class
Industry / original-data report
Transparency
Moderate - direct report landing and public prompt/sector counts; full prompt weighting and scoring proprietary.
Dataset family
SIMILARWEB_AI_BRAND_VISIBILITY_INDEX_2026
Provenance
Official Similarweb report hub and press release; JY Scauri review: retained source 12 of 34; overlaps existing project Similarweb source family.
Origin
master
Corrections applied
None
S026Decision: IncludeReview: AbstractEvidence: Brand recognition versus discoveryB2B: High

The Discovery Gap: How Product Hunt Startups Vanish in LLM Organic Discovery Queries

Amit Prakash Sharma / 2026-01 / Academic research

Open direct source
Sample or setting

112 Product Hunt startups tested across recognition and category-discovery prompts on two models

Key result

Models often knew individual startups when named but failed to surface them in unprompted category discovery, revealing a large recognition-to-shortlist gap.

Limitation

Highly relevant to startup brand visibility, but only two models and a selected startup cohort were tested.

Why this decision

Highly relevant to startup brand visibility, but only two models and a selected startup cohort were tested.

Source classification and provenance
Study design
Not stated
Evidence stream
Brand recognition versus discovery
B2B relevance
High
Source type
Preprint
Source class
Academic
Transparency
Medium
Dataset family
Product Hunt discovery gap
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S027Decision: IncludeReview: AbstractEvidence: Brand recommendation biasB2B: High

Global Is Good, Local Is Bad? Understanding Brand Bias in LLMs

Mahammed Kamruzzaman, Hieu Minh Nguyen, Gene Louis Kim; University of South Florida / 2024-11 / Academic research

Open direct source
Sample or setting

Curated prompts across four brand categories tested on multiple LLMs

Key result

Models associated global brands more positively than local brands and produced country-of-origin and income-linked recommendation effects.

Limitation

Direct brand-visibility and recommendation evidence, although it does not observe real buyers or live web retrieval.

Why this decision

Direct brand-visibility and recommendation evidence, although it does not observe real buyers or live web retrieval.

Source classification and provenance
Study design
Not stated
Evidence stream
Brand recommendation bias
B2B relevance
High
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
LLM brand-bias benchmark
Provenance
ACL Anthology search for LLM brand recommendations
Origin
master
Corrections applied
None
S028Decision: IncludeReview: FullEvidence: ChatGPT citation sourcesB2B: High for source strategy; indirect for decisions.

How ChatGPT Sources the Web

Profound / 2026-02-03 / Industry / practice research

Open direct source
Sample or setting

700,000 ChatGPT conversations with web citations from Q4 2025; conversation-turn and topic/source patterns analyzed.

Key result

Most citations appeared in turn one, Wikipedia functioned as a default knowledge layer and source clusters varied by topic.

Limitation

only citation-bearing conversations and no behavior/outcome link.

Why this decision

Large real-conversation citation analysis; only citation-bearing conversations and no behavior/outcome link.

Source classification and provenance
Study design
Not stated
Evidence stream
ChatGPT citation sources
B2B relevance
High for source strategy; indirect for decisions.
Source type
full vendor citation study
Source class
Industry / original-data report
Transparency
Moderate-high - full article and base; prompt sampling/classification detail proprietary.
Dataset family
PROFOUND_CHATGPT_CITATIONS_Q4_2025
Provenance
Official Profound research hub and direct page; Cyrus Shepard 54-source evidence sheet; Profound citation chain.
Origin
master
Corrections applied
None
S029Decision: IncludeReview: AbstractEvidence: Citation absorptionB2B: High

From Citation Selection to Citation Absorption: A Measurement Framework for Generative Engine Optimization Across AI Search Platforms

Kai Zhang, Xinyue He, Jingang Yao / 2026-04 / Academic research

Open direct source
Sample or setting

21,143 citations described with 72 observable features across AI search platforms

Key result

Citation presence and actual absorption into the generated answer diverged, and a multi-feature score could predict contribution, but the observational design cannot establish causal source use.

Limitation

Important distinction between decorative citation and substantive influence, with no causal validation.

Why this decision

Important distinction between decorative citation and substantive influence, with no causal validation.

Source classification and provenance
Study design
Not stated
Evidence stream
Citation absorption
B2B relevance
High
Source type
Preprint
Source class
Academic
Transparency
Medium
Dataset family
Citation absorption cross-platform audit
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S030Decision: IncludeReview: AbstractEvidence: Citation feature optimizationB2B: High

Think Before Writing: Feature-Level Multi-Objective Optimization for Generative Citation Visibility

Zikang Liu, Peilan Xu / 2026-07 / Academic research

Open direct source
Sample or setting

Feature-level optimization experiments spanning informational and lexical content features

Key result

Informational features were more useful than lexical tricks for citation visibility, although the approach was computationally costly and relied substantially on model judges.

Limitation

Directly informs what content attributes may matter, with explicit evaluation caveats.

Why this decision

Directly informs what content attributes may matter, with explicit evaluation caveats.

Source classification and provenance
Study design
Not stated
Evidence stream
Citation feature optimization
B2B relevance
High
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
Feature-level citation optimization
Provenance
Martinez 2026 core matrix; ACL Anthology; Olivier Martinez 45-study literature matrix; overlap with Cyrus Shepard 54-source sheet.
Origin
master
Corrections applied
None
S031Decision: IncludeReview: AbstractEvidence: Citation fidelityB2B: Medium

Evaluating Verifiability in Generative Search Engines

Nelson F. Liu, Tianyi Zhang, Percy Liang; Stanford / 2023-12 / Academic research

Open direct source
Sample or setting

Human audit of Bing Chat, NeevaAI, Perplexity and YouChat across diverse query sets

Key result

Only 51.5% of generated sentences were fully supported and 74.5% of citations supported their associated proposition, showing that visible citations are not equivalent to reliable evidence.

Limitation

Audits early commercial engines and citation correctness, but does not measure source selection causally or any downstream click, choice or revenue outcome.

Why this decision

Direct audit of commercial generative-search citations and essential constraint on citation-as-visibility metrics.

Source classification and provenance
Study design
Not stated
Evidence stream
Citation fidelity
B2B relevance
Medium
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
Commercial GSE verifiability audit
Provenance
Martinez 2026 core matrix; ACL Anthology; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S032Decision: IncludeReview: FullEvidence: citation fragmentationB2B: High - multi-platform planning.

Answer Engine Citation Overlap Strategy

Profound / 2025-07-01 / Industry / practice research

Open direct source
Sample or setting

100,000 distinct prompts run across ChatGPT and Perplexity; cited sources compared between engines.

Key result

The two engines cited substantially different parts of the web, supporting platform-specific source strategies.

Limitation

only two platforms and proprietary prompt selection.

Why this decision

Clear large cross-engine comparison; only two platforms and proprietary prompt selection.

Source classification and provenance
Study design
Not stated
Evidence stream
citation fragmentation
B2B relevance
High - multi-platform planning.
Source type
full vendor cross-platform study
Source class
Industry / original-data report
Transparency
Moderate-high - full article and prompt count; demand weighting/raw citation matrix unavailable.
Dataset family
PROFOUND_CITATION_OVERLAP_100K_2025
Provenance
Official Profound research hub and direct page
Origin
master
Corrections applied
None
S033Decision: IncludeReview: AbstractEvidence: Citation source characteristicsB2B: High

When Content Is Goliath and Algorithm Is David: The Style and Semantic Effects of Generative Search Engine

Lijia Ma, Juan Qin, Xingchen Xu, Yong Tan; UNC Charlotte, USTC, University of Washington / 2024-03 / Industry / practice research

Open direct source
Sample or setting

Approximately 10,000 websites collected from Google's generative and conventional search surfaces

Key result

Generative search cited pages with greater semantic similarity and higher language-model predictability than conventional search, indicating a distinct content-selection mechanism.

Limitation

Early large cross-surface empirical audit, but tied to a changing Google product snapshot.

Why this decision

Early large cross-surface empirical audit, but tied to a changing Google product snapshot.

Source classification and provenance
Study design
Not stated
Evidence stream
Citation source characteristics
B2B relevance
High
Source type
Working paper
Source class
Academic
Transparency
Medium
Dataset family
Google generative-versus-conventional source audit
Provenance
SSRN search for generative search citations
Origin
master
Corrections applied
None
S034Decision: IncludeReview: FullEvidence: citation source mix by industry and modelB2B: Very high - direct SaaS/software citation benchmark.

Where Do AI Citations Come From?

Profound / 2026-07-30 / Industry / practice research

Open direct source
Sample or setting

11.84B citations across eight models, 8,061 active categories April 16-July 16, 2026; 3.02M-domain classification map covering 98.3% of volume.

Key result

About 57% of citations were brand-owned sites overall; SaaS/software had only 11.4% earned-media share while model mixes varied widely.

Limitation

vendor-customer category selection and observational scope remain.

Why this decision

Exceptionally large, well-described current citation dataset with a SaaS cut; vendor-customer category selection and observational scope remain.

Source classification and provenance
Study design
Not stated
Evidence stream
citation source mix by industry and model
B2B relevance
Very high - direct SaaS/software citation benchmark.
Source type
full vendor observational research article
Source class
Industry / original-data report
Transparency
High-moderate - full methodology, thresholds and classification; prompt categories/customer selection are proprietary.
Dataset family
PROFOUND_CITATIONS_API_2026
Provenance
Official Profound research hub and full-page inspection
Origin
master
Corrections applied
None
S035Decision: IncludeReview: FullEvidence: citation sources and recommendation formatsB2B: High - B2B shortlists frequently use best/vendor-comparison prompts.

Do Self-Promotional Best Lists Boost ChatGPT Visibility?

Ahrefs / 2025-12-04 / Industry / practice research

Open direct source
Sample or setting

26,283 source URLs for top-of-funnel ChatGPT queries; includes 1,100 blog lists and 3,000 cited lists analyzed by authority/category.

Key result

Recently updated best-of lists were prominent ChatGPT sources and third-party list placement correlated with brand inclusion.

Limitation

correlational and focused on selected commercial prompts.

Why this decision

Large format/source analysis relevant to comparison content; correlational and focused on selected commercial prompts.

Source classification and provenance
Study design
Not stated
Evidence stream
citation sources and recommendation formats
B2B relevance
High - B2B shortlists frequently use best/vendor-comparison prompts.
Source type
full vendor source-URL study
Source class
Industry / original-data report
Transparency
Moderate-high - full article with URL counts; query universe and scoring details partly proprietary.
Dataset family
AHREFS_BRAND_RADAR_BEST_LISTS_2025
Provenance
Official Ahrefs data-studies search
Origin
master
Corrections applied
None
S036Decision: IncludeReview: FullEvidence: citation volatility and platform fragmentationB2B: High for measurement cadence and multi-platform scope.

AI Citation Drift: How Stable Are Sources in AI Search Results?

SISTRIX / 2026-05-01 / Industry / practice research

Open direct source
Sample or setting

82,619 qualified prompts, 1,548,213 snapshots, six countries, three platforms and 17 weeks from December 17, 2025-April 8, 2026; 2,556 URLs manually/classifier categorized.

Key result

Weekly domain drift was about 56%; platforms differed by more than 80% for the same questions and URL-level drift was higher.

Limitation

monitored prompts and ChatGPT source-attribution subset remain.

Why this decision

Strong transparent independent-vendor triangulation for instability; monitored prompts and ChatGPT source-attribution subset remain.

Source classification and provenance
Study design
Not stated
Evidence stream
citation volatility and platform fragmentation
B2B relevance
High for measurement cadence and multi-platform scope.
Source type
full vendor longitudinal research article
Source class
Industry / original-data report
Transparency
High - full bases, dates, countries, metrics and stated limitations.
Dataset family
SISTRIX_AI_RESEARCH_INDEX_2025_2026
Provenance
Official SISTRIX search
Origin
master
Corrections applied
None
S037Decision: IncludeReview: AbstractEvidence: Clicks and zero-click behaviorB2B: High

Investigating Click Behaviors on Google Search Result Pages That Produce an AI Overview

Athena Chapekis, Anna Lieb, Sono Shah, Aaron Smith; Pew Research Center / 2026-08 / Academic research

Open direct source
Sample or setting

One month of browsing data from a representative panel of 900 US adults; mixed-effects logistic regression controlling for panelist and query attributes

Key result

Clicks to AIO-cited sources occurred in about 1% of AIO visits, and AIO exposure remained associated with fewer clicks and more session endings after query controls.

Limitation

Representative behavioral panel with explicit modeling; observational exposure still limits pure causal interpretation.

Why this decision

Representative behavioral panel with explicit modeling; observational exposure still limits pure causal interpretation.

Source classification and provenance
Study design
Not stated
Evidence stream
Clicks and zero-click behavior
B2B relevance
High
Source type
Preprint
Source class
Academic
Transparency
High
Dataset family
Pew AIO browsing panel
Provenance
Forward search from Pew AIO report; arXiv
Origin
master
Corrections applied
None
S038Decision: IncludeReview: AbstractEvidence: Clicks, zero-click and user experienceB2B: High

The Impact of Google AI Overviews on Publisher Traffic and User Experience: Evidence from a Field Experiment

Saharsh Agarwal, Ananya Sen; Indian School of Business, Carnegie Mellon / 2026-04 / Industry / practice research

Open direct source
Sample or setting

Randomized field experiment using a Chrome extension to show Google with or without AIOs

Key result

Conditional on appearing, AIOs reduced outbound organic clicks by 39.8% and increased zero-click searches by 34.5% without measurable improvement in perceived search quality.

Limitation

Extension-mediated exposure and a short experimental window may differ from native rollout behavior; no purchase or revenue outcome is observed.

Why this decision

Direct randomized field evidence on traffic and experience, unusually strong for this literature.

Source classification and provenance
Study design
Not stated
Evidence stream
Clicks, zero-click and user experience
B2B relevance
High
Source type
Working paper
Source class
Academic
Transparency
High
Dataset family
AIO Chrome field experiment
Provenance
SSRN search for AIO field experiments
Origin
master
Corrections applied
None
S039Decision: IncludeReview: AbstractEvidence: Commercial outcomesB2B: Medium

Generative Engine Optimization: A VLM and Agent Framework for Pinterest Acquisition Growth

Faye Zhang et al., Pinterest / 2026 / Academic research

Open direct source
Sample or setting

Production deployment at Pinterest comparing an optimization framework with a control at platform scale.

Key result

The authors report about 20% production traffic lift versus control.

Limitation

Rare production outcome evidence; assignment mechanics and full denominator are insufficiently disclosed for causal certainty.

Why this decision

Rare production outcome evidence; assignment mechanics and full denominator are insufficiently disclosed for causal certainty. Martinez matrix row 41.

Source classification and provenance
Study design
Not stated
Evidence stream
Commercial outcomes
B2B relevance
Medium
Source type
industry preprint
Source class
Prior-review mined
Transparency
Moderate
Dataset family
PINTEREST-GEO-PRODUCTION-2026
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S040Decision: IncludeReview: Landing-pageEvidence: Commercial outcomesB2B: Medium

Own the Agentic Commerce Journey

IBM Institute for Business Value and National Retail Federation / 2026-01 / Industry / practice research

Open direct source
Sample or setting

About 18,000 consumers in 23 countries plus approximately 200 executives.

Key result

Forty-five percent of consumers reported using AI somewhere in the purchase journey.

Limitation

Large multinational buyer-adoption evidence; self-report and retail scope mean it supports transfer hypotheses, not B2B causal claims.

Why this decision

Large multinational buyer-adoption evidence; self-report and retail scope mean it supports transfer hypotheses, not B2B causal claims. JY source 7.

Source classification and provenance
Study design
Not stated
Evidence stream
Commercial outcomes
B2B relevance
Medium
Source type
industry survey
Source class
Prior-review mined
Transparency
Moderate
Dataset family
IBM-NRF-AGENTIC-COMMERCE-2026
Provenance
JY Scauri review: retained source 7 of 34.
Origin
master
Corrections applied
None
S041Decision: IncludeReview: AbstractEvidence: Commercial recommendation manipulationB2B: Medium

Adversarial Search Engine Optimization for Large Language Models

Fredrik Nestaas, Edoardo Debenedetti, Florian Tramer; ETH Zurich / 2025-04 / Academic research

Open direct source
Sample or setting

Controlled pages tested on Bing, Perplexity and LLM browsing plugins

Key result

Adversarial page content can steer LLM search preferences across several commercial systems, establishing that retrieved webpages can manipulate downstream recommendations.

Limitation

Cross-system evidence of commercial vulnerability, though controlled pages are not a normal marketing intervention.

Why this decision

Cross-system evidence of commercial vulnerability, though controlled pages are not a normal marketing intervention.

Source classification and provenance
Study design
Not stated
Evidence stream
Commercial recommendation manipulation
B2B relevance
Medium
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
ASEO commercial attack study
Provenance
Martinez 2026 core matrix; OpenReview; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S042Decision: IncludeReview: AbstractEvidence: Competitive citation factorsB2B: High

What Gets Cited: Competitive GEO in AI Answer Engines

Rahul Vishwakarma, Shushant Kumar, Ratnesh Jamidar / 2026-05 / Academic research

Open direct source
Sample or setting

252,000 controlled trials injecting two competing documents across answer-engine settings

Key result

Topical relevance and context position dominated citation probability, while many surface-level content tactics were smaller or inconsistent.

Limitation

Injected competing documents bypass organic crawling and retrieval, so the large controlled effects are not production visibility estimates.

Why this decision

Large controlled factor study with direct competitive relevance; injected documents still bypass organic retrieval.

Source classification and provenance
Study design
Not stated
Evidence stream
Competitive citation factors
B2B relevance
High
Source type
Accepted conference preprint
Source class
Academic
Transparency
High
Dataset family
Competitive GEO trials
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix; also surfaced in Cyrus Shepard citation chain.
Origin
master
Corrections applied
None
S043Decision: IncludeReview: AbstractEvidence: Competitive content optimizationB2B: High

C-SEO Bench: Does Conversational SEO Work?

Haritz Puerto et al.; NAVER AI Lab and collaborators / 2025-12 / Academic research

Open direct source
Sample or setting

54 intervention-domain cases across multiple actors and QA settings

Key result

Only three of 54 cases showed a positive effect, none of the QA cases did, and gains approached zero when many competitors adopted the same intervention.

Limitation

Only 54 intervention-domain cases and largely benchmarked competition; production equilibrium and downstream outcomes remain unobserved.

Why this decision

Important counterevidence to universal GEO tactics and one of the few studies modeling competitive adoption.

Source classification and provenance
Study design
Not stated
Evidence stream
Competitive content optimization
B2B relevance
High
Source type
Peer-reviewed conference benchmark paper
Source class
Academic
Transparency
High
Dataset family
C-SEO Bench
Provenance
Martinez 2026 core matrix; NeurIPS proceedings; Olivier Martinez 45-study literature matrix; overlap with Cyrus Shepard 54-source sheet.
Origin
master
Corrections applied
None
S044Decision: IncludeReview: FullEvidence: consumer adoption and AI search exposureB2B: Medium-high - employed-adult work cut, not B2B buying.

Americans and AI 2026: Chatbots, Smart Devices and Views on Impact

Pew Research Center / 2026-06-17 / Industry / practice research

Open direct source
Sample or setting

Survey of 5,119 US adults conducted February 17-23, 2026; full topline and data downloads available.

Key result

44% reported chatbot use; 42% used chatbots to search for information; 38% of employed adults used them for work and 60% read AI search summaries.

Limitation

self-report and broad consumer population.

Why this decision

Independent current adoption denominator with transparent survey base; self-report and broad consumer population.

Source classification and provenance
Study design
Not stated
Evidence stream
consumer adoption and AI search exposure
B2B relevance
Medium-high - employed-adult work cut, not B2B buying.
Source type
full independent survey report with data
Source class
Industry / original-data report
Transparency
High - full report, dates, n, questionnaire and downloadable data.
Dataset family
PEW_AMERICANS_AI_2026
Provenance
Official Pew search and full report
Origin
master
Corrections applied
None
S045Decision: IncludeReview: AbstractEvidence: Consumer decision behaviorB2B: High

Comparing Traditional and LLM-Based Search for Consumer Choice: A Randomized Experiment

Sofia Eleni Spatharioti, David Rothschild, Daniel Goldstein, Jake Hofman; Microsoft Research / 2023-12 / Academic research

Open direct source
Sample or setting

Randomized online product-comparison tasks using LLM search or traditional search, plus confidence-highlighting intervention

Key result

LLM users finished faster with fewer and more complex queries and similar accuracy when answers were reliable, but over-relied on incorrect outputs; confidence highlighting improved error detection.

Limitation

Direct decision experiment supporting both efficiency and overreliance mechanisms; initial preprint predates the window but was revised within it.

Why this decision

Direct decision experiment supporting both efficiency and overreliance mechanisms; initial preprint predates the window but was revised within it.

Source classification and provenance
Study design
Not stated
Evidence stream
Consumer decision behavior
B2B relevance
High
Source type
Preprint
Source class
Academic
Transparency
High
Dataset family
Microsoft consumer-choice experiment
Provenance
Backward search from 2026 AI-search experiments
Origin
master
Corrections applied
None
S046Decision: IncludeReview: FullEvidence: Content optimization and citation prominenceB2B: Medium

GEO: Generative Engine Optimization

Pranjal Aggarwal et al.; Princeton, Georgia Tech, IIT Delhi, Allen Institute / 2024-08 / Academic research

Open direct source
Sample or setting

GEO-bench with 10,000 queries; five fixed Google-result documents; GPT-3.5 generations; nine rewrite strategies; 200 post-retrieval Perplexity tests

Key result

Quotations, statistics and cited sources can raise a document's relative answer share after it is already in context, with the famous roughly 40% result referring to relative position-adjusted word share rather than traffic or retrieval.

Limitation

Foundational controlled GEO experiment, but its fixed retrieved context sharply limits organic-discoverability claims.

Why this decision

Foundational controlled GEO experiment, but its fixed retrieved context sharply limits organic-discoverability claims.

Source classification and provenance
Study design
Not stated
Evidence stream
Content optimization and citation prominence
B2B relevance
Medium
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
GEO-Bench foundational
Provenance
Martinez 2026 core matrix; DOI record; Olivier Martinez 45-study literature matrix; also present in Cyrus Shepard 54-source evidence sheet.
Origin
master
Corrections applied
None
S047Decision: IncludeReview: AbstractEvidence: Content structure and citationsB2B: High

Structural Feature Engineering for Generative Engine Optimization: How Content Structure Shapes Citation Behavior

Junwei Yu, Mufeng Yang, Yepeng Ding, Hiroyuki Sato / 2026-03 / Academic research

Open direct source
Sample or setting

Structural page interventions tested across six generative engines

Key result

Headings, answer-first organization and other structural changes were associated with citation gains across six engines, but the external validation and human quality checks were limited.

Limitation

Direct cross-engine structural test with practical value, retained as provisional until stronger validation.

Why this decision

Direct cross-engine structural test with practical value, retained as provisional until stronger validation.

Source classification and provenance
Study design
Not stated
Evidence stream
Content structure and citations
B2B relevance
High
Source type
Preprint
Source class
Academic
Transparency
Medium
Dataset family
Six-engine structure experiment
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S048Decision: IncludeReview: AbstractEvidence: Content-platform engagementB2B: Medium

The Impact of AI Search on the Online Content Ecosystem: Evidence from Google and Reddit

Peibo Zhang, Ruomeng Cui, Dennis J. Zhang; Emory University, Washington University in St. Louis / 2026-05 / Academic research

Open direct source
Sample or setting

Difference-in-differences exploiting SFW versus NSFW Reddit eligibility for Google AI Overviews

Key result

AIO eligibility increased daily comments and commenting users by about 12%, especially for experiential content, but subsequent AI Mode largely eliminated those gains.

Limitation

Reddit eligibility is an imperfect proxy for overview exposure, and platform-specific community content may not generalize to B2B publishers.

Why this decision

Causal evidence that interface design and content type can reverse whether AI search complements or substitutes for a source platform.

Source classification and provenance
Study design
Not stated
Evidence stream
Content-platform engagement
B2B relevance
Medium
Source type
Preprint
Source class
Academic
Transparency
High
Dataset family
Reddit AIO policy natural experiment
Provenance
SSRN and arXiv search for AI search ecosystem effects
Origin
master
Corrections applied
None
S049Decision: IncludeReview: FullEvidence: conversion, engagement and machine readabilityB2B: High as commerce/complexity transfer evidence; indirect for SaaS.

AI Traffic Grows but Retail Sites Lag in AI Search Visibility

Adobe Digital Insights / 2026-04-16 / Industry / practice research

Open direct source
Sample or setting

More than 1T US retail visits plus a 5,000+ respondent US survey; January-March 2026; Adobe AI Content Visibility Checker benchmark.

Key result

In March 2026 AI-referred visits converted 42% better than non-AI visits; engagement was 12% higher and product-page machine-readability averaged 66%.

Limitation

comparator mixes channels and remains retail-only.

Why this decision

Large transaction base and latest wave show the conversion reversal; comparator mixes channels and remains retail-only.

Source classification and provenance
Study design
Not stated
Evidence stream
conversion, engagement and machine readability
B2B relevance
High as commerce/complexity transfer evidence; indirect for SaaS.
Source type
full first-party analytics and survey article
Source class
Industry / original-data report
Transparency
Moderate-high - full article and broad bases; absolute rates and merchant composition are not public.
Dataset family
ADOBE_DIGITAL_INSIGHTS_ROLLING_2024_2026
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official Adobe page
Origin
master
Corrections applied
None
S050Decision: IncludeReview: AbstractEvidence: Cross-engine source auditB2B: High

Generative Engine Optimization: How to Dominate AI Search

Mahe Chen, Xiaoxuan Wang, Kaiwen Chen, Nick Koudas; University of Toronto / 2025-09 / Academic research

Open direct source
Sample or setting

Large controlled audit across AI search platforms, verticals, languages and query paraphrases

Key result

AI search disproportionately surfaced earned-media sources over brand-owned and social sources, while source diversity, freshness and paraphrase stability varied materially by engine.

Limitation

Direct multi-platform audit with clear practical relevance, but still an observational snapshot.

Why this decision

Direct multi-platform audit with clear practical relevance, but still an observational snapshot.

Source classification and provenance
Study design
Not stated
Evidence stream
Cross-engine source audit
B2B relevance
High
Source type
Preprint
Source class
Academic
Transparency
Medium
Dataset family
AI Search source-mix audit
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix; overlap with Cyrus Shepard 54-source sheet.
Origin
master
Corrections applied
None
S051Decision: IncludeReview: AbstractEvidence: Cross-engine source selectionB2B: Medium

Generative AI Search Engines as Arbiters of Public Knowledge: An Audit of Bias and Authority

Alice Li, Luanne Sinnamon; University of British Columbia / 2024-10 / Academic research

Open direct source
Sample or setting

1,008 responses overall; 672-response Bing-Perplexity comparison phase

Key result

Bing and Perplexity used materially different source pools, with only about 26% domain overlap in the paired phase, undermining any assumption of a single universal AI ranking.

Limitation

Direct commercial-engine source audit with a clearly reported comparison base.

Why this decision

Direct commercial-engine source audit with a clearly reported comparison base.

Source classification and provenance
Study design
Not stated
Evidence stream
Cross-engine source selection
B2B relevance
Medium
Source type
Peer-reviewed journal proceedings paper
Source class
Academic
Transparency
Medium
Dataset family
Bing-Perplexity authority audit
Provenance
Martinez 2026 core matrix; DOI record; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S052Decision: IncludeReview: FullEvidence: cross-platform mention and citation fragmentationB2B: High - monitoring and model-selection implications.

Variability of Google Models: Gemini vs AIO vs AI Mode

Profound / 2026-07-14 / Industry / practice research

Open direct source
Sample or setting

15,155 brand-category pairs tracked daily in May 2026; 2,346 complete prompts, 218,178 responses and 2.64B citations across three Google surfaces.

Key result

Brands faced a median 8-point visibility gap across Google models; pairwise entity overlap was only about one-third and citation mixes differed.

Limitation

Large transparent within-provider comparison proves that one Google metric cannot represent all surfaces.

Why this decision

Large transparent within-provider comparison proves that one Google metric cannot represent all surfaces.

Source classification and provenance
Study design
Not stated
Evidence stream
cross-platform mention and citation fragmentation
B2B relevance
High - monitoring and model-selection implications.
Source type
full vendor longitudinal research article
Source class
Industry / original-data report
Transparency
High-moderate - full methods and bases; designed/customer prompt universe remains proprietary.
Dataset family
PROFOUND_GOOGLE_MODELS_MAY_2026
Provenance
Official Profound research hub and full-page inspection
Origin
master
Corrections applied
None
S053Decision: IncludeReview: AbstractEvidence: Cross-platform source selection and stabilityB2B: High

Characterizing Web Search in the Age of Generative AI

Elisabeth Kirsten et al.; Max Planck Institute and collaborators / 2026-07 / Academic research

Open direct source
Sample or setting

4,706 queries comparing Google organic with five generative systems from Google, OpenAI and Perplexity

Key result

Source footprints differed sharply by engine, and controllable surfaces changed 9% to 28% of repeated decisions, making AI visibility a distribution rather than a stable rank.

Limitation

Cross-sectional overlap audit identifies differences but cannot explain causality, stability or downstream user response.

Why this decision

Strong multi-surface, repeated-query audit directly relevant to measurement strategy.

Source classification and provenance
Study design
Not stated
Evidence stream
Cross-platform source selection and stability
B2B relevance
High
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
Generative web-search comparison
Provenance
Martinez 2026 core matrix; ACL Anthology; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S054Decision: IncludeReview: AbstractEvidence: Demand-guided content optimizationB2B: High

Mind Reader: Latent User Demand-Guided Content Optimization for Generative Search Engine

Tong Chen et al.; collaborators in China / 2026-07 / Academic research

Open direct source
Sample or setting

LLM-derived latent user demands used to generate and evaluate optimized content

Key result

Demand-guided optimization produced large reported visibility gains, but models generated much of both the demand signal and the evaluation.

Limitation

Direct content-demand hypothesis with circularity risk from LLM-generated inputs and judges.

Why this decision

Direct content-demand hypothesis with circularity risk from LLM-generated inputs and judges.

Source classification and provenance
Study design
Not stated
Evidence stream
Demand-guided content optimization
B2B relevance
High
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
Mind Reader demand benchmark
Provenance
Martinez 2026 core matrix; ACL Anthology; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S055Decision: IncludeReview: Landing-pageEvidence: Discovery and evaluationB2B: Direct B2B

Responsive Inside the B2B Buyer's Mind

Responsive / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

350 strategic enterprise procurement participants

Key result

AI use was more common in discovery and evaluation than in final decision-making; public summaries report different rates by stage.

Limitation

Vendor white paper and procurement-heavy population; exact question bases require full report extraction.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
B2B buyer research
Evidence stream
Discovery and evaluation
B2B relevance
Direct B2B
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
RESPONSIVE-2025
Provenance
AI Visibility Evidence Register v0.1
Origin
master
Corrections applied
None
S056Decision: IncludeReview: FullEvidence: Discovery and validationB2B: Direct B2B

Omniscient Digital / Wynter B2B SaaS buyer panel

Omniscient / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

100 B2B SaaS leaders in a message-testing panel

Key result

42 mentioned LLMs during first-stage research, but only 11 used one as the absolute first step; vendor sites, reviews and peers remained important validation sources.

Limitation

Small qualitative/quantified panel; useful for process texture, not prevalence estimation.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
B2B buyer research
Evidence stream
Discovery and validation
B2B relevance
Direct B2B
Source type
Commercial agencies
Source class
Baseline evidence register
Transparency
Medium
Dataset family
OMNISCIENT-WYNTER-2025
Provenance
AI Visibility Evidence Register v0.1
Origin
master
Corrections applied
None
S057Decision: IncludeReview: Landing-pageEvidence: Downstream influenceB2B: Indirect / transfer

Matomo measurement documentation

Matomo / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

Analytics classification rules rather than outcome research

Key result

Human AI-referral visits, zero-click influence and bot/agent traffic are separate classes. Its AI Assistant channel is not retroactive; historical visits remain in their original Referral/Direct classifications; server logs are required for bots that do not execute JavaScript

Limitation

Measurement guidance, not evidence of market size or commercial impact. Useful for the instrumentation chapter only.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Downstream journeys and attribution leakage
Evidence stream
Downstream influence
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Low
Dataset family
D5
Provenance
AI Visibility Evidence Register v0.1
Origin
master
Corrections applied
None
S058Decision: IncludeReview: FullEvidence: Downstream influenceB2B: Indirect / transfer

Profound “AI Mention Effect”; joined AI-conversation and browsing panel with backward-placebo design

Profound / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

More than 2M AI conversations and associated browsing; double-opt-in US panel; ChatGPT, Gemini and Google AI Overviews; January-June 2026. Exposure: brand introduced in the answer, not prompt. Outcome: own-site visit within seven days

Key result

Treated versus forecast visit rates: Gemini 5.42% vs 2.21%, +3.21pp; AIO 7.79% vs 4.83%, +2.96pp; ChatGPT 6.39% vs 4.33%, +2.07pp. Software lift: +4.0pp AIO, +3.0pp ChatGPT, +3.6pp Gemini. Only about 1-2.5% of ChatGPT downstream visits carried a matched AI-referral parameter; more than 97% did not. 42% of visits occurred within 24 hours

Limitation

Observational, not randomised; page visit, not conversion; users/exposures/brands per cell and panel provider not disclosed. Commercial vendor interest. Independence from the Similarweb and Scrunch joined panels is not yet verified.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Downstream journeys and attribution leakage
Evidence stream
Downstream influence
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
D1
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official Profound research hub
Origin
master
Corrections applied
None
S059Decision: IncludeReview: FullEvidence: Downstream influenceB2B: Indirect / transfer

Scrunch AI preprint; production joined conversation-clickstream panel with matched backward placebos, stance classifier and same-response category controls

Scrunch / 2026 or earlier / Academic research

Open direct source
Sample or setting

Two undisclosed English-speaking markets; ChatGPT, Claude and Gemini; early 2026; consumer-brand lexicon. Aggregate panel size and cell sizes withheld. Outcome window: seven days

Key result

For recommendations to people with no recent observed brand engagement, branded Google search rose from 2.9% to 7.2%, +4.3pp; own-site visits 3.1% to 5.5%, +2.4pp; retailer-page visits 0.8% to 1.8%, +1.0pp. Neutral name-drops produced only +1.8/+1.1/+0.3pp. Unnamed same-category brands in the same response were near flat

Limitation

Preprint, not peer reviewed; consumer brands; no transactions; a brand-specific within-session intent shock remains unidentified. Panel sizes/markets withheld. Stronger controls than vendor blog studies, but independence from D1/D3 must be confirmed.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Downstream journeys and attribution leakage
Evidence stream
Downstream influence
B2B relevance
Indirect / transfer
Source type
Independent / platform
Source class
Baseline evidence register
Transparency
High
Dataset family
D2
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official author preprint
Origin
master
Corrections applied
None
S060Decision: IncludeReview: FullEvidence: Downstream influenceB2B: Indirect / transfer

Similarweb Downstream Impact of AI Visibility; joined ChatGPT recommendation and later browsing

Similarweb / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

“Thousands” of real journeys in Finance, Travel and Beauty; public description indicates US desktop, six months, seven-day follow-up; excludes prior brand visits and prompts naming the brand

Key result

Users shown a ChatGPT-recommended brand were 2.5 times more likely to visit it than users shown a competitor recommendation; 55.9% of AI-influenced visits arrived through search; influenced visitors viewed about twice the pages and spent about twice the time

Limitation

Exact exposed/control bases, CIs, date range and matching details are not in the public blog. Observational and no purchase/revenue outcome. The 2.5x headline resembles D1 but the estimand differs; do not call it an independent replication until panel lineage is confirmed.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Downstream journeys and attribution leakage
Evidence stream
Downstream influence
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
SIMILARWEB-DOWNSTREAM
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official Similarweb page
Origin
master
Corrections applied
None
S061Decision: IncludeReview: Landing-pageEvidence: Downstream influenceB2B: Direct B2B

The Road from $4M to $5M ARR

Tally / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

Tally onboarding surveys and internal growth reporting

Key result

Tally reported AI tools as its largest acquisition source by April 2026 via onboarding survey.

Limitation

Methods, absolute counts and attribution wording are incomplete; company-specific and highly selected. Useful as existence proofs only, not benchmarks or causal evidence.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Downstream journeys and attribution leakage
Evidence stream
Downstream influence
B2B relevance
Direct B2B
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Low
Dataset family
D4
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official company blog
Origin
master
Corrections applied
None
S062Decision: IncludeReview: Landing-pageEvidence: Downstream influenceB2B: Direct B2B

How We're Adapting SEO for LLMs and AI Search

Vercel / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

Vercel internal acquisition analytics

Key result

Vercel reported ChatGPT referred about 10% of new signups in June 2025, up from 4.8% one month earlier and 1% six months earlier.

Limitation

Methods, absolute counts and attribution wording are incomplete; company-specific and highly selected. Useful as existence proofs only, not benchmarks or causal evidence.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Downstream journeys and attribution leakage
Evidence stream
Downstream influence
B2B relevance
Direct B2B
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Low
Dataset family
D4
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official company article
Origin
master
Corrections applied
None
S063Decision: IncludeReview: AbstractEvidence: E-commerce product visibilityB2B: High

E-GEO: A Testbed for Generative Engine Optimization in E-Commerce

Puneet S. Bagga et al.; MIT and collaborators / 2025-11 / Academic research

Open direct source
Sample or setting

13,747 consumer queries, ten Amazon listings each, five engines, seven rewriters and fifteen hand-written heuristics

Key result

Ten of fifteen common heuristics were neutral or harmful, while meta-optimized prompts performed better and converged on a more stable cross-domain structure.

Limitation

Large direct product-visibility testbed with unusually relevant commercial tasks, though it uses fixed retrieved listings.

Why this decision

Large direct product-visibility testbed with unusually relevant commercial tasks, though it uses fixed retrieved listings.

Source classification and provenance
Study design
Not stated
Evidence stream
E-commerce product visibility
B2B relevance
High
Source type
Preprint
Source class
Academic
Transparency
High
Dataset family
E-GEO e-commerce benchmark
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix; overlap with Cyrus Shepard 54-source sheet.
Origin
master
Corrections applied
None
S064Decision: IncludeReview: FullEvidence: early referrals, engagement and conversionB2B: High as temporal/complexity evidence; indirect retail population.

Traffic to U.S. Retail Websites from Generative AI Sources Jumps 1,200 Percent

Adobe Analytics / 2025-03-17 / Industry / practice research

Open direct source
Sample or setting

More than 1T US retail visits plus a 5,000-person US survey; July 2024-February 2025.

Key result

AI referral traffic rose 1,200% from July 2024; conversion remained 9% below non-AI sources in February 2025, improving from a 43% gap.

Limitation

same Adobe rolling family, not an independent confirmation.

Why this decision

Essential baseline showing conversion changed sign over time; same Adobe rolling family, not an independent confirmation.

Source classification and provenance
Study design
Not stated
Evidence stream
early referrals, engagement and conversion
B2B relevance
High as temporal/complexity evidence; indirect retail population.
Source type
full first-party analytics and survey article
Source class
Industry / original-data report
Transparency
Moderate-high - full article, survey base and visit scale; no absolute channel rates.
Dataset family
ADOBE_DIGITAL_INSIGHTS_ROLLING_2024_2026
Provenance
Official Adobe search; predecessor wave identified during lineage review
Origin
master
Corrections applied
None
S065Decision: IncludeReview: FullEvidence: ecommerce conversion and order valueB2B: High as conversion/complexity transfer; direct population is ecommerce.

AI-Referred Shoppers Convert Better and Spend More: What Shopify's Early Data Shows

Shopify / 2026 / Industry / practice research

Open direct source
Sample or setting

Shopify Q1 2026 commerce data; AI-referred versus organic-search product-detail-page sessions across 25 merchant categories; January 2025-March 2026 growth window.

Key result

AI-referred PDP sessions converted nearly 50% better, had 14% higher AOV and outperformed organic in 23 of 25 categories.

Limitation

merchant/session counts, absolute rates and selection method are not disclosed.

Why this decision

Large-platform commerce evidence with category consistency; merchant/session counts, absolute rates and selection method are not disclosed.

Source classification and provenance
Study design
Not stated
Evidence stream
ecommerce conversion and order value
B2B relevance
High as conversion/complexity transfer; direct population is ecommerce.
Source type
full platform first-party analytics article
Source class
Industry / original-data report
Transparency
Moderate-low - full article and definitions but no absolute sample or rates.
Dataset family
SHOPIFY_COMMERCE_AI_Q1_2026
Provenance
Official Shopify Enterprise search
Origin
master
Corrections applied
None
S066Decision: IncludeReview: FullEvidence: End-to-end retrieval and generationB2B: High

SAGEO Arena: A Realistic Environment for Evaluating Search-Augmented Generative Engine Optimization

Sunghwan Kim et al.; Yonsei University and collaborators / 2026-02 / Academic research

Open direct source
Sample or setting

Reproducible full pipeline over a large web corpus retaining structural page information

Key result

Techniques that help downstream generation often degrade retrieval or reranking, while structural information partly mitigates the trade-off, demonstrating that citation optimization can reduce discoverability.

Limitation

Runs in a reproducible non-commercial pipeline whose retriever, corpus and generator may not represent proprietary answer engines.

Why this decision

One of the strongest stage-separated end-to-end GEO studies, albeit in a non-commercial pipeline.

Source classification and provenance
Study design
Not stated
Evidence stream
End-to-end retrieval and generation
B2B relevance
High
Source type
Accepted conference preprint
Source class
Academic
Transparency
High
Dataset family
SAGEO Arena
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix; overlap with Cyrus Shepard 54-source sheet.
Origin
master
Corrections applied
None
S067Decision: IncludeReview: Landing-pageEvidence: EvaluationB2B: Direct B2B

MarketingGraham tech-buyer behaviour survey

MarketingGraham / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

792 respondents; five questions fielded November 2025

Key result

Reports buyer-rated usefulness of AI in technology research and related evaluation behaviour.

Limitation

Recruitment, geography, weighting and several denominator details are limited; directional only until full extraction.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
B2B buyer research
Evidence stream
Evaluation
B2B relevance
Direct B2B
Source type
Independent consultancy
Source class
Baseline evidence register
Transparency
Low
Dataset family
MARKETINGGRAHAM-2026
Provenance
AI Visibility Evidence Register v0.1
Origin
master
Corrections applied
None
S068Decision: IncludeReview: FullEvidence: Evaluation and validationB2B: Direct B2B

Gartner B2B buyer survey

Gartner / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

645 B2B buyers; 2026 newsroom summary

Key result

45% used generative AI in a recent purchase, while 69% preferred to validate AI-generated insights with a sales representative.

Limitation

Public newsroom summary omits full questionnaire and sampling detail. Supports AI-plus-human validation, not AI causality.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
B2B buyer research
Evidence stream
Evaluation and validation
B2B relevance
Direct B2B
Source type
Commercial analyst
Source class
Baseline evidence register
Transparency
Medium
Dataset family
GARTNER-B2B-2026
Provenance
AI Visibility Evidence Register v0.1; Existing synthesis; official Gartner newsroom
Origin
master
Corrections applied
None
S069Decision: IncludeReview: FullEvidence: Evaluation and validationB2B: Direct B2B

TrustRadius 2026 B2B Buying Disconnect

TrustRadius / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

1,862 verified technology buyers plus 444 technology vendors; global network; online survey in January 2026 about a purchase in the prior year

Key result

63% of buyers used AI during a technology purchase; 94% of AI-using buyers fact-checked the output at least sometimes.

Limitation

Self-reported behaviour from a review-platform network. Useful for verification behaviour, not causal pipeline impact.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
B2B buyer research
Evidence stream
Evaluation and validation
B2B relevance
Direct B2B
Source type
Commercial marketplace
Source class
Baseline evidence register
Transparency
High
Dataset family
TRUSTRADIUS-2026
Provenance
AI Visibility Evidence Register v0.1; Existing synthesis; official TrustRadius page
Origin
master
Corrections applied
None
S070Decision: IncludeReview: FullEvidence: fan-out queries and retrieval mechanicsB2B: High for content/topic strategy; no commercial outcome.

What AI Engines Actually Search For and Why ChatGPT Never Searches the Same Way Twice

Profound / 2026-04-30 / Industry / practice research

Open direct source
Sample or setting

Random sample of 10,000 prompts over 14 days in late March-mid April 2026 across ChatGPT, Perplexity and Copilot; captured internal search queries.

Key result

91% of ChatGPT fan-out queries were unique across runs versus 14% for Perplexity; original-prompt word overlap was 13% for ChatGPT versus 88% for Perplexity.

Limitation

models and platform implementation can change quickly.

Why this decision

Strong cross-engine retrieval evidence; models and platform implementation can change quickly.

Source classification and provenance
Study design
Not stated
Evidence stream
fan-out queries and retrieval mechanics
B2B relevance
High for content/topic strategy; no commercial outcome.
Source type
full vendor longitudinal retrieval study
Source class
Industry / original-data report
Transparency
High-moderate - full methodology, dates and base; raw prompts are proprietary and some classification uses LLMs.
Dataset family
PROFOUND_FANOUT_10K_2026
Provenance
Official Profound research hub and full-page inspection
Origin
master
Corrections applied
None
S071Decision: IncludeReview: AbstractEvidence: Global exposure and source diversityB2B: Medium

The Rise of AI Search: Implications for Information Markets and Human Judgement at Scale

Sinan Aral, Haiwen Li, Rui Zuo; MIT / 2026-02 / Academic research

Open direct source
Sample or setting

24,000 queries in 243 countries producing 2.8 million AI and traditional search results across 2024 and 2025

Key result

AI-search exposure expanded sharply across countries and topics while surfacing fewer long-tail sources, less response variety and more low-credibility sources than traditional search.

Limitation

Global query-result audits measure exposure and concentration, not whether users noticed, trusted, clicked or purchased.

Why this decision

Large global longitudinal audit connecting rollout policy to source concentration.

Source classification and provenance
Study design
Not stated
Evidence stream
Global exposure and source diversity
B2B relevance
Medium
Source type
Preprint
Source class
Academic
Transparency
High
Dataset family
Global AI search exposure panel
Provenance
Forward citation from Human Trust in AI Search
Origin
master
Corrections applied
None
S072Decision: IncludeReview: AbstractEvidence: Information exposure and coverageB2B: Medium

Answer Bubbles: Information Exposure in AI-Mediated Search

Michelle Huang, Agam Goyal, Koustuv Saha, Eshwar Chandrasekharan; University of Illinois Chicago / 2026-03 / Academic research

Open direct source
Sample or setting

Cross-system comparison using Natural Questions-derived tasks and atomic content-unit coverage metrics

Key result

Generative and traditional search can provide similar topical coverage while exposing users to different and narrower combinations of sources and source content.

Limitation

Adds coverage and exposure metrics beyond simple citation counts; dataset age limits current-market inference.

Why this decision

Adds coverage and exposure metrics beyond simple citation counts; dataset age limits current-market inference.

Source classification and provenance
Study design
Not stated
Evidence stream
Information exposure and coverage
B2B relevance
Medium
Source type
Preprint
Source class
Academic
Transparency
High
Dataset family
Answer Bubbles exposure benchmark
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S073Decision: IncludeReview: FullEvidence: interface effects on referral attributionB2B: High for attribution-system interpretation.

ChatGPT Referral Traffic Near Triples Overnight

Similarweb / 2026-05-25 / Industry / practice research

Open direct source
Sample or setting

Similarweb desktop panel; ChatGPT referrals April 30-May 20, 2026 around a May 7 homepage-link interface change.

Key result

Week-on-week referrals rose 157.7% and homepage referrals 354.7%; homepage share moved to roughly 60%.

Limitation

short uncontrolled window.

Why this decision

Strong demonstration that product UI can change measurable referrals without equivalent demand change; short uncontrolled window.

Source classification and provenance
Study design
Not stated
Evidence stream
interface effects on referral attribution
B2B relevance
High for attribution-system interpretation.
Source type
full vendor quasi-experiment article
Source class
Industry / original-data report
Transparency
Moderate - full article gives window and indexed change; absolute bases are missing.
Dataset family
SIMILARWEB_CHATGPT_UI_REFERRAL_2026
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official Similarweb article
Origin
master
Corrections applied
None
S074Decision: IncludeReview: AbstractEvidence: Latent source preferenceB2B: High

In Agents We Trust, but Who Do Agents Trust? Latent Source Preferences Steer LLM Generations

Mohammad Aflah Khan et al.; Max Planck Institute, Adobe Research and collaborators / 2026-02 / Academic research

Open direct source
Sample or setting

Controlled synthetic and real tasks across 12 LLMs from six providers

Key result

Several models had strong predictable source preferences that could outweigh content, changed with framing and persisted despite instructions to avoid bias.

Limitation

Controlled source cues, including synthetic tasks, do not show live-web retrieval, stable production behavior or buyer outcomes.

Why this decision

Broad controlled confirmation that source identity itself can steer what agents present.

Source classification and provenance
Study design
Not stated
Evidence stream
Latent source preference
B2B relevance
High
Source type
Preprint
Source class
Academic
Transparency
High
Dataset family
Twelve-model latent-source audit
Provenance
OpenReview and arXiv search for source preference
Origin
master
Corrections applied
None
S075Decision: IncludeReview: AbstractEvidence: LLM ranker manipulationB2B: Medium

The Ranking Blind Spot: Decision Hijacking in LLM-Based Text Ranking

Yaoyao Qian et al.; Penn State and collaborators / 2025-11 / Academic research

Open direct source
Sample or setting

Adversarial experiments on LLM-based text-ranking tasks

Key result

LLM rankers contain a decision-making blind spot that lets optimized text hijack ranking outcomes without equivalent merit gains.

Limitation

Mechanistically relevant to answer-engine reranking but not a production web audit.

Why this decision

Mechanistically relevant to answer-engine reranking but not a production web audit.

Source classification and provenance
Study design
Not stated
Evidence stream
LLM ranker manipulation
B2B relevance
Medium
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
Decision hijacking ranker tests
Provenance
Martinez 2026 core matrix; ACL Anthology; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S076Decision: IncludeReview: AbstractEvidence: Longitudinal AI Overview auditB2B: High

Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact

Haofei Xu, Umar Iqbal, Jacob M. Montgomery; Washington University in St. Louis / 2026-05 / Academic research

Open direct source
Sample or setting

55,393 trending queries across 19 categories over 40 days; 98,020 atomic claims

Key result

AIO activation averaged 13.7% but 64.7% for question queries; nearly 30% of cited domains were absent from co-displayed results and 11% of claims lacked support from cited pages.

Limitation

Forty-day trending-query window may overrepresent newsworthy demand, and the audit does not observe user clicks or commercial outcomes.

Why this decision

Large longitudinal audit covering activation, distinct source selection and citation fidelity.

Source classification and provenance
Study design
Not stated
Evidence stream
Longitudinal AI Overview audit
B2B relevance
High
Source type
Preprint
Source class
Academic
Transparency
High
Dataset family
AIO longitudinal audit
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S077Decision: IncludeReview: FullEvidence: market/language fragmentation and citationsB2B: High for international SaaS.

How Query Language Reshapes AI Citations

Profound / 2026-04-21 / Industry / practice research

Open direct source
Sample or setting

3.25B citations across seven models, 14 countries and native-language prompts in March 2026; source types classified.

Key result

Social citation rates varied materially by engine and language; AI Overviews used social sources at 15.3% versus Gemini 3.6% overall.

Limitation

customer prompt demand may not represent populations.

Why this decision

Large current evidence that English-only/global tracking is unsafe; customer prompt demand may not represent populations.

Source classification and provenance
Study design
Not stated
Evidence stream
market/language fragmentation and citations
B2B relevance
High for international SaaS.
Source type
full vendor cross-market citation study
Source class
Industry / original-data report
Transparency
High-moderate - full method, countries and counts; category composition and raw data are proprietary.
Dataset family
PROFOUND_CITATIONS_API_LANGUAGE_MARCH_2026
Provenance
Official Profound research hub and full-page inspection; Cyrus Shepard 54-source evidence sheet; Profound citation chain.
Origin
master
Corrections applied
None
S078Decision: IncludeReview: Landing-pageEvidence: Measurement/reliabilityB2B: High

AI Brand Recommendation Inconsistency

SparkToro and Gumshoe.ai / 2026-01 / Industry / practice research

Open direct source
Sample or setting

600 volunteers, 12 product categories, three AI tools and 2,961 completed prompt runs.

Key result

Brand recommendations varied substantially across repeated identical prompts and platforms.

Limitation

Rare human-run repeated-measures evidence that one-shot visibility scores are unstable; convenience volunteers and limited categories remain.

Why this decision

Rare human-run repeated-measures evidence that one-shot visibility scores are unstable; convenience volunteers and limited categories remain. JY source 27.

Source classification and provenance
Study design
Not stated
Evidence stream
Measurement/reliability
B2B relevance
High
Source type
industry repeated-measures experiment
Source class
Prior-review mined
Transparency
High
Dataset family
SPARKTORO-GUMSHOE-INCONSISTENCY-2026
Provenance
JY Scauri review: retained source 27 of 34.
Origin
master
Corrections applied
None
S079Decision: IncludeReview: AbstractEvidence: Mechanism/optimizationB2B: Medium

From Experience to Skill: Multi-Agent Generative Engine Optimization via Reusable Strategy Learning

Beining Wu et al. / 2026 / Academic research

Open direct source
Sample or setting

Twin-Branch protocol across three engines using reusable strategy memory.

Key result

Reusable agent strategies improved visibility while attempting to preserve attribution fidelity.

Limitation

Multi-engine optimization evidence; frozen evaluation contexts and automated metrics limit live-market inference.

Why this decision

Multi-engine optimization evidence; frozen evaluation contexts and automated metrics limit live-market inference. Martinez matrix row 24.

Source classification and provenance
Study design
Not stated
Evidence stream
Mechanism/optimization
B2B relevance
Medium
Source type
peer-reviewed conference paper
Source class
Prior-review mined
Transparency
High
Dataset family
MAGEO-2026
Provenance
Martinez 2026 core matrix; ACL Anthology; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S080Decision: IncludeReview: Landing-pageEvidence: Mechanism/optimizationB2B: High

How Big Are Google's Grounding Chunks?

Dan Petrovic / DEJAN / 2026 / Industry / practice research

Open direct source
Sample or setting

7,060 queries with at least three sources, 2,275 tokenized pages and 883,262 observed snippets from multiple client verticals.

Key result

Median grounding budget was about 1,929 words per query; rank-one sources received a median 531 words versus 266 for rank five, and long pages had lower coverage.

Limitation

Detailed primary page and discussion clarify measurement; private client query selection, no confounder controls and no public data limit causal conclusions.

Why this decision

Detailed primary page and discussion clarify measurement; private client query selection, no confounder controls and no public data limit causal conclusions. JY source 32.

Source classification and provenance
Study design
Not stated
Evidence stream
Mechanism/optimization
B2B relevance
High
Source type
industry technical audit
Source class
Prior-review mined
Transparency
Moderate
Dataset family
DEJAN-GROUNDING-CHUNKS-2026
Provenance
JY Scauri review: retained source 32 of 34; primary technical page checked including methodology discussion.
Origin
master
Corrections applied
None
S081Decision: IncludeReview: AbstractEvidence: Mechanism/optimizationB2B: High

IF-GEO: Conflict-Aware Instruction Fusion for Multi-Query Generative Engine Optimization

Heyang Zhou et al. / 2026 / Academic research

Open direct source
Sample or setting

Multi-query optimization using instruction fusion and downside-risk estimation.

Key result

Optimizing across prompt sets can improve average visibility while controlling losses on conflicting queries.

Limitation

Supports portfolio-level prompt measurement rather than single-prompt optimization.

Why this decision

Supports portfolio-level prompt measurement rather than single-prompt optimization. Martinez matrix row 23.

Source classification and provenance
Study design
Not stated
Evidence stream
Mechanism/optimization
B2B relevance
High
Source type
peer-reviewed conference paper
Source class
Prior-review mined
Transparency
High
Dataset family
IF-GEO-2026
Provenance
Martinez 2026 core matrix; ACL Anthology; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S082Decision: IncludeReview: AbstractEvidence: Mechanism/optimizationB2B: Medium

AgenticGEO: A Self-Evolving Agentic System for Generative Engine Optimization

Jiaqi Yuan et al. / 2026 / Academic research

Open direct source
Sample or setting

Fourteen baselines across GEO-Bench, MS MARCO and an Amazon-derived dataset on two open-weight engines.

Key result

Agentic optimization improved multi-objective visibility-quality frontiers over static baselines.

Limitation

Broad benchmark evidence but open-model laboratory results may not transfer to commercial engines.

Why this decision

Broad benchmark evidence but open-model laboratory results may not transfer to commercial engines. Martinez matrix row 20; Cyrus overlap.

Source classification and provenance
Study design
Not stated
Evidence stream
Mechanism/optimization
B2B relevance
Medium
Source type
preprint
Source class
Prior-review mined
Transparency
High
Dataset family
AGENTICGEO-2026
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix; overlap with Cyrus Shepard 54-source sheet.
Origin
master
Corrections applied
None
S083Decision: IncludeReview: AbstractEvidence: Mechanism/optimizationB2B: Medium

What Generative Search Engines Like and How to Optimize Web Content Cooperatively

Yujiang Wu et al. / 2026 / Academic research

Open direct source
Sample or setting

AutoGEO preference-learning system with five fixed documents, one optimized target and multiple generators.

Key result

Reported mean visibility gains of 35.99% in fixed-context tests while balancing utility constraints.

Limitation

Rigorous optimization evidence, but fixed retrieved sets are not the open web.

Why this decision

Rigorous optimization evidence, but fixed retrieved sets are not the open web. Martinez matrix row 12; overlap with Cyrus Shepard sheet.

Source classification and provenance
Study design
Not stated
Evidence stream
Mechanism/optimization
B2B relevance
Medium
Source type
peer-reviewed conference paper
Source class
Prior-review mined
Transparency
High
Dataset family
AUTOGEO-2026
Provenance
Martinez 2026 core matrix; ICLR proceedings; Olivier Martinez 45-study literature matrix; overlap with Cyrus Shepard 54-source sheet.
Origin
master
Corrections applied
None
S084Decision: IncludeReview: AbstractEvidence: Mechanism/optimizationB2B: Medium

Diagnosing and Repairing Citation Failures in Generative Engine Optimization

Zhihua Tian et al. / 2026 / Academic research

Open direct source
Sample or setting

AgentGEO evaluated on MIMIQ documents and HTML pages with diagnosis and constrained edits.

Key result

Reported more than 40% relative citation gain while modifying about 5% of text, though reported counts require reconciliation.

Limitation

Promising constrained-edit evidence; counting inconsistencies prevent high-confidence use.

Why this decision

Promising constrained-edit evidence; counting inconsistencies prevent high-confidence use. Martinez matrix row 19; Cyrus overlap.

Source classification and provenance
Study design
Not stated
Evidence stream
Mechanism/optimization
B2B relevance
Medium
Source type
preprint
Source class
Prior-review mined
Transparency
Moderate
Dataset family
AGENTGEO-MIMIQ-2026
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix; overlap with Cyrus Shepard 54-source sheet.
Origin
master
Corrections applied
None
S085Decision: IncludeReview: FullEvidence: Multilingual source attributionB2B: Medium

AI Chatbot Accountability in the Age of Algorithmic Gatekeeping: Comparing Generative Search Engine Political Information Retrieval Across Five Languages

Joanne Kuai et al.; Karlstad University and collaborators / 2025-02 / Academic research

Open direct source
Sample or setting

Copilot audit on the 2024 Taiwan election: 200 prompt-level cases and 1,501 linked-statement cases in five languages

Key result

Nearly half of sampled answers contained factual errors, only about 63% of referenced content was both factually correct and accurately sourced, and performance varied sharply by language.

Limitation

Single Copilot election snapshot in a political domain limits transfer to commercial recommendations and current models.

Why this decision

Detailed multilingual commercial-engine source audit with manual coding and strong intercoder reliability.

Source classification and provenance
Study design
Not stated
Evidence stream
Multilingual source attribution
B2B relevance
Medium
Source type
Peer-reviewed journal article
Source class
Academic
Transparency
High
Dataset family
Five-language Copilot election audit
Provenance
Journal search for generative search gatekeeping
Origin
master
Corrections applied
None
S086Decision: IncludeReview: AbstractEvidence: Off-page signals and business visibilityB2B: High

Off-Page Signals Have Model-Specific Effects on Generative AI Search Visibility: Evidence from a Cross-Platform Audit of 2,729 Businesses Across Five Generative AI Systems

Joel House; independent researcher / 2026-05 / Industry / practice research

Open direct source
Sample or setting

2,729 businesses, about 95 prompts each, five systems and 266,844 business-model-prompt observations enriched with web and platform signals

Key result

Off-page associations varied by engine rather than supporting one universal authority recipe, with signals such as reviews, directories and social presence showing model-specific effects.

Limitation

Large directly commercial cross-platform audit, retained with independence and observational-confounding caveats.

Why this decision

Large directly commercial cross-platform audit, retained with independence and observational-confounding caveats.

Source classification and provenance
Study design
Not stated
Evidence stream
Off-page signals and business visibility
B2B relevance
High
Source type
Working paper with archived dataset
Source class
Academic
Transparency
Medium
Dataset family
Five-engine business visibility audit
Provenance
SSRN search for business AI visibility audits
Origin
master
Corrections applied
None
S087Decision: IncludeReview: FullEvidence: on-site behavior and intentB2B: High for visit-quality interpretation; site mix not disclosed.

AI Visitors Show Stronger Intent Signals Than Traditional Channels

Microsoft Clarity / 2026-04-14 / Industry / practice research

Open direct source
Sample or setting

More than 30B sessions from the 2,000 Clarity projects receiving the most LLM traffic; October 2025-March 2026; quick backs, duration, pages, scroll and rage clicks.

Key result

AI-referred traffic grew 22% and showed fewer early abandonment signals and deeper session intent than traditional channels, while using fewer pages.

Limitation

deliberately selects top AI-traffic projects and does not measure pipeline.

Why this decision

Massive session base and explicit behavioral metrics; deliberately selects top AI-traffic projects and does not measure pipeline.

Source classification and provenance
Study design
Not stated
Evidence stream
on-site behavior and intent
B2B relevance
High for visit-quality interpretation; site mix not disclosed.
Source type
full first-party behavioral analytics study
Source class
Industry / original-data report
Transparency
High-moderate - full method and session/project bases; project composition and exact weighted rates are proprietary.
Dataset family
MICROSOFT_CLARITY_TOP_AI_PROJECTS_2025_2026
Provenance
Official Microsoft Clarity search
Origin
master
Corrections applied
None
S088Decision: IncludeReview: FullEvidence: on-site engagementB2B: Medium-high - cross-site engagement benchmark without B2B weighting.

AI Visitors Visit Fewer Pages and Bounce More Often Than Traditional Search Visitors

Ahrefs / 2025 / Industry / practice research

Open direct source
Sample or setting

Ahrefs Web Analytics rolling sample of 81,947 sites; May-June 2025; AI referrals compared with search and all traffic.

Key result

AI visitors averaged 4.0 pages versus 5.2 for search and had higher bounce, contradicting a universal engagement premium.

Limitation

same exact dataset as IND-008 and cannot count as independent.

Why this decision

Important negative finding from a large panel; same exact dataset as IND-008 and cannot count as independent.

Source classification and provenance
Study design
Not stated
Evidence stream
on-site engagement
B2B relevance
Medium-high - cross-site engagement benchmark without B2B weighting.
Source type
full vendor research article
Source class
Industry / original-data report
Transparency
Moderate-high - full article and sample; mix/weighting remain unclear.
Dataset family
AHREFS_WEB_ANALYTICS_ROLLING
Provenance
Existing evidence register; official Ahrefs canonical article verified
Origin
master
Corrections applied
None
S089Decision: IncludeReview: Landing-pageEvidence: On-site outcomesB2B: Indirect / transfer

Adobe AI-sourced traffic report

Adobe / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

More than 1T US retail visits and 100M SKUs; industry data October 2024-December 2025; survey of 1,000+ consumers in November 2025

Key result

2025 holiday AI retail traffic converted 31% better than non-AI and produced 32% more revenue per visit. Technology/software had the highest AI-driven visit share, +120% YoY, and AI visits showed +35% engagement, 43% lower bounce, +46% time and +22% pages

Limitation

Tech/software outcome is engagement, not SaaS purchase or pipeline. Consumer and last-click. Overlaps O1 as an earlier wave; do not count the two Adobe percentages as independent studies.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
On-site engagement, conversion and revenue
Evidence stream
On-site outcomes
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
O2
Provenance
AI Visibility Evidence Register v0.1
Origin
master
Corrections applied
None
S090Decision: IncludeReview: FullEvidence: On-site outcomesB2B: Indirect / transfer

Ahrefs Web Analytics engagement analysis - same 81,947-site panel as T3

Ahrefs / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

81,947 sites; May-June 2025; AI referrals versus search and all traffic

Key result

AI visitors averaged 4.0 pages versus 5.2 for search; 86 seconds versus 78; and 67.8% bounce versus 63.7% for search and 62.4% overall

Limitation

Cross-site means conceal mix and landing-page differences; conversion not measured. Directly contradicts a universal “AI visitors are higher quality” claim. Same dataset as T3.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
On-site engagement, conversion and revenue
Evidence stream
On-site outcomes
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
T3
Provenance
AI Visibility Evidence Register v0.1
Origin
master
Corrections applied
None
S091Decision: IncludeReview: Landing-pageEvidence: On-site outcomesB2B: Direct B2B

Ahrefs.com first-party SaaS case

Ahrefs.com / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

Ahrefs.com; 30 days ending 16 June 2025

Key result

AI referrals were 0.5% of visits but generated 12.1% of signups, a reported 23-times higher signup-per-visit ratio than traditional organic search

Limitation

One brand, no absolute counts, signup rather than paid revenue, likely high prior brand/product awareness and power-user selection. Do not generalise 23x.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
On-site engagement, conversion and revenue
Evidence stream
On-site outcomes
B2B relevance
Direct B2B
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Low
Dataset family
O5
Provenance
AI Visibility Evidence Register v0.1
Origin
master
Corrections applied
None
S092Decision: IncludeReview: AbstractEvidence: On-site outcomesB2B: Indirect / transfer

Similarweb ecommerce referral estimates

Similarweb / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

Ecommerce sites; June 2025

Key result

AI-referral visits reportedly converted at 11.4%, versus 9.3% paid search and 5.3% organic search

Limitation

Public article gives insufficient sampling, site, geography and model detail. Referral only and likely part of Similarweb's State of Ecommerce/rolling panel.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
On-site engagement, conversion and revenue
Evidence stream
On-site outcomes
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Low
Dataset family
O6
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official Similarweb page
Origin
master
Corrections applied
None
S093Decision: IncludeReview: FullEvidence: organic ranking and AI citation overlapB2B: High for integrated SEO/AEO strategy.

76% of AI Overview Citations Pull from the Top 10

Ahrefs / 2025 / Industry / practice research

Open direct source
Sample or setting

1.9M AI Overview citations matched to Google organic search positions in Ahrefs data.

Key result

76% of AI Overview citations came from pages in Google's top ten for the associated search.

Limitation

AIO-specific and does not show causality or ChatGPT behavior.

Why this decision

Large direct overlap measure supports SEO as a foundation; AIO-specific and does not show causality or ChatGPT behavior.

Source classification and provenance
Study design
Not stated
Evidence stream
organic ranking and AI citation overlap
B2B relevance
High for integrated SEO/AEO strategy.
Source type
full vendor SERP/citation study
Source class
Industry / original-data report
Transparency
Moderate-high - full article and citation n; query sampling/geography need caution.
Dataset family
AHREFS_AIO_CITATION_RANK_OVERLAP_2025
Provenance
Official Ahrefs data-study search
Origin
master
Corrections applied
None
S094Decision: IncludeReview: AbstractEvidence: Outbound clicks and search displacementB2B: High

Answering Without Referring: How AI Search Rewrites the Web's Economic Bargain

Qiaoni Shi, Kai Zhu, Kai Gu; Bocconi University / 2026-07 / Industry / practice research

Open direct source
Sample or setting

URL-level Comscore US desktop clickstream comparing ChatGPT and Google, with access expansions used for causal estimation

Key result

Only 5.2% of ChatGPT conversation sessions produced an outbound click, and wider ChatGPT Search access reduced traditional search use by 9.4%, with largest referral losses in informational categories.

Limitation

US desktop clickstream and staggered access shocks may not transfer to mobile, later product versions or B2B revenue journeys.

Why this decision

Rare clickstream and quasi-experimental evidence directly measuring the referral gap.

Source classification and provenance
Study design
Not stated
Evidence stream
Outbound clicks and search displacement
B2B relevance
High
Source type
Working paper
Source class
Academic
Transparency
Medium
Dataset family
Comscore ChatGPT referral study
Provenance
SSRN forward search on AI search traffic
Origin
master
Corrections applied
None
S095Decision: IncludeReview: FullEvidence: platform adoption and use casesB2B: High for work/decision-support context; indirect for B2B purchasing.

How People Use ChatGPT

OpenAI Economic Research, Duke and Harvard researchers / 2025-09-15 / Industry / practice research

Open direct source
Sample or setting

Privacy-preserving automated classification of a representative sample of consumer ChatGPT conversations through July 2025; product growth tracked from launch; Harvard IRB approval.

Key result

Practical Guidance, Seeking Information and Writing comprised nearly 80% of conversations; decision support was a major economic-use mechanism.

Limitation

consumer ChatGPT messages do not identify brand visibility or buying outcomes.

Why this decision

Unusually strong first-party platform-use evidence and downloadable methodology; consumer ChatGPT messages do not identify brand visibility or buying outcomes.

Source classification and provenance
Study design
Not stated
Evidence stream
platform adoption and use cases
B2B relevance
High for work/decision-support context; indirect for B2B purchasing.
Source type
full first-party research paper and downloadable data
Source class
Industry / original-data report
Transparency
High - full paper, IRB, methodology and downloadable aggregate data; OpenAI is the platform owner.
Dataset family
OPENAI_CHATGPT_USAGE_2025
Provenance
Official OpenAI Signals and direct paper; JY Scauri review: retained source 13 of 34.
Origin
master
Corrections applied
None
S096Decision: IncludeReview: AbstractEvidence: Post-retrieval ranking manipulationB2B: Medium

Ranking Manipulation for Conversational Search Engines

Samuel Pfrommer et al.; UC Berkeley / 2024-11 / Academic research

Open direct source
Sample or setting

Controlled ranking attacks plus Perplexity validation using explicitly supplied URLs

Key result

Injected ranking instructions can influence conversational search ordering and transferred to Perplexity when target URLs were directly provided.

Limitation

Shows a real attack surface but does not demonstrate organic crawling or retrieval.

Why this decision

Shows a real attack surface but does not demonstrate organic crawling or retrieval.

Source classification and provenance
Study design
Not stated
Evidence stream
Post-retrieval ranking manipulation
B2B relevance
Medium
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
Conversational ranking injection
Provenance
Martinez 2026 core matrix; ACL Anthology; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S097Decision: IncludeReview: AbstractEvidence: Product recommendation manipulationB2B: High

Manipulating Large Language Models to Increase Product Visibility

Aounon Kumar, Himabindu Lakkaraju; Harvard University / 2024-04 / Academic research

Open direct source
Sample or setting

Controlled experiments on fictitious product catalogs with strategic text sequences

Key result

Strategic additions to product text can move items upward in LLM-generated recommendations, but the result was produced in a synthetic catalog rather than a live search product.

Limitation

Directly relevant product-visibility mechanism with limited external validity.

Why this decision

Directly relevant product-visibility mechanism with limited external validity.

Source classification and provenance
Study design
Not stated
Evidence stream
Product recommendation manipulation
B2B relevance
High
Source type
Preprint
Source class
Academic
Transparency
High
Dataset family
Synthetic product visibility manipulation
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix; also surfaced in Cyrus Shepard evidence sheet.
Origin
master
Corrections applied
None
S098Decision: IncludeReview: AbstractEvidence: Product recommendation manipulationB2B: High

Bias Beware: The Impact of Cognitive Biases on LLM-Driven Product Recommendations

Giorgos Filandrianos et al.; National Technical University of Athens / 2025-11 / Academic research

Open direct source
Sample or setting

Product-description interventions based on cognitive-bias cues across LLMs of several sizes

Key result

Social-proof language consistently increased recommendation rate and rank, whereas scarcity and exclusivity cues often reduced visibility, showing that familiar persuasion tactics transfer unpredictably.

Limitation

Prompted product descriptions and model outputs do not establish organic web retrieval, durable visibility or actual purchases.

Why this decision

Direct controlled product-visibility evidence with clear implications for content experiments.

Source classification and provenance
Study design
Not stated
Evidence stream
Product recommendation manipulation
B2B relevance
High
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
Cognitive-bias product benchmark
Provenance
ACL Anthology search for LLM product recommendation bias
Origin
master
Corrections applied
None
S099Decision: IncludeReview: AbstractEvidence: Product-ranking manipulationB2B: High

Controlling Output Rankings in Generative Engines for LLM-Based Search

Haibo Jin et al.; Penn State and collaborators / 2026-02 / Academic research

Open direct source
Sample or setting

ProductBench with 15 categories and 200 products each; top-ten Amazon lists; four search-capable LLMs

Key result

CORE reported promotion success of 91.4% at top five and 80.3% at top one, but each model received fixed top-ten product lists and some review-based variants may introduce fabricated content.

Limitation

Large direct product-ranking experiment with unusually high effects that require fixed-context and integrity caveats.

Why this decision

Large direct product-ranking experiment with unusually high effects that require fixed-context and integrity caveats.

Source classification and provenance
Study design
Not stated
Evidence stream
Product-ranking manipulation
B2B relevance
High
Source type
Preprint
Source class
Academic
Transparency
High
Dataset family
ProductBench CORE
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S100Decision: IncludeReview: FullEvidence: Publisher economicsB2B: Indirect / transfer

Cloudflare network logs; crawl-to-refer ratio

Cloudflare / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

Worldwide Cloudflare-served HTML requests; January-July 2025

Key result

July crawl requests per attributed referral: Anthropic 38,065.7:1, OpenAI 1,091.4:1, Perplexity 194.8:1, Microsoft 40.7:1, Google 5.4:1

Limitation

Requests are not people, sessions, use or value. Ratios are unstable when referrals are sparse and bot identity is imperfect. Supply-side publisher economics must never be presented as buyer-influence evidence.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Supply-side crawling and the click bargain
Evidence stream
Publisher economics
B2B relevance
Indirect / transfer
Source type
Independent / platform
Source class
Baseline evidence register
Transparency
High
Dataset family
S1
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official Cloudflare page
Origin
master
Corrections applied
None
S101Decision: IncludeReview: FullEvidence: publisher referrals and conversionB2B: Medium-high - digital conversion mechanism, not software buying.

AI Traffic Converts at 3x the Rate of Other Channels

Microsoft Clarity / 2025-11-06 / Industry / practice research

Open direct source
Sample or setting

1,277 publisher/news domains in Microsoft Clarity; eight months of traffic growth and one month of Smart Events conversion analysis.

Key result

AI referrals were below 1% of traffic but grew 155.6%; sign-up CTR was 1.66% versus search 0.15%, and 52% of domains recorded an AI conversion.

Limitation

publisher-heavy selection and automatic Smart Events differ from SaaS pipeline.

Why this decision

Large cross-domain first-party conversion study; publisher-heavy selection and automatic Smart Events differ from SaaS pipeline.

Source classification and provenance
Study design
Not stated
Evidence stream
publisher referrals and conversion
B2B relevance
Medium-high - digital conversion mechanism, not software buying.
Source type
full first-party platform analytics study
Source class
Industry / original-data report
Transparency
High-moderate - full article, domain count and rates; traffic/session totals and weighting absent.
Dataset family
MICROSOFT_CLARITY_PUBLISHERS_2025
Provenance
Official Microsoft Clarity search
Origin
master
Corrections applied
None
S102Decision: IncludeReview: AbstractEvidence: Publisher trafficB2B: Medium

Impact of AI Search Summaries on Website Traffic: Evidence from Google AI Overviews and Wikipedia

Mehrzad Khosravi, Hema Yoganarasimhan; University of Washington / 2026-02 / Industry / practice research

Open direct source
Sample or setting

Difference-in-differences across 161,382 matched Wikipedia article-language pairs using staggered AIO rollout

Key result

AIO exposure reduced daily English Wikipedia article traffic by about 15%, with larger relative declines in culture than STEM content.

Limitation

Wikipedia is a distinctive publisher and the staggered language comparison depends on parallel-trends assumptions; results are not a universal site effect.

Why this decision

Strong quasi-experimental downstream traffic evidence with a transparent comparison design.

Source classification and provenance
Study design
Not stated
Evidence stream
Publisher traffic
B2B relevance
Medium
Source type
Working paper
Source class
Academic
Transparency
High
Dataset family
Wikipedia AIO natural experiment
Provenance
SSRN search for AI Overviews publisher traffic
Origin
master
Corrections applied
None
S103Decision: IncludeReview: FullEvidence: Query decomposition and retrievalB2B: High

How AI Platforms Search: Fan-Out Query Behavior Across Intent Types, Verticals, and Platforms

Anthony Lee; independent researcher / 2026-04 / Industry / practice research

Open direct source
Sample or setting

1,323 fan-out queries from 540 parent prompts across ChatGPT, Gemini and Perplexity, ten verticals and five intents

Key result

Intent and platform predicted fan-out type more than vertical; discovery prompts triggered 3.3 times more entity injection than informational prompts, while exact query strings were highly stochastic.

Limitation

Directly measures the hidden retrieval queries that determine discoverability, with limited per-vertical power and independent provenance.

Why this decision

Directly measures the hidden retrieval queries that determine discoverability, with limited per-vertical power and independent provenance.

Source classification and provenance
Study design
Not stated
Evidence stream
Query decomposition and retrieval
B2B relevance
High
Source type
Working paper
Source class
Academic
Transparency
Medium
Dataset family
Cross-platform fan-out audit
Provenance
SSRN search for AI search retrieval behavior
Origin
master
Corrections applied
None
S104Decision: IncludeReview: AbstractEvidence: Recommendations and purchase valueB2B: High

The Price of Advice: Experimental Evidence on the Effects of AI Recommenders

Amit Zac, Michal Gal; ETH Zurich, University of Amsterdam, University of Haifa / 2025-10 / Industry / practice research

Open direct source
Sample or setting

Laboratory purchase experiment assigning traditional search, GPT, Gemini or a steering GPT, plus large API audits

Key result

Conversational recommenders increased consumer spending, with the steering GPT highest; the effect reflected linguistic framing and premium-brand exposure rather than perceived quality or generalized trust.

Limitation

Laboratory purchases and an intentionally steering model may overstate effects relative to natural, high-stakes B2B evaluation.

Why this decision

Rare controlled downstream purchasing evidence that connects model recommendations to transaction value.

Source classification and provenance
Study design
Not stated
Evidence stream
Recommendations and purchase value
B2B relevance
High
Source type
Working paper
Source class
Academic
Transparency
High
Dataset family
AI recommender purchase experiment
Provenance
SSRN search for LLM recommendation experiments
Origin
master
Corrections applied
None
S105Decision: IncludeReview: FullEvidence: recommendations, choice and zero-click decisionsB2B: High as decision mechanism; consumer task categories.

The Shortlist Is the New Shelf

Profound, Kevin Indig and Clickstream Solutions / 2026-07-14 / Industry / practice research

Open direct source
Sample or setting

56 participants completed 221 real shopping tasks in ChatGPT; behavior/transcripts coded and matched to Profound share-of-voice data.

Key result

Chosen brands had 24% answer share versus 11% for passed-over brands; choice correlation was 0.57 and 92.8% of tasks ended without an open-web click.

Limitation

small consumer sample, task prompting and correlation prevent broad causality.

Why this decision

Rare direct observed link between AI visibility and choice; small consumer sample, task prompting and correlation prevent broad causality.

Source classification and provenance
Study design
Not stated
Evidence stream
recommendations, choice and zero-click decisions
B2B relevance
High as decision mechanism; consumer task categories.
Source type
full behavioral study article
Source class
Industry / original-data report
Transparency
High-moderate - full method and appendix-level bases; participant recruitment/category cells are small.
Dataset family
PROFOUND_KEVIN_CLICKSTREAM_SHORTLIST_2026
Provenance
Official Profound research hub and full-page inspection
Origin
master
Corrections applied
None
S106Decision: IncludeReview: AbstractEvidence: Referral concentration and crawler blockingB2B: High

LLMs as Gatekeepers: Source Concentration, Factual Quality, and Political Slant in Information Search

Pengxiang Zhou, Davide Proserpio, Ali Goli; USC, University of Rochester / 2026-06 / Industry / practice research

Open direct source
Sample or setting

48 months of referral traffic for 8,082 news domains across five AI platforms and organic search, linked to bot-blocking records

Key result

AI referrals were more concentrated than organic search; blocking Perplexity's crawler reduced the blocking domain's referral traffic by 46-48% and redistributed traffic mainly to already large accessible outlets.

Limitation

News-domain referral markets and crawler blocking are not B2B discovery; blocking decisions may also correlate with unobserved publisher strategy.

Why this decision

Rare multi-year referral and crawler-access evidence with direct publisher-market implications.

Source classification and provenance
Study design
Not stated
Evidence stream
Referral concentration and crawler blocking
B2B relevance
High
Source type
Working paper
Source class
Academic
Transparency
High
Dataset family
News referral and crawler-blocking panel
Provenance
SSRN search for AI referral concentration
Origin
master
Corrections applied
None
S107Decision: IncludeReview: FullEvidence: referral growth, complexity and conversionB2B: High as complexity-transfer support.

Q2 2025 Insights: AI Referrals Surge Across Industries

Adobe Digital Insights / 2025 / Industry / practice research

Open direct source
Sample or setting

Adobe US retail and travel analytics; July 2024-May 2025; proprietary customer transaction base.

Key result

By May 2025 AI retail conversion was 22% below non-AI, while consumer-electronics AI traffic share was 4x apparel and home goods 3x apparel.

Limitation

overlaps its rolling dataset and remains non-B2B.

Why this decision

Strongest Adobe wave for the considered-purchase mechanism; overlaps its rolling dataset and remains non-B2B.

Source classification and provenance
Study design
Not stated
Evidence stream
referral growth, complexity and conversion
B2B relevance
High as complexity-transfer support.
Source type
full first-party analytics article
Source class
Industry / original-data report
Transparency
Moderate - full article reports time series and categories; exact site/sample denominators are not restated.
Dataset family
ADOBE_DIGITAL_INSIGHTS_ROLLING_2024_2026
Provenance
Official Adobe search; rolling-family lineage audit
Origin
master
Corrections applied
None
S108Decision: IncludeReview: FullEvidence: referrals, conversion and revenue per visitB2B: High for technology engagement context; consumer and last-click outcomes.

AI-Driven Traffic Surges Across Industries with Retail Experiencing Biggest Gains

Adobe Digital Insights / 2026-01-12 / Industry / practice research

Open direct source
Sample or setting

More than 1T US retail visits; October 2024-December 2025; companion survey of 1,000+ US consumers.

Key result

Holiday 2025 AI referrals converted 31% better than non-AI traffic; technology/software had the highest AI-driven visit share, up 120% year over year.

Limitation

overlaps the Adobe rolling panel and is not an independent replication.

Why this decision

Material earlier wave and cross-industry cut; overlaps the Adobe rolling panel and is not an independent replication.

Source classification and provenance
Study design
Not stated
Evidence stream
referrals, conversion and revenue per visit
B2B relevance
High for technology engagement context; consumer and last-click outcomes.
Source type
full first-party analytics and survey article
Source class
Industry / original-data report
Transparency
Moderate-high - full article and date range; exact site and rate denominators remain proprietary.
Dataset family
ADOBE_DIGITAL_INSIGHTS_ROLLING_2024_2026
Provenance
Official Adobe research search; lineage matched to existing register
Origin
master
Corrections applied
None
S109Decision: IncludeReview: AbstractEvidence: Search and purchase behaviorB2B: High

Reasoning AI, Consumer Search, and Purchase: Evidence from an Online Platform Field Experiment

Se Yan et al.; Peking University, University of Toronto, Zhejiang University / 2025-10 / Industry / practice research

Open direct source
Sample or setting

More than 510,000 travel-platform users randomized between DeepSeek-R1 and DeepSeek-V3 assistants

Key result

The reasoning assistant reduced hotel bookings by 2.5%, alongside fewer searches, hotel views and clicks, while increasing later chat engagement.

Limitation

One travel platform and two DeepSeek assistants make the direction and magnitude highly context-specific; hotel bookings are not B2B pipeline.

Why this decision

Large randomized commercial field test showing that better AI assistance can narrow search and reduce platform purchases.

Source classification and provenance
Study design
Not stated
Evidence stream
Search and purchase behavior
B2B relevance
High
Source type
Working paper
Source class
Academic
Transparency
High
Dataset family
Travel reasoning-assistant experiment
Provenance
SSRN search for AI consumer search and purchase
Origin
master
Corrections applied
None
S110Decision: IncludeReview: AbstractEvidence: Search substitution and publisher trafficB2B: High

The Impact of LLM Adoption on Online User Behavior

Nicolas Padilla, H. Tai Lam, Anja Lambrecht, Brett Hollenbeck; UCLA, London Business School / 2025-12 / Industry / practice research

Open direct source
Sample or setting

Detailed 2022-2023 clickstream following individual LLM adoption

Key result

Traditional search eventually fell more than 20% below adopters' pre-use level; small websites, education sites, Stack Overflow and display-ad exposure declined while frequently visited sites were less affected.

Limitation

Strong individual-level adoption design with downstream site effects, though it reflects an early ChatGPT period.

Why this decision

Strong individual-level adoption design with downstream site effects, though it reflects an early ChatGPT period.

Source classification and provenance
Study design
Not stated
Evidence stream
Search substitution and publisher traffic
B2B relevance
High
Source type
Working paper
Source class
Academic
Transparency
High
Dataset family
Early-adopter clickstream panel
Provenance
Marketing Science references and SSRN
Origin
master
Corrections applied
None
S111Decision: IncludeReview: AbstractEvidence: Search substitution and web explorationB2B: High

Beyond Search: LLM Adoption and Web Traffic Concentration

Samira Gholami, Cristiana Firullo, Cristobal Cheyre, Alessandro Acquisti; Stanford, Cornell, Carnegie Mellon / 2026-02 / Industry / practice research

Open direct source
Sample or setting

Nationally representative Comscore panel data from 2019-2024 plus a session-level field experiment

Key result

LLM use was usually complementary to web exploration rather than LLM-only, and adopters visited more unique sites, complicating the simple zero-click displacement narrative.

Limitation

Most behavioral data end in 2024 and capture early standalone-LLM adoption, before newer integrated AI-search interfaces.

Why this decision

Large behavioral panel plus field evidence that tests substitution and concentration directly.

Source classification and provenance
Study design
Not stated
Evidence stream
Search substitution and web exploration
B2B relevance
High
Source type
Working paper
Source class
Academic
Transparency
Medium
Dataset family
Comscore plus field experiment
Provenance
SSRN search for LLM adoption and traffic
Origin
master
Corrections applied
None
S112Decision: IncludeReview: FullEvidence: search-to-AI citation overlapB2B: High for SEO-to-GEO transfer assumptions.

ChatGPT May Scrape Google, but the Results Don't Match

Ahrefs / 2025-09-03 / Industry / practice research

Open direct source
Sample or setting

3,311 short-tail keywords across four intents run through ChatGPT, Perplexity and Google's top 100, building on long-tail/fan-out experiments.

Key result

Citation overlap with Google remained low even when AI systems retrieved from search indexes, implying an additional selection layer.

Limitation

one keyword sample and changing retrieval stack limit durability.

Why this decision

Well-defined cross-platform/query design; one keyword sample and changing retrieval stack limit durability.

Source classification and provenance
Study design
Not stated
Evidence stream
search-to-AI citation overlap
B2B relevance
High for SEO-to-GEO transfer assumptions.
Source type
full vendor retrieval-overlap study
Source class
Industry / original-data report
Transparency
High-moderate - full methods and n; exact query/citation outputs not downloadable.
Dataset family
AHREFS_BRAND_RADAR_SEARCH_CITATION_OVERLAP_2025
Provenance
Official Ahrefs search and full-page inspection
Origin
master
Corrections applied
None
S113Decision: IncludeReview: FullEvidence: Shortlist and decisionB2B: Direct B2B

G2 2026 AI Search Insight survey

G2 / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

1,076 global B2B software decision-makers and influencers surveyed March 2026

Key result

51% started software research with an AI chatbot more often than Google; 71% used chatbots in the process; 54% cited AI as a shortlist influence. On vendor selection: 33% chose a previously unconsidered vendor and 36% chose a familiar but different vendor.

Limitation

Self-report and survey recruitment details constrain causal interpretation. The 69% combines two response categories and means a different vendor than initially planned, not AI-caused revenue.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
B2B buyer research
Evidence stream
Shortlist and decision
B2B relevance
Direct B2B
Source type
Commercial marketplace
Source class
Baseline evidence register
Transparency
Medium
Dataset family
G2-B2B-2026
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official G2 report page
Origin
master
Corrections applied
None
S114Decision: IncludeReview: FullEvidence: Shortlist and decisionB2B: Direct B2B

Semrush survey of B2B professionals

Semrush / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

643 US B2B professionals surveyed Mar-Apr 2026; 21 quality failures removed; 622 valid; subsequent published results use 519 respondents who use AI at work

Key result

Among the 519 AI-at-work users: 66% regularly used AI for vendor research, 92% said it influenced the shortlist, 83% said it influenced the final decision, and 32% said the influence was major.

Limitation

Question wording, response scale and raw numerators are not public. Published percentages condition on AI users; approximately 69% of all 622 is a rounded reconstruction, not a directly reported result.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
B2B buyer research
Evidence stream
Shortlist and decision
B2B relevance
Direct B2B
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
SEMRUSH-B2B-2026
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official Semrush study page
Origin
master
Corrections applied
None
S115Decision: IncludeReview: FullEvidence: Shortlist and pipeline mechanismB2B: Direct B2B

6sense Buyer Experience Report 2025

6sense / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

Main study: 3,744 B2B buyers across North America, Europe, APAC and UK/Ireland; purchases at least $25k. Companion: 766 respondents. Software median purchase roughly $200k-$300k.

Key result

94% used LLMs somewhere in the journey. Buyers considered 5.1 vendors and had 3.6 on the Day One shortlist; 95% bought from that shortlist, and the preferred pre-contact vendor ultimately won 77% of the time.

Limitation

Does not observe AI visibility experimentally. Strong buying-mechanism bridge, not proof that an AI mention caused shortlist or revenue.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
B2B buyer research
Evidence stream
Shortlist and pipeline mechanism
B2B relevance
Direct B2B
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
High
Dataset family
6SENSE-BUYER-2025
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official 6sense report
Origin
master
Corrections applied
None
S116Decision: IncludeReview: AbstractEvidence: Source credibility and groundednessB2B: Medium

Assessing Web Search Credibility and Response Groundedness in Chat Assistants

Ivan Vykopal et al.; Kempelen Institute and collaborators / 2026-03 / Academic research

Open direct source
Sample or setting

100 claims across five misinformation-prone topics tested on GPT-4o, GPT-5, Perplexity and Qwen Chat

Key result

Perplexity used the most credible sources in this restricted test, while GPT-4o cited more non-credible sources on sensitive topics and groundedness varied separately from credibility.

Limitation

Direct modern commercial-assistant audit, although restricted topics limit generalization.

Why this decision

Direct modern commercial-assistant audit, although restricted topics limit generalization.

Source classification and provenance
Study design
Not stated
Evidence stream
Source credibility and groundedness
B2B relevance
Medium
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
Chat-assistant credibility audit
Provenance
Martinez 2026 core matrix; ACL Anthology; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S117Decision: IncludeReview: FullEvidence: Source identity and citation preferenceB2B: High

Media Source Matters More Than Content: Unveiling Political Bias in LLM-Generated Citations

Sunhao Dai et al.; Renmin University, USTC, Chinese Academy of Sciences, NUS / 2025-11 / Academic research

Open direct source
Sample or setting

AllSides-2024 with 1,340 queries, 2,680 paired passages and 169 media outlets; controlled source-name swaps

Key result

LLMs cited left-labelled outlets more than traditional retrievers, and swapping publisher names nearly reversed the bias, showing source identity can outweigh matched content.

Limitation

Political passage pairs isolate source-name effects but do not test commercial recommendations, live retrieval or buyers.

Why this decision

Strong controlled evidence that publisher brand itself can drive citation selection.

Source classification and provenance
Study design
Not stated
Evidence stream
Source identity and citation preference
B2B relevance
High
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
AllSides-2024 citation bias
Provenance
ACL Anthology search for citation source bias
Origin
master
Corrections applied
None
S118Decision: IncludeReview: AbstractEvidence: Source influence and evidence selectionB2B: Medium

What Evidence Do Language Models Find Convincing?

Alexander Wan, Eric Wallace, Dan Klein; UC Berkeley / 2024-08 / Academic research

Open direct source
Sample or setting

Controlled ConflictingQA contexts testing evidence attributes and position across language models

Key result

Topical relevance and context position dominated many stylistic cues in determining which conflicting evidence a model adopted.

Limitation

Uses controlled conflicting evidence already placed in context; it does not test organic retrieval, live answer engines or user behavior.

Why this decision

Strong mechanism evidence for why relevance and placement matter once content is retrieved.

Source classification and provenance
Study design
Not stated
Evidence stream
Source influence and evidence selection
B2B relevance
Medium
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
ConflictingQA evidence influence
Provenance
Martinez 2026 core matrix; ACL Anthology; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S119Decision: IncludeReview: AbstractEvidence: Source influence benchmarkB2B: Medium

CC-GSEO-Bench: A Content-Centric Benchmark for Measuring Source Influence in Generative Search Engines

Qiyuan Chen et al.; Zhejiang University and collaborators / 2025-09 / Academic research

Open direct source
Sample or setting

Retrieved web contexts, LLM-generated article-centric queries and predominantly automated evaluation

Key result

The benchmark measures how source content shapes generated answers, but heavy use of generated queries and automated judges leaves open whether gains reflect real user demand or human-perceived quality.

Limitation

Direct source-influence benchmark, useful with explicit judge and query-generation caveats.

Why this decision

Direct source-influence benchmark, useful with explicit judge and query-generation caveats.

Source classification and provenance
Study design
Not stated
Evidence stream
Source influence benchmark
B2B relevance
Medium
Source type
Preprint
Source class
Academic
Transparency
High
Dataset family
CC-GSEO-Bench
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix; overlap with Cyrus Shepard 54-source sheet.
Origin
master
Corrections applied
None
S120Decision: IncludeReview: AbstractEvidence: Source quality and synthetic contentB2B: Medium

Synthetic Sources? Auditing Generative Search Engine Citations for Evidence of AI-Generated Sources

Mowafak Allaham, Nicholas Diakopoulos; Northwestern University / 2026-05 / Academic research

Open direct source
Sample or setting

712 real-world queries across politics, health and environment on ChatGPT, Copilot, Gemini and Perplexity; 19,154 retrieved classified textual pages

Key result

Roughly 16% of classified cited pages showed evidence of AI generation across all four engines, while source use combined a narrow repeatedly cited core with a long tail.

Limitation

Direct cross-engine source-quality audit; detector error and inaccessible URLs require sensitivity treatment.

Why this decision

Direct cross-engine source-quality audit; detector error and inaccessible URLs require sensitivity treatment.

Source classification and provenance
Study design
Not stated
Evidence stream
Source quality and synthetic content
B2B relevance
Medium
Source type
Preprint
Source class
Academic
Transparency
High
Dataset family
Synthetic-source GSE audit
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S121Decision: IncludeReview: FullEvidence: TrafficB2B: Indirect / transfer

Ahrefs Web Analytics rolling cross-site panel

Ahrefs / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

81,947 sites; May-June 2025; referrers from ChatGPT, Perplexity, Copilot and Gemini

Key result

Average AI traffic was 0.25% of site traffic; the three largest assistants were about 0.19%; Google sent about 210 times as much traffic as those three; ChatGPT supplied more than 80% of AI traffic

Limitation

Voluntary Ahrefs analytics customers, site-level average unclear, no B2B weighting. The earlier 3,000- and 35,000-site articles are earlier waves of this same rolling dataset, not independent replications.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Attributable AI referrals and channel share
Evidence stream
Traffic
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
T3
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official Ahrefs study page
Origin
master
Corrections applied
None
S122Decision: IncludeReview: FullEvidence: TrafficB2B: Direct B2B

Conductor 2026 AEO/GEO Benchmarks; anonymised enterprise-customer analytics

Conductor / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

1,215 US enterprise-customer domains; 3.3B sessions, including 35.7M LLM/chatbot sessions; May-September 2025

Key result

AI referrals were 1.08% of all sessions. Information Technology averaged 2.80%; software and services 3.37%; semiconductors 4.09%; hardware 0.99%. ChatGPT supplied 87.4% of AI referrals overall and 88.5% in IT

Limitation

Enterprise-customer selection, older 2025 observation window, and referral-only measurement. The report also contains a separate synthetic-prompt/citation dataset; do not merge its denominator with traffic sessions.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Attributable AI referrals and channel share
Evidence stream
Traffic
B2B relevance
Direct B2B
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
CONDUCTOR-TRAFFIC
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official report/PDF verification; JY Scauri review: retained source 14 of 34; already represented in project baseline lineage.
Origin
master
Corrections applied
None
S123Decision: IncludeReview: FullEvidence: TrafficB2B: Indirect / transfer

Peer-reviewed Marketing Science analysis using first-party Google Analytics supplied by Grips Intelligence

Peer-reviewed / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

973 ecommerce websites, 24 categories and all continents; 10.51B sessions, 164.9M transactions and $20.6B revenue over 28 July 2024-2 August 2025; main model uses six months ending 2 August 2025

Key result

ChatGPT referrals were under 0.2% of traffic, about 200 times smaller than organic Google. The 12-month data contained 4.92M ChatGPT-referral sessions and 50,251 transactions

Limitation

Ecommerce and last-click only; selection into Grips customer data; the first author worked for Grips. Still the strongest directly observed cross-site transaction dataset. Independent of the vendor panels above as far as public disclosures show.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Attributable AI referrals and channel share
Evidence stream
Traffic
B2B relevance
Indirect / transfer
Source type
Independent / platform
Source class
Baseline evidence register
Transparency
High
Dataset family
T2
Provenance
AI Visibility Evidence Register v0.1; Existing project source; Marketing Science primary article; Existing evidence register; publisher DOI page
Origin
master
Corrections applied
None
S124Decision: IncludeReview: FullEvidence: TrafficB2B: Indirect / transfer

Similarweb top-1,000-site referral estimates

Similarweb / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

Worldwide top 1,000 sites; June 2025

Key result

AI platforms generated an estimated 1.13B referral visits, +357% YoY, versus 191B referrals from Google search; ChatGPT supplied more than 80% of AI referrals

Limitation

Concentrated in the world's largest sites, so unsuitable as a typical-site share. Same rolling Similarweb referral family as the 2025 Generative AI Report and related “top sites” articles.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Attributable AI referrals and channel share
Evidence stream
Traffic
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
SIMILARWEB-REFERRALS
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official Similarweb article
Origin
master
Corrections applied
None
S125Decision: IncludeReview: FullEvidence: Traffic disruptionB2B: Indirect / transfer

Ahrefs matched keyword/GSC analysis

Ahrefs / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

300,000 keywords: 150,000 with an AI Overview in December 2025 and 150,000 informational controls without one; aggregated desktop GSC CTR; December 2023 versus December 2025

Key result

Position-one CTR fell from 0.073 to 0.016 for the later-AIO cohort. After applying the control trend, observed CTR was about 58% below the forecast no-AIO CTR. Estimated impact declined from -58% at position one to -19.4% at position ten

Limitation

Matched observational cohorts, not randomised; geography/language not clearly disclosed; post-period AIO status defines the cohort. The earlier 34.5% article is the first wave of this same study, not a replication.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Click suppression and zero-click behaviour
Evidence stream
Traffic disruption
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
Z2
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official Ahrefs article; JY Scauri review: retained source 3 of 34; also present in existing project evidence register.
Origin
master
Corrections applied
None
S126Decision: IncludeReview: FullEvidence: Traffic disruptionB2B: Indirect / transfer

Pew Research Center metered browsing panel

Pew / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

900 consenting US adults; browser activity on personal desktop, laptop and mobile; 2,457,176 page visits during March 2025

Key result

AI summaries appeared on 18% of Google searches. Users clicked a traditional result on 8% of visits with an AI summary versus 15% without one; only 1% clicked a link inside the summary; 26% ended the session versus 16% without a summary

Limitation

Observational: AI summaries appear on systematically different queries, especially longer ones. The 8/15 comparison is per Google-result page visit, not a keyword-position CTR. Independent primary panel.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Click suppression and zero-click behaviour
Evidence stream
Traffic disruption
B2B relevance
Indirect / transfer
Source type
Independent / platform
Source class
Baseline evidence register
Transparency
High
Dataset family
Z1
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official Pew page
Origin
master
Corrections applied
None
S127Decision: IncludeReview: FullEvidence: Traffic disruptionB2B: Indirect / transfer

Semrush/Datos early AI Mode clickstream study

Semrush/Datos / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

Nearly 69M US desktop Google sessions; 1 May-5 July 2025

Key result

AI Mode rose from 0.25% to just over 1% of Google sessions. Only 6-8% of AI Mode sessions reached an external domain, implying 92-94% zero-click

Limitation

Early-adopter/post-launch period, desktop only and session definitions may differ from normal Google search. Same Semrush/Datos clickstream family.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Click suppression and zero-click behaviour
Evidence stream
Traffic disruption
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
DATOS-CLICKSTREAM
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official Semrush page
Origin
master
Corrections applied
None
S128Decision: IncludeReview: FullEvidence: Traffic disruptionB2B: Indirect / transfer

SparkToro analysis of Similarweb clickstream

SparkToro / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

US Google searches; desktop plus mobile browser, excluding the Google mobile app; January-April 2026

Key result

68.01% of Google searches ended without any click, versus 60.45% in the authors' 2024 analysis; AI Mode itself represented only 0.34% of searches

Limitation

Similarweb panel; mobile session endings use a ten-second inactivity rule and mobile paid-click share is modelled. Zero-click predates generative AI, so the 7.56-point increase cannot be assigned causally to AI. Not independent of Similarweb.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Click suppression and zero-click behaviour
Evidence stream
Traffic disruption
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
SIMILARWEB-CLICKSTREAM
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official partner analysis
Origin
master
Corrections applied
None
S129Decision: IncludeReview: Landing-pageEvidence: Traffic/clicksB2B: High

Google AI Overviews: New Research Reveals How to Navigate Click Drop-Off

Amsive / 2025 / Industry / practice research

Open direct source
Sample or setting

Approximately 700,000 keywords across ten sites and five industries, comparing AIO and non-AIO click behavior.

Key result

Average organic click-through fell 15.49% in the sample, while branded queries rose 18.68%, revealing strong query-type heterogeneity.

Limitation

Useful counterexample to a universal click-loss claim; small site cohort and observational design limit causal inference.

Why this decision

Useful counterexample to a universal click-loss claim; small site cohort and observational design limit causal inference. JY source 16.

Source classification and provenance
Study design
Not stated
Evidence stream
Traffic/clicks
B2B relevance
High
Source type
industry observational study
Source class
Prior-review mined
Transparency
Moderate
Dataset family
AMSIVE-AIO-CTR-2025
Provenance
JY Scauri review: retained source 16 of 34.
Origin
master
Corrections applied
None
S130Decision: IncludeReview: Landing-pageEvidence: Traffic/clicksB2B: High

AI Overviews at the One-Year Mark: Presence, Size, and What They're Citing

BrightEdge / 2026 / Industry / practice research

Open direct source
Sample or setting

Twelve months of AI Overview presence and size plus citation overlap across nine verticals using BrightEdge Generative Parser.

Key result

AI Overview presence and real-estate expanded while citation overlap with organic top ten remained low and highly vertical-specific.

Limitation

Useful longitudinal and B2B-tech cut; proprietary tracked keyword composition is not published.

Why this decision

Useful longitudinal and B2B-tech cut; proprietary tracked keyword composition is not published. JY source 28.

Source classification and provenance
Study design
Not stated
Evidence stream
Traffic/clicks
B2B relevance
High
Source type
industry longitudinal SERP analysis
Source class
Prior-review mined
Transparency
Moderate
Dataset family
BRIGHTEDGE-AIO-LONGITUDINAL-2025-26
Provenance
JY Scauri review: retained source 28 of 34; primary BrightEdge page verified.
Origin
master
Corrections applied
None
S131Decision: IncludeReview: Landing-pageEvidence: Traffic/clicksB2B: High

Debunking the Myth That SEO Traffic Has Dramatically Declined

Graphite and Similarweb / 2026-01 / Industry / practice research

Open direct source
Sample or setting

More than 40,000 large US websites using Similarweb estimates, validated against Google Search Console with median correlation of 0.86.

Key result

Estimated SEO traffic declined 2.5% year over year overall while Google traffic rose 0.8%, with losses concentrated outside the very largest sites.

Limitation

Large denominator and validation make this useful market context; proprietary panel and site-selection bias remain. JY retained study 1; overlaps existing project traffic register.

Why this decision

Large denominator and validation make this useful market context; proprietary panel and site-selection bias remain. JY retained study 1; overlaps existing project traffic register.

Source classification and provenance
Study design
Not stated
Evidence stream
Traffic/clicks
B2B relevance
High
Source type
industry panel analysis
Source class
Prior-review mined
Transparency
Moderate
Dataset family
SIMILARWEB-GRAPHITE-SEARCH-2026
Provenance
JY Scauri review: retained source 1 of 34 from 77 screened.
Origin
master
Corrections applied
None
S132Decision: IncludeReview: AbstractEvidence: Traffic/clicksB2B: Medium

Disentangling Answer Engine Optimization from Platform Growth: A Log-Based Natural Experiment on ChatGPT Referral Traffic

Keisuke Watanabe and Kazuki Nakayashiki / 2026 / Academic research

Open direct source
Sample or setting

Server logs from one site with an internal control and interrupted time-series analysis.

Key result

Reported a 1.82-times referral ratio after intervention, but the placebo test was not conventionally significant at p=.16.

Limitation

Rare quasi-experimental traffic evidence; single-site design and weak placebo result require low-confidence interpretation.

Why this decision

Rare quasi-experimental traffic evidence; single-site design and weak placebo result require low-confidence interpretation. Martinez matrix row 39.

Source classification and provenance
Study design
Not stated
Evidence stream
Traffic/clicks
B2B relevance
Medium
Source type
preprint natural experiment
Source class
Prior-review mined
Transparency
High
Dataset family
AEO-TRAFFIC-NATURAL-EXPERIMENT-2026
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S133Decision: IncludeReview: Landing-pageEvidence: Traffic/clicksB2B: High

2025 State of AI Discovery Report: What 1.96 Million LLM Sessions Tell Us About the Future of Search

Previsible / 2026-01 / Industry / practice research

Open direct source
Sample or setting

1,963,544 identifiable LLM-referred sessions across 12 months and sites in SaaS, e-commerce, finance, legal, health and publishing.

Key result

AI referrals were 0.13% of sessions overall but disproportionately reached industry, tool and pricing pages; SaaS had 16.4% of AI traffic landing on pricing pages.

Limitation

One of the strongest B2B-adjacent traffic-intent datasets; identifiable referrals miss zero-click influence and site selection is undisclosed.

Why this decision

One of the strongest B2B-adjacent traffic-intent datasets; identifiable referrals miss zero-click influence and site selection is undisclosed. JY source 24.

Source classification and provenance
Study design
Not stated
Evidence stream
Traffic/clicks
B2B relevance
High
Source type
industry web-analytics study
Source class
Prior-review mined
Transparency
Moderate
Dataset family
PREVISIBLE-AI-DISCOVERY-2025
Provenance
JY Scauri review: retained source 24 of 34; stable canonical page verified after previsible.io redirect.
Origin
master
Corrections applied
None
S134Decision: IncludeReview: Landing-pageEvidence: Traffic/clicksB2B: Low

Journalism, Media, and Technology Trends and Predictions 2026

Reuters Institute with Chartbeat / 2026-01 / Industry / practice research

Open direct source
Sample or setting

Chartbeat traffic data covering more than 2,500 publisher sites plus industry survey material.

Key result

Google organic traffic to the covered publisher cohort fell about 33% from November 2024 to November 2025.

Limitation

Large publisher-specific outcome; do not generalize to all websites or B2B SaaS.

Why this decision

Large publisher-specific outcome; do not generalize to all websites or B2B SaaS. JY source 11.

Source classification and provenance
Study design
Not stated
Evidence stream
Traffic/clicks
B2B relevance
Low
Source type
industry and institute report
Source class
Prior-review mined
Transparency
Moderate
Dataset family
CHARTBEAT-PUBLISHER-2025
Provenance
JY Scauri review: retained source 11 of 34.
Origin
master
Corrections applied
None
S135Decision: IncludeReview: Landing-pageEvidence: Traffic/clicksB2B: High

AIO Impact on Google CTR: 2026 Update

Seer Interactive / 2026 / Industry / practice research

Open direct source
Sample or setting

Rolling client panel; current public edition reports 53 brands, 5.47 million tracked queries and 2.43 billion impressions.

Key result

Organic CTR fell materially on queries showing AI Overviews, with large differences by brand and query type.

Limitation

Large first-party panel, but changing cohort means the current page differs from JY's earlier 42-organization, 25.1M-impression snapshot. Treat as rolling lineage, not two studies.

Why this decision

Large first-party panel, but changing cohort means the current page differs from JY's earlier 42-organization, 25.1M-impression snapshot. Treat as rolling lineage, not two studies.

Source classification and provenance
Study design
Not stated
Evidence stream
Traffic/clicks
B2B relevance
High
Source type
industry longitudinal analysis
Source class
Prior-review mined
Transparency
Moderate
Dataset family
SEER-AIO-CTR-ROLLING
Provenance
JY Scauri review: retained source 4 of 34; public page has since been updated.
Origin
master
Corrections applied
None
S136Decision: IncludeReview: Landing-pageEvidence: Traffic/clicksB2B: Low

BREAKING! News Thrives in the Age of AI

Shahzad Abbas / Define Media Group / 2026-03-09 / Industry / practice research

Open direct source
Sample or setting

Google Search Console data from a 64-site core panel; content classification on the top 15 national and local news brands.

Key result

Organic search traffic was 42% below the pre-AIO baseline while breaking-news traffic grew 103%, driven mainly by Google Discover.

Limitation

High-volume first-party publisher evidence with a disclosed method; publisher and content-mix effects do not transfer directly to B2B SaaS.

Why this decision

High-volume first-party publisher evidence with a disclosed method; publisher and content-mix effects do not transfer directly to B2B SaaS. JY source 31.

Source classification and provenance
Study design
Not stated
Evidence stream
Traffic/clicks
B2B relevance
Low
Source type
industry panel analysis
Source class
Prior-review mined
Transparency
High
Dataset family
DEFINE-MEDIA-PUBLISHER-PANEL-2026
Provenance
JY Scauri review: retained source 31 of 34; original Define Media page independently located and checked.
Origin
master
Corrections applied
None
S137Decision: IncludeReview: AbstractEvidence: Trust, clicks and citationsB2B: High

Human Trust in AI Search: A Large-Scale Experiment

Haiwen Li, Sinan Aral; MIT / 2025-04 / Academic research

Open direct source
Sample or setting

About 12,000 queries across seven countries yielding about 80,000 results, plus a preregistered randomized US-representative experiment

Key result

Links and citations increased trust even when wrong or hallucinated, and higher trust predicted more clicking and less time evaluating the answer.

Limitation

Controlled trust tasks establish interface causality, not sustained real-world buying behavior or pipeline impact.

Why this decision

Causal interface evidence showing why citation visibility can influence behavior without corresponding factual quality.

Source classification and provenance
Study design
Not stated
Evidence stream
Trust, clicks and citations
B2B relevance
High
Source type
Preprint
Source class
Academic
Transparency
High
Dataset family
MIT AI search trust experiment
Provenance
arXiv search for AI search user trust
Origin
master
Corrections applied
None
S138Decision: IncludeReview: AbstractEvidence: User behavior and citation usabilityB2B: Medium

Search Engines in the AI Era: A Qualitative Understanding to the False Promise of Factual and Verifiable Source-Cited Responses in LLM-Based Search

Pranav Narayanan Venkit et al.; Salesforce Research / 2025-06 / Academic research

Open direct source
Sample or setting

Three-person pilot, 21-person main qualitative study, plus automated audit of You.com, Perplexity and Bing Chat

Key result

Participants encountered 16 recurring answer-engine limitations, including hallucinated or misleading citations, and the automated audit reproduced many of those problems.

Limitation

Small qualitative sample and early-generation products limit prevalence claims and current-platform transfer.

Why this decision

Rare human-centered evidence on how source-cited answers are actually interpreted and verified.

Source classification and provenance
Study design
Not stated
Evidence stream
User behavior and citation usability
B2B relevance
Medium
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
Answer Engine Evaluation benchmark
Provenance
Martinez 2026 core matrix; ACM DOI; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S139Decision: IncludeReview: AbstractEvidence: User behavior and decision qualityB2B: High

Generative AI as a New Paradigm for Online Search: Evidence from a Large-Scale Experiment and Qualitative Interviews

Jakob Kaiser et al.; German research collaborators / 2026-05 / Industry / practice research

Open direct source
Sample or setting

Online experiment assigning ChatGPT or Google for smartphone and holiday-choice tasks, plus qualitative interviews

Key result

ChatGPT users found the optimal option more often and faster while consulting significantly fewer external sites, consistent with AI functioning as a one-stop decision intermediary.

Limitation

Short consumer-choice tasks under randomized tool assignment do not reproduce long-cycle, multi-person B2B buying.

Why this decision

Direct randomized comparison on complex consumer decisions with strong mechanism relevance.

Source classification and provenance
Study design
Not stated
Evidence stream
User behavior and decision quality
B2B relevance
High
Source type
Working paper
Source class
Academic
Transparency
Medium
Dataset family
ChatGPT-versus-Google task experiment
Provenance
SSRN search for generative AI search experiments
Origin
master
Corrections applied
None
S140Decision: IncludeReview: FullEvidence: VisibilityB2B: Indirect / transfer

Ahrefs Brand Radar cross-sectional correlations

Ahrefs / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

75,000 selected brands with domain rating above 40 and a high-volume keyword; millions of AI responses across ChatGPT, AI Mode and AIO

Key result

YouTube mentions correlated about 0.737 with AI visibility; branded web mentions about 0.66-0.71; page volume only about 0.194. Output overlap across platforms was about 0.779

Limitation

Highly selected established brands; correlational; common prompt pools and Google-owned YouTube may inflate relationships; no causal result.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Mentions, recommendations and citations
Evidence stream
Visibility
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
V4
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official Ahrefs page; JY Scauri review: retained source 20 of 34; same broad Ahrefs data lineage as REV-051 and REV-064.
Origin
master
Corrections applied
None
S141Decision: IncludeReview: FullEvidence: VisibilityB2B: Indirect / transfer

Semrush + Kevin Indig “Ghost Citations”; Semrush AI Visibility Toolkit

Semrush / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

3,981 domain appearances from 115 prompts in 14 countries across ChatGPT, AIO, Gemini and AI Mode

Key result

61.7% of all appearances were citation-only, 13.2% both cited and mentioned, and 25.1% mention-only. Thus 74.9% of all appearances were cited and 38.3% mentioned. Comparative prompts produced 2.4x as many mentions as informational prompts

Limitation

Important denominator error in the public framing: the defensible result is “61.7% of all appearances were ghost citations.” If the categories are exhaustive, 61.7/74.9 = **82.4% of cited appearances** lacked a mention, not 62%. Prompt selection and platform weighting are opaque. Same Semrush/Kevin dataset family as related Growth Memo work.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Mentions, recommendations and citations
Evidence stream
Visibility
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
V1
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official Semrush page
Origin
master
Corrections applied
None
S142Decision: IncludeReview: FullEvidence: VisibilityB2B: Indirect / transfer

Semrush + Kevin Indig ChatGPT topic-authority study

Semrush / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

1,094 US topical categories, five designed prompts each, tracked monthly January-June 2026; 50,000+ brands, 220,000+ domains and 600,000+ citations

Key result

Only 15.2% of categories had a clear mention leader; 53.7% had no brand appearing in at least three of five prompts. Only 21% of most-cited domains were also the most-mentioned brand; citation leadership and mention leadership correlated -0.229

Limitation

Prompts are a designed index, not observed user-prompt frequencies; ownership thresholds are author-defined; ChatGPT only. Likely shares Semrush AI Visibility Toolkit infrastructure with V1 but is a distinct larger panel/time series.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Mentions, recommendations and citations
Evidence stream
Visibility
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
V2
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official Semrush page
Origin
master
Corrections applied
None
S143Decision: IncludeReview: FullEvidence: VisibilityB2B: Indirect / transfer

Semrush AI citation-volatility study

Semrush / 2026 or earlier / Industry / practice research

Open direct source
Sample or setting

230,000 prompts, weekly snapshots for 13 weeks, ChatGPT Search, AI Mode and Perplexity; 14 July-12 October 2025; more than 100M citations

Key result

Reddit appeared in nearly 60% of ChatGPT responses in early August then about 10% by mid-September; Wikipedia fell from roughly 55% to under 20%, while other engines moved differently

Limitation

Designed/monitored prompts, source-share metric and no causal attribution to a specific platform change. Same Semrush AI Visibility data infrastructure; proves volatility, not buyer impact.

Why this decision

Already admitted to the evidence register

Source classification and provenance
Study design
Mentions, recommendations and citations
Evidence stream
Visibility
B2B relevance
Indirect / transfer
Source type
Commercial vendor
Source class
Baseline evidence register
Transparency
Medium
Dataset family
V3
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official Semrush page
Origin
master
Corrections applied
None
S144Decision: IncludeReview: AbstractEvidence: Visibility stabilityB2B: High

Don't Measure Once: Measuring Visibility in AI Search (GEO)

Julius Schulte, Malte Bleeker, Philipp Kaufmann; University of St. Gallen / 2026-04 / Academic research

Open direct source
Sample or setting

Repeated AI-search visibility measurements over a small Swiss business universe

Key result

Run-to-run and time variation was large enough that a single prompt execution misrepresented brand visibility, supporting repeated panel measurement rather than deterministic ranks.

Limitation

Directly relevant measurement design, though the sampled market is small and geographically narrow.

Why this decision

Directly relevant measurement design, though the sampled market is small and geographically narrow.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility stability
B2B relevance
High
Source type
Preprint
Source class
Academic
Transparency
Medium
Dataset family
Swiss repeated visibility audit
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix; overlap with Cyrus Shepard 54-source sheet.
Origin
master
Corrections applied
None
S145Decision: IncludeReview: Landing-pageEvidence: Visibility/citationB2B: High

AI Overview Citations Drop from 76% to 38% Organic Top-10 Overlap

Ahrefs / 2026-03 / Industry / practice research

Open direct source
Sample or setting

863,000 keywords and about four million AI Overview citation URLs.

Key result

Only about 38% of AI Overview citations came from pages ranking in Google's organic top ten in the updated snapshot.

Limitation

Large live-SERP evidence for citation-ranking divergence; it is one Ahrefs rolling dataset, not independent from related Brand Radar analyses.

Why this decision

Large live-SERP evidence for citation-ranking divergence; it is one Ahrefs rolling dataset, not independent from related Brand Radar analyses. JY source 6.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
industry SERP analysis
Source class
Prior-review mined
Transparency
Moderate
Dataset family
AHREFS-BRAND-RADAR-2025-26
Provenance
JY Scauri review: retained source 6 of 34.
Origin
master
Corrections applied
None
S146Decision: IncludeReview: Landing-pageEvidence: Visibility/citationB2B: High

Do AI Assistants Prefer to Cite Fresh Content?

Ahrefs / 2026 / Industry / practice research

Open direct source
Sample or setting

16.975 million cited URLs across seven AI surfaces compared with organic-result freshness.

Key result

AI-cited URLs were on average 25.7% fresher than organic results in the analyzed set.

Limitation

Very large sample, but freshness is correlational and the same Ahrefs Brand Radar family supplies several analyses.

Why this decision

Very large sample, but freshness is correlational and the same Ahrefs Brand Radar family supplies several analyses. JY source 19.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
industry citation analysis
Source class
Prior-review mined
Transparency
Moderate
Dataset family
AHREFS-BRAND-RADAR-2025-26
Provenance
JY Scauri review: retained source 19 of 34; same broad Ahrefs data lineage as REV-051 and REV-065.
Origin
master
Corrections applied
None
S147Decision: IncludeReview: Landing-pageEvidence: Visibility/citationB2B: High

From Retrieved to Cited: How Commercial Content Earns Citations in AI Search

AirOps / 2026-04-01 / Industry / practice research

Open direct source
Sample or setting

7,500 commercial prompts, 217,508 retrieved pages and 25 on-page signals.

Key result

Statistics, tables, lists and tighter sentence structure were associated with citation conditional on retrieval.

Limitation

Commercial-prompt subset is highly relevant, but it is a derivative analysis within the AirOps retrieval family and correlations are conditional on retrieval.

Why this decision

Commercial-prompt subset is highly relevant, but it is a derivative analysis within the AirOps retrieval family and correlations are conditional on retrieval.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
industry content-factor analysis
Source class
Prior-review mined
Transparency
Moderate
Dataset family
AIROPS-CHATGPT-RETRIEVAL-2026
Provenance
Cyrus Shepard 54-source evidence sheet; AirOps citation chain.
Origin
master
Corrections applied
None
S148Decision: IncludeReview: Landing-pageEvidence: Visibility/citationB2B: High

The Fan-Out Effect: What Happens Between a Query and a Citation

AirOps / 2026-04-13 / Industry / practice research

Open direct source
Sample or setting

16,851 original queries, 50,553 runs and 353,799 retrieved pages with citation outcomes by retrieval rank.

Key result

The rank-one retrieved page was cited in 58.4% of runs versus 14.2% for rank ten, and fan-out queries broadened the source pool.

Limitation

Detailed retrieval-stage evidence; part of the same AirOps 2026 program as REV-068 and later derivatives, not an independent confirmation.

Why this decision

Detailed retrieval-stage evidence; part of the same AirOps 2026 program as REV-068 and later derivatives, not an independent confirmation.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
industry retrieval analysis
Source class
Prior-review mined
Transparency
Moderate
Dataset family
AIROPS-CHATGPT-RETRIEVAL-2026
Provenance
Cyrus Shepard 54-source evidence sheet; AirOps citation chain.
Origin
master
Corrections applied
None
S149Decision: IncludeReview: Landing-pageEvidence: Visibility/citationB2B: High

The Influence of Retrieval, Fan-Out, and Google SERPs on ChatGPT Citations

AirOps / 2026-03 / Industry / practice research

Open direct source
Sample or setting

548,534 retrieved pages across 15,000 original prompts and 43,233 total fan-out queries.

Key result

About 85% of retrieved pages were not cited and roughly one-third of citations came through fan-out queries rather than the original prompt.

Limitation

Directly separates retrieval from final citation; same underlying AirOps research program as REV-080, REV-081 and Kevin's REV-089, so do not count as independent replications.

Why this decision

Directly separates retrieval from final citation; same underlying AirOps research program as REV-080, REV-081 and Kevin's REV-089, so do not count as independent replications. JY source 23.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
industry retrieval analysis
Source class
Prior-review mined
Transparency
Moderate
Dataset family
AIROPS-CHATGPT-RETRIEVAL-2026
Provenance
JY Scauri review: retained source 23 of 34; repeated in Cyrus Shepard evidence sheet through follow-on analyses.
Origin
master
Corrections applied
None
S150Decision: IncludeReview: Landing-pageEvidence: Visibility/citationB2B: High

The AI Citation Economy Report 2025

OtterlyAI / 2026-02 / Industry / practice research

Open direct source
Sample or setting

More than one million citations across ChatGPT, Perplexity and Google AI Overviews.

Key result

Third-party sources dominated citation share, with owned brand sites contributing a small minority.

Limitation

Supports ecosystem visibility thesis; prompt-selection and deduplication methods are not fully disclosed.

Why this decision

Supports ecosystem visibility thesis; prompt-selection and deduplication methods are not fully disclosed. JY source 22.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
industry citation analysis
Source class
Prior-review mined
Transparency
Partial
Dataset family
OTTERLY-CITATION-ECONOMY-2025
Provenance
JY Scauri review: retained source 22 of 34.
Origin
master
Corrections applied
None
S151Decision: IncludeReview: Landing-pageEvidence: Visibility/citationB2B: High

The YouTube Citation Study 2026

OtterlyAI / 2026-03 / Industry / practice research

Open direct source
Sample or setting

More than 100 million citation instances over 30 days across six AI platforms.

Key result

Long-form YouTube content represented 94% of cited videos while view and subscriber popularity showed little relationship.

Limitation

Unusually large platform comparison; proprietary prompt universe and classification details are only partially public.

Why this decision

Unusually large platform comparison; proprietary prompt universe and classification details are only partially public. JY source 21.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
industry citation analysis
Source class
Prior-review mined
Transparency
Partial
Dataset family
OTTERLY-YOUTUBE-2026
Provenance
JY Scauri review: retained source 21 of 34.
Origin
master
Corrections applied
None
S152Decision: IncludeReview: Landing-pageEvidence: Visibility/citationB2B: High

LinkedIn Is the Most-Cited Domain for Professional Queries in AI Search

Profound / 2026-03-09 / Industry / practice research

Open direct source
Sample or setting

1.4 million citations across six models from November 15 to February 15 using two proprietary datasets.

Key result

LinkedIn was the leading cited domain for professional-query categories in the analyzed data.

Limitation

Direct B2B and executive-authority relevance; proprietary definition of professional queries and uncertain overlap across datasets require caution.

Why this decision

Direct B2B and executive-authority relevance; proprietary definition of professional queries and uncertain overlap across datasets require caution.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
industry B2B citation analysis
Source class
Prior-review mined
Transparency
Partial
Dataset family
PROFOUND-PROFESSIONAL-QUERIES-2026
Provenance
Cyrus Shepard 54-source evidence sheet; Profound citation chain.
Origin
master
Corrections applied
None
S153Decision: IncludeReview: Landing-pageEvidence: Visibility/citationB2B: High

Who Shapes AI Answers? Enhanced Citation Categories

Profound / 2026-01 / Industry / practice research

Open direct source
Sample or setting

Approximately 27 million citations across ChatGPT, Gemini and Google AI Overviews categorized by source type.

Key result

Only about 4% of citations came from brand-owned websites, while third-party ecosystems dominated.

Limitation

Central ecosystem-visibility evidence; proprietary prompts, categories and potential overlap with other Profound analyses require sensitivity treatment.

Why this decision

Central ecosystem-visibility evidence; proprietary prompts, categories and potential overlap with other Profound analyses require sensitivity treatment. JY source 34.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
industry citation analysis
Source class
Prior-review mined
Transparency
Partial
Dataset family
PROFOUND-ENHANCED-CITATIONS-2026
Provenance
JY Scauri review: retained source 34 of 34.
Origin
master
Corrections applied
None
S154Decision: IncludeReview: Landing-pageEvidence: Visibility/citationB2B: Medium

Google AI Overview Citation Share - January 2026

Promptwatch / 2026-01 / Industry / practice research

Open direct source
Sample or setting

568,499 prompts and 5,585,499 citations captured from Google AI Overviews during January 2026.

Key result

A small group of large domains, including YouTube and Reddit, captured a substantial share of observed citations.

Limitation

Large real-interface sample with a precise time window; prompt universe and collection weighting are not disclosed.

Why this decision

Large real-interface sample with a precise time window; prompt universe and collection weighting are not disclosed. JY source 33.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
Medium
Source type
industry citation dashboard
Source class
Prior-review mined
Transparency
Moderate
Dataset family
PROMPTWATCH-AIO-JAN2026
Provenance
JY Scauri review: retained source 33 of 34; direct monthly data page checked.
Origin
master
Corrections applied
None
S155Decision: IncludeReview: Landing-pageEvidence: Visibility/citationB2B: Medium

Google Links in AI Mode Answers: Self-Citation Study

SE Ranking / 2026 / Industry / practice research

Open direct source
Sample or setting

68,313 keywords and 1,321,398 citations from Google AI Mode answers.

Key result

Google properties captured a material share of AI Mode citations and citation patterns varied by result component.

Limitation

Large surface-specific audit; proprietary prompt selection and fast-moving interface limit durability.

Why this decision

Large surface-specific audit; proprietary prompt selection and fast-moving interface limit durability. JY source 17.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
Medium
Source type
industry SERP analysis
Source class
Prior-review mined
Transparency
Moderate
Dataset family
SERANKING-AIMODE-LINKS-2026
Provenance
JY Scauri review: retained source 17 of 34.
Origin
master
Corrections applied
None
S156Decision: IncludeReview: Landing-pageEvidence: Visibility/citationB2B: High

How to Increase Visibility in AI Search Engines: Citation Factor Study

SE Ranking / 2026 / Industry / practice research

Open direct source
Sample or setting

ChatGPT: about 129,000 domains and 216,000 pages; AI Mode: about 295,000 domains and 2.3 million pages; more than 50 factors modeled with XGBoost and SHAP.

Key result

Citation correlates differed across ChatGPT and AI Mode, with no single tactic transferring uniformly.

Limitation

Large comparative dataset; correlational model, proprietary sampling and rolling page updates prevent causal tactic claims.

Why this decision

Large comparative dataset; correlational model, proprietary sampling and rolling page updates prevent causal tactic claims. JY source 18.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
industry machine-learning analysis
Source class
Prior-review mined
Transparency
Moderate
Dataset family
SERANKING-AI-CITATION-FACTORS-2025-26
Provenance
JY Scauri review: retained source 18 of 34; current page is a rolling public edition.
Origin
master
Corrections applied
None
S157Decision: IncludeReview: Landing-pageEvidence: Visibility/citationB2B: High

LLM Ghost Citations: Why Your Content Is Working and Your Brand Isn't

Seer Interactive using Scrunch data / 2026 / Industry / practice research

Open direct source
Sample or setting

541,213 responses across 20 brands and six AI platforms using six tests.

Key result

A recommended brand was also cited 53.1% of the time versus 10.6% when the brand was not recommended.

Limitation

Strong evidence that recommendation and citation are different outcomes; proprietary response panel and possible Scrunch lineage overlap must be disclosed.

Why this decision

Strong evidence that recommendation and citation are different outcomes; proprietary response panel and possible Scrunch lineage overlap must be disclosed. JY source 5.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
industry behavioral analysis
Source class
Prior-review mined
Transparency
Moderate
Dataset family
SCRUNCH-SEER-GHOST-CITATIONS-2026
Provenance
JY Scauri review: retained source 5 of 34.
Origin
master
Corrections applied
None
S158Decision: IncludeReview: Landing-pageEvidence: Visibility/citationB2B: High

AI Mode Comparison Study

Semrush / 2025-07-21 / Industry / practice research

Open direct source
Sample or setting

5,000 keywords, about 150,000 citations, four intent groups and four AI search platforms.

Key result

Google AI Mode's sidebar had about 51% domain overlap and 32% URL overlap with Google's top ten, while other platforms differed.

Limitation

Useful same-query cross-platform comparison; vendor-selected keywords and an older interface snapshot limit current generalization.

Why this decision

Useful same-query cross-platform comparison; vendor-selected keywords and an older interface snapshot limit current generalization.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
industry cross-platform citation study
Source class
Prior-review mined
Transparency
Moderate
Dataset family
SEMRUSH-AI-MODE-COMPARISON-2025
Provenance
Cyrus Shepard 54-source evidence sheet; Semrush citation chain.
Origin
master
Corrections applied
None
S159Decision: IncludeReview: Landing-pageEvidence: Visibility/citationB2B: High

AI Mode Rankings Overlap Study

seoClarity / 2025-09 / Industry / practice research

Open direct source
Sample or setting

1,000 transactional US desktop queries producing 12,011 AI Mode citations.

Key result

Only 19% of cited URLs also ranked in the organic top 20, much lower than the same vendor's AIO snapshot.

Limitation

Useful same-vendor contrast between AI surfaces; smaller transactional-only sample and platform drift remain.

Why this decision

Useful same-vendor contrast between AI surfaces; smaller transactional-only sample and platform drift remain.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
industry SERP analysis
Source class
Prior-review mined
Transparency
Moderate
Dataset family
SEOCLARITY-AIMODE-OVERLAP-2025
Provenance
Cyrus Shepard 54-source evidence sheet.
Origin
master
Corrections applied
None
S160Decision: IncludeReview: Landing-pageEvidence: Visibility/citationB2B: High

AI Overview Rankings Overlap Study

seoClarity / 2025-10 / Industry / practice research

Open direct source
Sample or setting

362,000 US desktop queries and about 5.1 million AI Overview citations.

Key result

Fifty-six percent of citations came from organic top-20 results and the organic number-one position appeared in 43% of AIOs in this snapshot.

Limitation

Large direct surface audit; vendor keyword set and one-country desktop snapshot limit generalization.

Why this decision

Large direct surface audit; vendor keyword set and one-country desktop snapshot limit generalization.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
industry SERP analysis
Source class
Prior-review mined
Transparency
Moderate
Dataset family
SEOCLARITY-AIO-OVERLAP-2025
Provenance
Cyrus Shepard 54-source evidence sheet.
Origin
master
Corrections applied
None
S161Decision: IncludeReview: Landing-pageEvidence: Visibility/citationB2B: High

Q1 2026 AI Citation Trends Report

Tinuiti and Profound / 2026-03 / Industry / practice research

Open direct source
Sample or setting

Mid- and lower-funnel prompts across seven AI platforms and nine commercial categories from October 2025 through January 2026.

Key result

Citation-source mix varied sharply by platform and category; social and marketplace domains changed share over time.

Limitation

Commercial-intent and B2B technology categories are relevant; prompt counts and possible overlap with Profound's wider corpus are not fully disclosed.

Why this decision

Commercial-intent and B2B technology categories are relevant; prompt counts and possible overlap with Profound's wider corpus are not fully disclosed. JY source 29.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
industry longitudinal citation report
Source class
Prior-review mined
Transparency
Partial
Dataset family
TINUITI-PROFOUND-CITATION-TRENDS-Q1-2026
Provenance
JY Scauri review: retained source 29 of 34; primary Tinuiti analysis page for the Q1 report.
Origin
master
Corrections applied
None
S162Decision: IncludeReview: Landing-pageEvidence: Visibility/citationB2B: High

The Content Types Most Cited by LLMs

Wix AI Search Lab and Peec AI / 2026 / Industry / practice research

Open direct source
Sample or setting

75,000 AI answers and 1,056,727 citations across ChatGPT, AI Mode and Perplexity.

Key result

Listicles, general articles and product pages accounted for about 52% of citations, with formats varying by query intent.

Limitation

Large intent-and-format dataset; proprietary prompt mix and classifier limit generalization to all categories.

Why this decision

Large intent-and-format dataset; proprietary prompt mix and classifier limit generalization to all categories. JY source 26.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
industry citation analysis
Source class
Prior-review mined
Transparency
Moderate
Dataset family
WIX-PEEC-CONTENT-TYPES-2026
Provenance
JY Scauri review: retained source 26 of 34.
Origin
master
Corrections applied
None
S163Decision: ContextReview: FullEvidence: adoption, platform fragmentation and high-consideration exposureB2B: Medium - consumer/high-consideration mechanism.

Comscore Q1 2026 AI Intelligence Report

Comscore / 2026-06-02 / Industry / practice research

Open direct source
Sample or setting

Comscore US desktop/mobile panel and search/ad-exposure data; Q1 2026 with comparisons to 2024/2025; full report details not public on release page.

Key result

AI assistants reached 36% of desktop and 23% of mobile users; Claude conversations grew sharply; about 25% of credit-card applicants had AIO exposure over three quarters.

Limitation

Valuable independent-ish panel and high-consideration exposure example, but only press-release depth and no causal purchase estimate.

Why this decision

Valuable independent-ish panel and high-consideration exposure example, but only press-release depth and no causal purchase estimate.

Source classification and provenance
Study design
Not stated
Evidence stream
adoption, platform fragmentation and high-consideration exposure
B2B relevance
Medium - consumer/high-consideration mechanism.
Source type
official report press release
Source class
Industry / original-data report
Transparency
Low-moderate - official release gives selected metrics but not full sample, weights or models.
Dataset family
COMSCORE_AI_INTELLIGENCE_Q1_2026
Provenance
Official Comscore search
Origin
master
Corrections applied
None
S164Decision: ContextReview: FullEvidence: adoption, prompt behavior and referralsB2B: Medium - all-market discovery behavior.

Stepping Into the Conversation: Insights from the 2025 GenAI Landscape Report

Similarweb / 2025-11-25 / Industry / practice research

Open direct source
Sample or setting

Billions of global web and app signals in Similarweb's 2025 Generative AI Landscape report; 2025 platform, session and referral patterns.

Key result

GenAI web visits rose 76% and app downloads 319%; reported average ChatGPT prompts around 60 words versus Google queries around 3.4.

Limitation

Useful predecessor baseline but overlaps the Similarweb panel and summary page lacks the full report's exact denominators.

Why this decision

Useful predecessor baseline but overlaps the Similarweb panel and summary page lacks the full report's exact denominators.

Source classification and provenance
Study design
Not stated
Evidence stream
adoption, prompt behavior and referrals
B2B relevance
Medium - all-market discovery behavior.
Source type
full official report summary tied to downloadable report
Source class
Industry / original-data report
Transparency
Moderate - full official summary and report lineage; proprietary methodology and some bases gated.
Dataset family
SIMILARWEB_GENAI_LANDSCAPE_2025
Provenance
Official Similarweb AI-news index and report press release
Origin
master
Corrections applied
None
S165Decision: ContextReview: Landing-pageEvidence: Adoption/reachB2B: Medium

State of Consumer AI 2025: Product Hits, Misses, and What's Next

Andreessen Horowitz using YipitData / 2025-12 / Industry / practice research

Open direct source
Sample or setting

YipitData usage, subscription and retention panels across consumer AI products.

Key result

Usage is fragmenting across products while many users remain loyal to a primary platform.

Limitation

Useful platform context but commercial venture perspective and opaque panel methods limit claim weight.

Why this decision

Useful platform context but commercial venture perspective and opaque panel methods limit claim weight. JY source 9.

Source classification and provenance
Study design
Not stated
Evidence stream
Adoption/reach
B2B relevance
Medium
Source type
industry market analysis
Source class
Prior-review mined
Transparency
Partial
Dataset family
YIPIT-CONSUMER-AI-2025
Provenance
JY Scauri review: retained source 9 of 34.
Origin
master
Corrections applied
None
S166Decision: ContextReview: FullEvidence: AI adoption versus traditional searchB2B: Medium-high - market context; no B2B cut.

State of Search Q1 2026: Behaviors, Trends, and Clicks Across the US and Europe

Datos and Rand Fishkin / 2026 / Industry / practice research

Open direct source
Sample or setting

Large-scale clickstream from millions of real users across the US, EU and UK; platform usage, AI Mode and discovery behavior.

Key result

The report benchmarks AI-tool adoption against search and shows regional/device differences; exact tables require the downloaded report.

Limitation

Highly relevant common-denominator panel, but only landing-page depth was available during this screen.

Why this decision

Highly relevant common-denominator panel, but only landing-page depth was available during this screen.

Source classification and provenance
Study design
Not stated
Evidence stream
AI adoption versus traditional search
B2B relevance
Medium-high - market context; no B2B cut.
Source type
gated official report landing page
Source class
Industry / original-data report
Transparency
Low at screen - official gated landing page states panel scope but not result tables or exact dates.
Dataset family
DATOS_STATE_OF_SEARCH_Q1_2026
Provenance
Official Datos report search; JY Scauri review: retained source 15 of 34; already represented in project traffic register family.
Origin
master
Corrections applied
None
S167Decision: ContextReview: AbstractEvidence: AI Overview snippet manipulationB2B: Low

Exploring LLM Biases to Manipulate AI Search Overview

Roman Smirnov / 2026-05 / Academic research

Open direct source
Sample or setting

Reinforcement learning against a simulated overview and comparative preference signal

Key result

An RL policy exploited model comparison preferences to promote snippets in a simulated overview, showing a possible mechanism rather than a live Google effect.

Limitation

Interesting ranking mechanism with limited ecological validity.

Why this decision

Interesting ranking mechanism with limited ecological validity.

Source classification and provenance
Study design
Not stated
Evidence stream
AI Overview snippet manipulation
B2B relevance
Low
Source type
Preprint
Source class
Academic
Transparency
Medium
Dataset family
Simulated AI overview RL
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S168Decision: ContextReview: FullEvidence: AI shopping trigger mechanicsB2B: Medium - explicitly shows non-transfer of the shopping carousel to SaaS.

We Reverse-Engineered ChatGPT's Shopping Trigger

Profound / 2026-03-17 / Industry / practice research

Open direct source
Sample or setting

1.18M prompts analyzed with a 7,500-prompt labeled ground-truth sample; classifier reproduced Shopping-card behavior at about 95-97% accuracy.

Key result

Physical-product category was far more predictive than purchase-intent wording; software, services, travel and finance almost never triggered Shopping.

Limitation

Useful boundary: commerce surfaces are not evidence for B2B SaaS answer behavior.

Why this decision

Useful boundary: commerce surfaces are not evidence for B2B SaaS answer behavior.

Source classification and provenance
Study design
Not stated
Evidence stream
AI shopping trigger mechanics
B2B relevance
Medium - explicitly shows non-transfer of the shopping carousel to SaaS.
Source type
full vendor classification experiment
Source class
Industry / original-data report
Transparency
High-moderate - full method and labeled base; sample construction/model update sensitivity remain.
Dataset family
PROFOUND_CHATGPT_SHOPPING_2026
Provenance
Official Profound research hub and full-page inspection
Origin
master
Corrections applied
None
S169Decision: ContextReview: AbstractEvidence: Answer coverageB2B: Medium

Do RAG Systems Cover What Matters? Evaluating and Optimizing Responses with Sub-Question Coverage

Kaige Xie et al.; Salesforce Research / 2025-04 / Academic research

Open direct source
Sample or setting

Complex questions decomposed into subquestions to evaluate and optimize RAG response coverage

Key result

Citation volume alone can conceal missing parts of an information need; subquestion coverage provided a more decision-relevant measure and could be optimized directly.

Limitation

Not an organic visibility study, but supplies a necessary metric for whether cited sources actually cover a buyer's task.

Why this decision

Not an organic visibility study, but supplies a necessary metric for whether cited sources actually cover a buyer's task.

Source classification and provenance
Study design
Not stated
Evidence stream
Answer coverage
B2B relevance
Medium
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
Subquestion coverage benchmark
Provenance
Martinez 2026 references; ACL Anthology
Origin
master
Corrections applied
None
S170Decision: ContextReview: AbstractEvidence: Answer robustness and fidelityB2B: Medium

Evaluating Robustness of Generative Search Engine on Adversarial Factoid Questions

Xuming Hu et al.; multiple universities and industry labs / 2024-08 / Academic research

Open direct source
Sample or setting

Human evaluation of Bing Chat, Perplexity and YouChat under black-box adversarial factual queries

Key result

Subtle adversarial factoid questions induced incorrect answers across engines, and retrieval-augmented configurations were more susceptible than comparable models without retrieval.

Limitation

Direct commercial-engine reliability evidence, but the adversarial query setting is not an ordinary buyer journey.

Why this decision

Direct commercial-engine reliability evidence, but the adversarial query setting is not an ordinary buyer journey.

Source classification and provenance
Study design
Not stated
Evidence stream
Answer robustness and fidelity
B2B relevance
Medium
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
Adversarial factoid GSE audit
Provenance
ACL Anthology search for generative search evaluation
Origin
master
Corrections applied
None
S171Decision: ContextReview: AbstractEvidence: Attack and white-hat benchmarkB2B: Medium

GEO-Bench: Benchmarking Ranking Manipulation in Generative Engine Optimization

Ojas Nimase et al.; Michigan State University and collaborators / 2026-05 / Academic research

Open direct source
Sample or setting

Benchmark comparison of adversarial and white-hat methods against one LLM ranker

Key result

Several ranking interventions moved target items in the tested ranker, but evidence rests on one ranking system and the benchmark name can be confused with the original GEO-bench.

Limitation

Useful comparative screen, insufficient for a cross-platform practical recommendation.

Why this decision

Useful comparative screen, insufficient for a cross-platform practical recommendation.

Source classification and provenance
Study design
Not stated
Evidence stream
Attack and white-hat benchmark
B2B relevance
Medium
Source type
Preprint
Source class
Academic
Transparency
High
Dataset family
GEO-Bench ranking manipulation
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S172Decision: ContextReview: AbstractEvidence: Authorship metadata and attributionB2B: Medium

Evaluation of Attribution Bias in Generator-Aware Retrieval-Augmented Large Language Models

Amin Abolghasemi et al.; University of Strathclyde, University of Amsterdam, Leiden University / 2025-07 / Academic research

Open direct source
Sample or setting

Counterfactual tests of three LLMs with human versus AI authorship metadata on retrieved documents

Key result

Adding authorship information changed attribution quality by 3% to 18%, with models often preferring explicitly human-authored sources.

Limitation

Clean mechanism evidence that metadata changes source use, though not a live answer-engine audit.

Why this decision

Clean mechanism evidence that metadata changes source use, though not a live answer-engine audit.

Source classification and provenance
Study design
Not stated
Evidence stream
Authorship metadata and attribution
B2B relevance
Medium
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
Generator-aware attribution benchmark
Provenance
ACL Anthology search for source attribution bias
Origin
master
Corrections applied
None
S173Decision: ContextReview: FullEvidence: B2B buying groups, GenAI and purchase riskB2B: Very high - direct business buying.

The State of Business Buying 2026

Forrester / 2026-01-21 / Industry / practice research

Open direct source
Sample or setting

Forrester 2026 Buyer Insights program; public release reports stakeholder/group/trial findings but omits survey n and instrument.

Key result

GenAI searches were described as a starting point while the typical decision involved 13 internal stakeholders and nine external influencers; trials rose with deal value.

Limitation

likely overlaps the 2025 Buyers' Journey Survey.

Why this decision

Useful board-level synthesis but press-release depth cannot support precise independent estimates; likely overlaps the 2025 Buyers' Journey Survey.

Source classification and provenance
Study design
Not stated
Evidence stream
B2B buying groups, GenAI and purchase risk
B2B relevance
Very high - direct business buying.
Source type
official report press release
Source class
Industry / original-data report
Transparency
Low - official press release; full client report and method unavailable publicly.
Dataset family
FORRESTER_BUYERS_JOURNEY_2025
Provenance
Official Forrester press newsroom
Origin
master
Corrections applied
None
S174Decision: ContextReview: AbstractEvidence: Citation attributionB2B: Low

CiteME: Can Language Models Accurately Cite Scientific Claims?

Kyle Lo et al.; Allen Institute for AI and University of Washington / 2024-12 / Academic research

Open direct source
Sample or setting

Scientific claim-to-paper identification benchmark comparing frontier LMs, humans and a search-enabled agent

Key result

Frontier LMs achieved only 4.2% to 18.5% accuracy versus 69.7% for humans; a search-and-read agent improved to 35.3% but remained far behind.

Limitation

Shows the limits of identifying the right source even in a bounded scholarly domain; indirect for web visibility.

Why this decision

Shows the limits of identifying the right source even in a bounded scholarly domain; indirect for web visibility.

Source classification and provenance
Study design
Not stated
Evidence stream
Citation attribution
B2B relevance
Low
Source type
Peer-reviewed benchmark paper
Source class
Academic
Transparency
High
Dataset family
CiteME benchmark
Provenance
NeurIPS search for citation attribution
Origin
master
Corrections applied
None
S175Decision: ContextReview: AbstractEvidence: Citation generation and evaluationB2B: Medium

Enabling Large Language Models to Generate Text with Citations

Tianyu Gao, Howard Yen, Jiatong Yu, Danqi Chen; Princeton University / 2023-12 / Academic research

Open direct source
Sample or setting

ALCE benchmark spanning diverse questions and retrieval corpora, with automatic metrics validated against human judgments

Key result

Even the best tested systems lacked complete citation support 50% of the time on ELI5, exposing substantial room between citation display and complete evidence support.

Limitation

Foundational citation benchmark that defines a quality constraint but does not measure publisher visibility.

Why this decision

Foundational citation benchmark that defines a quality constraint but does not measure publisher visibility.

Source classification and provenance
Study design
Not stated
Evidence stream
Citation generation and evaluation
B2B relevance
Medium
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
ALCE citation benchmark
Provenance
Backward citation from citation-bias papers; ACL Anthology
Origin
master
Corrections applied
None
S176Decision: ContextReview: AbstractEvidence: Citation generation and groundingB2B: Low

Training Language Models to Generate Text with Citations via Fine-Grained Rewards

Chengyu Huang, Zeqiu Wu, Yushi Hu, Wenya Wang; Nanyang Technological University / 2024-08 / Academic research

Open direct source
Sample or setting

Fine-grained reward training on ALCE QA datasets with generalization tested on EXPERTQA

Key result

Fine-grained citation rewards improved support, relevance and answer correctness; a 7B model outperformed GPT-3.5-turbo on the tested benchmarks.

Limitation

Useful evidence that citations are an optimizable system behavior rather than a fixed proxy for authority; not a visibility audit.

Why this decision

Useful evidence that citations are an optimizable system behavior rather than a fixed proxy for authority; not a visibility audit.

Source classification and provenance
Study design
Not stated
Evidence stream
Citation generation and grounding
B2B relevance
Low
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
ALCE and EXPERTQA citation training
Provenance
ACL Anthology search for citation grounding
Origin
master
Corrections applied
None
S177Decision: ContextReview: AbstractEvidence: Citation quality measurementB2B: Medium

CiteEval: Principle-Driven Citation Evaluation for Source Attribution

Yumo Xu et al.; AWS AI Labs, Orby.ai, Google / 2025-07 / Academic research

Open direct source
Sample or setting

Multi-domain CiteBench with human annotations and automated metrics evaluated across diverse citation systems

Key result

Fine-grained evaluation using the full retrieval and response context aligned better with human judgments than simple entailment-based support scores.

Limitation

Improves how the meta-study should grade citation quality, but does not estimate brand visibility.

Why this decision

Improves how the meta-study should grade citation quality, but does not estimate brand visibility.

Source classification and provenance
Study design
Not stated
Evidence stream
Citation quality measurement
B2B relevance
Medium
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
CiteBench and CiteEval
Provenance
ACL Anthology search for citation evaluation
Origin
master
Corrections applied
None
S178Decision: ContextReview: FullEvidence: citation volatilityB2B: High for monitoring design.

AI Search Volatility: Why AI Search Results Keep Changing

Profound / 2025-07-17 / Industry / practice research

Open direct source
Sample or setting

Repeated prompt monitoring across major AI platforms; public article reports month-to-month citation drift but keeps exact full panel proprietary.

Key result

AI citations could change by as much as 60% in one month, undermining one-off prompt checks.

Limitation

public sample and volatility denominator are less transparent than newer Sistrix/Semrush studies.

Why this decision

Supports polling-style measurement; public sample and volatility denominator are less transparent than newer Sistrix/Semrush studies.

Source classification and provenance
Study design
Not stated
Evidence stream
citation volatility
B2B relevance
High for monitoring design.
Source type
full vendor longitudinal citation study
Source class
Industry / original-data report
Transparency
Low-moderate - full article, but insufficient bases for a flagship numeric claim.
Dataset family
PROFOUND_AEI_VOLATILITY_2025
Provenance
Official Profound research hub and direct page
Origin
master
Corrections applied
None
S179Decision: ContextReview: FullEvidence: citations versus brand mentionsB2B: High for SaaS measurement concept; weak external validity.

AI Citations vs. Impressions: Ahrefs Brand Study

Ahrefs / 2025 / Industry / practice research

Open direct source
Sample or setting

More than 31,000 Ahrefs brand mentions located in a 150M-prompt Brand Radar database across multiple engines.

Key result

The share of brand mentions that also cited Ahrefs ranged from 10.7% in AI Overviews to 51.6% in Perplexity.

Limitation

Useful platform-mechanics example but a single brand cannot support a general citation-to-mention benchmark.

Why this decision

Useful platform-mechanics example but a single brand cannot support a general citation-to-mention benchmark.

Source classification and provenance
Study design
Not stated
Evidence stream
citations versus brand mentions
B2B relevance
High for SaaS measurement concept; weak external validity.
Source type
single-brand database analysis
Source class
Industry / original-data report
Transparency
Moderate - full article and mention base; database weighting is not actual exposure.
Dataset family
AHREFS_BRAND_RADAR_SINGLE_BRAND
Provenance
AI Visibility Evidence Register v0.1; Existing evidence register; official Ahrefs page
Origin
master
Corrections applied
None
S180Decision: ContextReview: AbstractEvidence: Competitive attack dynamicsB2B: Low

Dynamics of Adversarial Attacks on Large Language Model-Based Search Engines

Xiyang Hu; Michigan State University / 2025-01 / Academic research

Open direct source
Sample or setting

Game-theoretic repeated-prisoner's-dilemma model of publishers choosing attacks

Key result

Theoretical equilibria show that individually attractive manipulation can create collectively worse outcomes, but conclusions depend on stylized payoff assumptions rather than observed markets.

Limitation

Useful competition lens, not primary observational evidence of visibility.

Why this decision

Useful competition lens, not primary observational evidence of visibility.

Source classification and provenance
Study design
Not stated
Evidence stream
Competitive attack dynamics
B2B relevance
Low
Source type
Workshop preprint
Source class
Academic
Transparency
High
Dataset family
Adversarial search game model
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S181Decision: ContextReview: AbstractEvidence: Consumer product recommendation biasB2B: Medium

Gender and Race Bias in Consumer Product Recommendations by Large Language Models

Ke Xu, Shera Potka, Alex Thomo; University of Victoria / 2025-04 / Academic research

Open direct source
Sample or setting

Prompted product suggestions across demographic groups analyzed with marked words, SVMs and Jensen-Shannon divergence

Key result

Recommendation language and product patterns differed significantly by race and gender prompts, indicating that measured brand visibility may depend on simulated persona.

Limitation

Supports persona-stratified monitoring but provides no real purchase or retrieval outcome.

Why this decision

Supports persona-stratified monitoring but provides no real purchase or retrieval outcome.

Source classification and provenance
Study design
Not stated
Evidence stream
Consumer product recommendation bias
B2B relevance
Medium
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
Demographic product-recommendation audit
Provenance
arXiv and Springer record
Origin
master
Corrections applied
None
S182Decision: ContextReview: AbstractEvidence: Content rewritingB2B: Low

Beyond SEO: A Transformer-Based Approach for Reinventing Web Content Optimisation

Florian Luttgenau, Imar Colic, Gervasio Ramirez / 2025-07 / Academic research

Open direct source
Sample or setting

Transformer rewrites evaluated mainly on synthetic travel content with a small extrinsic test

Key result

Model-generated rewrites improved measured generative visibility on synthetic travel pages, but the small external test cannot establish general web effects.

Limitation

Relevant technique with weak external validity and limited real-engine testing.

Why this decision

Relevant technique with weak external validity and limited real-engine testing.

Source classification and provenance
Study design
Not stated
Evidence stream
Content rewriting
B2B relevance
Low
Source type
Preprint
Source class
Academic
Transparency
Medium
Dataset family
Transformer travel-content optimization
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S183Decision: ContextReview: FullEvidence: crawlability and technical optimizationB2B: Medium - prevents wasteful technical prescriptions.

We Ran a Controlled Experiment on Markdown vs. HTML for AI Bots

Profound / 2026-02-17 / Industry / practice research

Open direct source
Sample or setting

381 pages randomized/tested across three weeks; AI-bot traffic compared under Markdown versus HTML delivery.

Key result

The controlled test did not support broad claims that serving Markdown automatically increases meaningful AI-bot traffic.

Limitation

Useful technical debunking but crawler activity is not brand mention, referral or pipeline impact.

Why this decision

Useful technical debunking but crawler activity is not brand mention, referral or pipeline impact.

Source classification and provenance
Study design
Not stated
Evidence stream
crawlability and technical optimization
B2B relevance
Medium - prevents wasteful technical prescriptions.
Source type
controlled technical experiment
Source class
Industry / original-data report
Transparency
High-moderate - full experiment article with page/time bases; site context limits generalization.
Dataset family
PROFOUND_MARKDOWN_HTML_EXPERIMENT_2026
Provenance
Official Profound research hub and direct page
Origin
master
Corrections applied
None
S184Decision: ContextReview: AbstractEvidence: Defense against source manipulationB2B: Low

GRADA: Graph-Based Reranking Against Adversarial Documents Attack

Jingjie Zheng et al.; University of Edinburgh and collaborators / 2025-11 / Academic research

Open direct source
Sample or setting

Graph-based defensive reranking evaluated against adversarial-document attacks

Key result

GRADA substantially reduced attack success with limited accuracy loss, showing that engine defenses can erase tactics that appear effective against undefended rankers.

Limitation

Useful for judging durability of GEO effects rather than a visibility outcome by itself.

Why this decision

Useful for judging durability of GEO effects rather than a visibility outcome by itself.

Source classification and provenance
Study design
Not stated
Evidence stream
Defense against source manipulation
B2B relevance
Low
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
GRADA defense benchmark
Provenance
Martinez 2026 core matrix; ACL Anthology; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S185Decision: ContextReview: FullEvidence: device mix and AI referralsB2B: High - B2B research is desktop-heavy, but this is referral-only.

The Open Frontier of Mobile AI Search

BrightEdge / 2025 / Industry / practice research

Open direct source
Sample or setting

BrightEdge referral data across traditional and generative engines in North America and Europe; brand-site referrals split by device.

Key result

More than 90% of AI referral traffic originated on desktop; ChatGPT was 94% desktop and Perplexity 96.5% desktop.

Limitation

Useful device-bias explanation but absolute bases, dates and sampled brands are not public.

Why this decision

Useful device-bias explanation but absolute bases, dates and sampled brands are not public.

Source classification and provenance
Study design
Not stated
Evidence stream
device mix and AI referrals
B2B relevance
High - B2B research is desktop-heavy, but this is referral-only.
Source type
full industry research article/report
Source class
Industry / original-data report
Transparency
Low-moderate - full article with platform shares; method denominator is opaque.
Dataset family
BRIGHTEDGE_AI_CATALYST_2025
Provenance
Official BrightEdge report index and direct page
Origin
master
Corrections applied
None
S186Decision: ContextReview: FullEvidence: early AI referral penetrationB2B: Medium - cross-site baseline without buyer type.

63% of Websites Receive AI Traffic: Study of 3,000 Sites

Ahrefs / 2025-02-06 / Industry / practice research

Open direct source
Sample or setting

Anonymized Ahrefs Web Analytics sample of 3,000 sites, segmented by traffic size; seven AI chatbots.

Key result

63% received at least one AI visit; the average site got 0.17% of traffic from chatbots and smaller sites had a higher share.

Limitation

retain as longitudinal publication, not an independent dataset.

Why this decision

Valid early wave but superseded by the 81,947-site update; retain as longitudinal publication, not an independent dataset.

Source classification and provenance
Study design
Not stated
Evidence stream
early AI referral penetration
B2B relevance
Medium - cross-site baseline without buyer type.
Source type
full vendor research article
Source class
Industry / original-data report
Transparency
High-moderate - full article and sample; voluntary customer selection persists.
Dataset family
AHREFS_WEB_ANALYTICS_ROLLING
Provenance
Official Ahrefs search; rolling-family deduplication
Origin
master
Corrections applied
None
S187Decision: ContextReview: AbstractEvidence: Financial product recommendation biasB2B: Medium

Exposing Product Bias in LLM Investment Recommendation

Yuhan Zhi et al.; Xi'an Jiaotong University and collaborators / 2025-03 / Academic research

Open direct source
Sample or setting

567,000 generated recommendations across stocks, funds, crypto, savings and portfolios

Key result

Models systematically favored specific assets such as Apple and Microsoft, and the product preferences persisted under tested debiasing methods.

Limitation

Large recommendation-bias test with no human choice or web-retrieval outcome.

Why this decision

Large recommendation-bias test with no human choice or web-retrieval outcome.

Source classification and provenance
Study design
Not stated
Evidence stream
Financial product recommendation bias
B2B relevance
Medium
Source type
Preprint
Source class
Academic
Transparency
High
Dataset family
Investment recommendation bias benchmark
Provenance
arXiv search for LLM product recommendation bias
Origin
master
Corrections applied
None
S188Decision: ContextReview: AbstractEvidence: Fine-grained citation generationB2B: Low

LongCite: Enabling LLMs to Generate Fine-Grained Citations in Long-Context QA

Jiajie Zhang et al.; Tsinghua University / 2025-07 / Academic research

Open direct source
Sample or setting

LongBench-Cite evaluation, automatically constructed LongCite-45k training set and trained 8B/9B models

Key result

Sentence-level citation training improved both citation quality and answer correctness, with the trained models outperforming GPT-4o on the paper's citation benchmark.

Limitation

Benchmark and trained-model results show citation engineering capability, not organic source discovery or publisher outcomes in deployed engines.

Why this decision

Shows citation quality can be engineered independently of organic source selection; benchmark evidence only.

Source classification and provenance
Study design
Not stated
Evidence stream
Fine-grained citation generation
B2B relevance
Low
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
LongBench-Cite and LongCite-45k
Provenance
ACL Anthology search for fine-grained citations
Origin
master
Corrections applied
None
S189Decision: ContextReview: AbstractEvidence: Governance/theoryB2B: Medium

Position: Generative Engine Optimization Creates Underexamined Risks, Governance Must Target Concentration, Disclosure, and Academic Blind Spots

Yizhu Wen et al. / 2026 / Academic research

Open direct source
Sample or setting

Normative synthesis of concentration, disclosure, integrity and research risks.

Key result

Argues that GEO may concentrate visibility power and create incentives that existing search governance does not address.

Limitation

Not primary empirical evidence; useful for limitations and ethics only.

Why this decision

Not primary empirical evidence; useful for limitations and ethics only. Martinez matrix row 44.

Source classification and provenance
Study design
Not stated
Evidence stream
Governance/theory
B2B relevance
Medium
Source type
peer-reviewed position paper
Source class
Prior-review mined
Transparency
High
Dataset family
GEO-GOVERNANCE-POSITION-2026
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S190Decision: ContextReview: AbstractEvidence: High-consideration recommendation biasB2B: Medium

Where Should I Study? Biased Language Models Decide! Evaluating Fairness in LMs for Academic Recommendations

Krithi Shailya et al.; Indian Institute of Technology Madras / 2025-12 / Academic research

Open direct source
Sample or setting

360 simulated profiles and more than 25,000 university recommendations from three open-source LLMs

Key result

Models disproportionately favored Global North institutions, repeated the same institutions and produced gender, nationality and income-related disparities.

Limitation

High-consideration shortlist mechanism is relevant, but education choices and open models do not directly measure B2B vendor discovery.

Why this decision

High-consideration shortlist mechanism is relevant, but education choices and open models do not directly measure B2B vendor discovery.

Source classification and provenance
Study design
Not stated
Evidence stream
High-consideration recommendation bias
B2B relevance
Medium
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
University recommendation fairness benchmark
Provenance
ACL Anthology search for LLM recommendations
Origin
master
Corrections applied
None
S191Decision: ContextReview: AbstractEvidence: LLM ranker manipulationB2B: Low

Are LLMs Reliable Rankers? Rank Manipulation via Two-Stage Token Optimization

Tiancheng Xing et al.; collaborators in the United States / 2026-07 / Academic research

Open direct source
Sample or setting

Two-stage token optimization producing short naturalistic suffixes in simplified ranking contexts

Key result

Natural-looking suffixes could promote targets in LLM rankings, indicating that content-level rank manipulation does not require overt prompt injection.

Limitation

Mechanistically important, but simplified ranker context and adversarial objective limit marketing transfer.

Why this decision

Mechanistically important, but simplified ranker context and adversarial objective limit marketing transfer.

Source classification and provenance
Study design
Not stated
Evidence stream
LLM ranker manipulation
B2B relevance
Low
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
Two-stage rank attack
Provenance
Martinez 2026 core matrix; ACL Anthology; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S192Decision: ContextReview: Landing-pageEvidence: Mechanism/optimizationB2B: High

Shorter, Focused Content Wins in ChatGPT

Kevin Indig / Growth Memo / 2026-04-13 / Industry / practice research

Open direct source
Sample or setting

815,000 query-page pairs from AirOps retrieval research.

Key result

Shorter, tightly focused pages outperformed broad ultimate-guide formats on conditional ChatGPT citation measures.

Limitation

This is a derivative interpretation of the AirOps retrieval dataset, not independent evidence; use the primary AirOps reports for quantitative claims.

Why this decision

This is a derivative interpretation of the AirOps retrieval dataset, not independent evidence; use the primary AirOps reports for quantitative claims.

Source classification and provenance
Study design
Not stated
Evidence stream
Mechanism/optimization
B2B relevance
High
Source type
industry re-analysis
Source class
Prior-review mined
Transparency
Moderate
Dataset family
AIROPS-CHATGPT-RETRIEVAL-2026
Provenance
Cyrus Shepard 54-source evidence sheet; Kevin Indig synthesis of AirOps data.
Origin
master
Corrections applied
None
S193Decision: ContextReview: AbstractEvidence: Personalized recommendation biasB2B: Medium

Stereotype or Personalization? User Identity Biases Chatbot Recommendations

Anjali Kantharuban et al.; Carnegie Mellon University / 2025-07 / Academic research

Open direct source
Sample or setting

Controlled explicit and implicit identity cues across popular consumer LLMs and four US racial groups

Key result

Revealed identity significantly changed recommendations across models, while answers generally failed to disclose that identity had influenced the result.

Limitation

Demonstrates personalization opacity relevant to monitoring, but not brand or traffic outcomes.

Why this decision

Demonstrates personalization opacity relevant to monitoring, but not brand or traffic outcomes.

Source classification and provenance
Study design
Not stated
Evidence stream
Personalized recommendation bias
B2B relevance
Medium
Source type
Peer-reviewed conference paper
Source class
Academic
Transparency
High
Dataset family
Chatbot identity recommendation study
Provenance
ACL Anthology search for chatbot recommendation bias
Origin
master
Corrections applied
None
S194Decision: ContextReview: AbstractEvidence: Popularity bias in recommendationsB2B: Low

Large Language Models as Recommender Systems: A Study of Popularity Bias

Jan Malte Lichtenberg, Alexander Buchholz, Pola Schwoebel; University of Cambridge and collaborators / 2024-06 / Academic research

Open direct source
Sample or setting

MovieLens 10M recommendation task comparing a simple LLM recommender with traditional systems

Key result

The tested LLM recommender displayed moderate popularity bias, generally less than traditional collaborative filtering, and prompting could reduce it further at an accuracy cost.

Limitation

Useful counterexample to the claim that LLM recommendations always amplify incumbents; movie recommendations are distant from B2B buying.

Why this decision

Useful counterexample to the claim that LLM recommendations always amplify incumbents; movie recommendations are distant from B2B buying.

Source classification and provenance
Study design
Not stated
Evidence stream
Popularity bias in recommendations
B2B relevance
Low
Source type
Workshop paper
Source class
Academic
Transparency
High
Dataset family
MovieLens LLM recommender benchmark
Provenance
arXiv search for LLM recommendation popularity bias
Origin
master
Corrections applied
None
S195Decision: ContextReview: AbstractEvidence: Publisher incentives and click economicsB2B: Medium

Do AI Overviews Benefit Search Engines? An Ecosystem Perspective

Yihang Wu et al.; Yale and collaborators / 2026-01 / Academic research

Open direct source
Sample or setting

Game-theoretic creator-effort model calibrated and evaluated with real click data

Key result

The model predicts that AIO traffic diversion can reduce long-run platform profit by weakening creator incentives, while citation or compensation mechanisms can improve outcomes.

Limitation

Uses real click inputs but produces model-dependent ecosystem conclusions rather than a measured market effect.

Why this decision

Uses real click inputs but produces model-dependent ecosystem conclusions rather than a measured market effect.

Source classification and provenance
Study design
Not stated
Evidence stream
Publisher incentives and click economics
B2B relevance
Medium
Source type
Preprint
Source class
Academic
Transparency
High
Dataset family
AIO creator ecosystem model
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S196Decision: ContextReview: FullEvidence: referral traffic, engagement and conversionB2B: High if client is B2B, but industry is undisclosed.

Case Study: Six Learnings About How Traffic from ChatGPT Converts

Seer Interactive / 2025-06-03 / Industry / practice research

Open direct source
Sample or setting

One client GA4 account; October 1, 2024-April 30, 2025; about 11,000 AI sessions versus almost 14M organic sessions; key events as conversions.

Key result

ChatGPT conversion was 15.9% versus Google organic 1.76%; all AI traffic was only 0.07% of organic traffic.

Limitation

Exact bases expose both strong rate and tiny volume, but one anonymous client/event definition cannot be generalized.

Why this decision

Exact bases expose both strong rate and tiny volume, but one anonymous client/event definition cannot be generalized.

Source classification and provenance
Study design
Not stated
Evidence stream
referral traffic, engagement and conversion
B2B relevance
High if client is B2B, but industry is undisclosed.
Source type
single-client first-party case study
Source class
Industry / original-data report
Transparency
Moderate-high - full case method, dates and session totals; client and conversion definition remain confidential.
Dataset family
SEER_SINGLE_CLIENT_AI_GA4_2025
Provenance
Official Seer search
Origin
master
Corrections applied
None
S197Decision: ContextReview: AbstractEvidence: Retrieval and ad inclusionB2B: Medium

Rewrite-to-Rank: Optimizing Ad Visibility via Retrieval-Aware Text Rewriting

Chloe Ho et al.; Carnegie Mellon University and Google / 2025-07 / Academic research

Open direct source
Sample or setting

Offline ad-retrieval pipeline optimized with PPO and evaluated with ranking metrics

Key result

Retrieval-aware rewriting improved ad inclusion, although the absolute mean reciprocal rank gain was small and the system was not a commercial answer engine.

Limitation

Useful upstream retrieval mechanism but indirect evidence for organic AI visibility.

Why this decision

Useful upstream retrieval mechanism but indirect evidence for organic AI visibility.

Source classification and provenance
Study design
Not stated
Evidence stream
Retrieval and ad inclusion
B2B relevance
Medium
Source type
Workshop paper
Source class
Academic
Transparency
High
Dataset family
Rewrite-to-Rank ad benchmark
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S198Decision: ContextReview: FullEvidence: retrieval source alignmentB2B: High for SEO-to-AI transfer assumptions.

AI Search Shift: ChatGPT's Growing Alignment with Google's Index

Profound / 2025-08-06 / Industry / practice research

Open direct source
Sample or setting

Millions of prompts/responses in Profound Answer Engine Insights across ChatGPT and search indexes; public article describes a time-dependent source-overlap shift.

Key result

ChatGPT citation behavior moved materially toward Google's index rather than remaining stable with Bing-like results.

Limitation

Strategically important retrieval change but public method/sample cells are too vague for a strong conclusion.

Why this decision

Strategically important retrieval change but public method/sample cells are too vague for a strong conclusion.

Source classification and provenance
Study design
Not stated
Evidence stream
retrieval source alignment
B2B relevance
High for SEO-to-AI transfer assumptions.
Source type
full vendor longitudinal citation/index study
Source class
Industry / original-data report
Transparency
Low-moderate - full article but exact prompts, periods and overlap denominator are insufficiently public.
Dataset family
PROFOUND_AEI_INDEX_ALIGNMENT_2025
Provenance
Official Profound research hub and direct page
Origin
master
Corrections applied
None
S199Decision: ContextReview: FullEvidence: SaaS signup conversionB2B: High - SaaS signup outcome.

Does AI Search Traffic Convert Better Than Traditional Search? For Ahrefs, Yes

Ahrefs / 2025-06-16 / Industry / practice research

Open direct source
Sample or setting

Ahrefs.com Web Analytics; 30 days ending June 16, 2025; AI referrals versus organic search visits and signups.

Key result

AI referrals were 0.5% of visits but 12.1% of signups, a reported 23x signup-per-visit ratio versus organic search.

Limitation

never use as a benchmark.

Why this decision

Direct B2B SaaS existence proof but one self-selected brand with no absolute counts; never use as a benchmark.

Source classification and provenance
Study design
Not stated
Evidence stream
SaaS signup conversion
B2B relevance
High - SaaS signup outcome.
Source type
single-company first-party case study
Source class
Industry / original-data report
Transparency
Moderate - full case article and window; absolute counts and paid conversion are missing.
Dataset family
AHREFS_FIRST_PARTY_SIGNUPS_2025
Provenance
Existing evidence register; official Ahrefs canonical page verification
Origin
master
Corrections applied
None
S200Decision: ContextReview: AbstractEvidence: Search behavior and task performanceB2B: Medium

ChatGPT vs. Google: A Comparative Study of Search Performance and User Experience

Ruiyun Rayna Xu, Yue Katherine Feng, Hailiang Chen; Miami University, Hong Kong Polytechnic University, University of Hong Kong / 2023-12 / Industry / practice research

Open direct source
Sample or setting

Two randomized online experiments comparing ChatGPT, Google and combined use across information-retrieval tasks

Key result

ChatGPT users completed tasks faster with no overall performance penalty, while combined-tool behavior varied by task.

Limitation

Useful early user-behavior evidence; initial version predates the main window and product capabilities are now stale.

Why this decision

Useful early user-behavior evidence; initial version predates the main window and product capabilities are now stale.

Source classification and provenance
Study design
Not stated
Evidence stream
Search behavior and task performance
B2B relevance
Medium
Source type
Working paper
Source class
Academic
Transparency
Medium
Dataset family
Early ChatGPT-Google experiments
Provenance
SSRN backward search from 2026 user experiments
Origin
master
Corrections applied
None
S201Decision: ContextReview: FullEvidence: shopping offers and purchase routingB2B: Medium-low - useful recommendation-surface analogy.

Which Retailers Does ChatGPT Actually Send Shoppers To?

Profound / 2026-03-26 / Industry / practice research

Open direct source
Sample or setting

22.5M ChatGPT Shopping buy offers over March 10-20, 2026; thousands of prompts repeatedly run; product/title stability assessed.

Key result

Top 20 merchants captured about 40% of offers; 95% of product titles appeared in fewer than 30% of repeated runs.

Limitation

Strong direct AI-commerce shelf mechanics but physical retail routing does not transfer cleanly to SaaS.

Why this decision

Strong direct AI-commerce shelf mechanics but physical retail routing does not transfer cleanly to SaaS.

Source classification and provenance
Study design
Not stated
Evidence stream
shopping offers and purchase routing
B2B relevance
Medium-low - useful recommendation-surface analogy.
Source type
full vendor commerce-analysis article
Source class
Industry / original-data report
Transparency
High-moderate - full article, dates and offer denominator; prompt universe proprietary.
Dataset family
PROFOUND_CHATGPT_SHOPPING_2026
Provenance
Official Profound research hub and full-page inspection
Origin
master
Corrections applied
None
S202Decision: ContextReview: FullEvidence: software buyer behavior and AI investmentB2B: High - direct software buyers; indirect study question.

G2 Buyer Behavior Report 2024

G2 / 2024 / Industry / practice research

Open direct source
Sample or setting

Behavior survey of more than 1,900 B2B software buyers; annual G2 buyer research wave.

Key result

Businesses were actively evaluating AI capabilities and changing software investment behavior; the report predates the sharper AI-search questions in 2026.

Limitation

Useful pre-period buyer baseline but limited direct evidence about AI visibility or recommendation influence.

Why this decision

Useful pre-period buyer baseline but limited direct evidence about AI visibility or recommendation influence.

Source classification and provenance
Study design
Not stated
Evidence stream
software buyer behavior and AI investment
B2B relevance
High - direct software buyers; indirect study question.
Source type
downloadable annual buyer report PDF
Source class
Industry / original-data report
Transparency
Moderate-high - direct full-report PDF with n; detailed instrument requires PDF extraction.
Dataset family
G2_BUYER_BEHAVIOR_2024
Provenance
Official G2 PDF search
Origin
master
Corrections applied
None
S203Decision: ContextReview: FullEvidence: software buying and AI procurementB2B: High - direct B2B software buyers.

G2 Buyer Behavior Report 2025: AI Always Included

G2 / 2025 / Industry / practice research

Open direct source
Sample or setting

Global B2B software-buyer survey reported in the PDF; annual wave focused on budgets, AI functionality and buyer behavior; exact recruitment detail is in the report.

Key result

AI functionality had become a routine requirement, especially for self-identified AI power users; report focuses more on purchasing AI-enabled software than AI-mediated discovery.

Limitation

retain as a separate wave, not replication.

Why this decision

Relevant annual buyer baseline but not the same question as the 2026 AI-search study; retain as a separate wave, not replication.

Source classification and provenance
Study design
Not stated
Evidence stream
software buying and AI procurement
B2B relevance
High - direct B2B software buyers.
Source type
downloadable annual buyer report PDF
Source class
Industry / original-data report
Transparency
Moderate-high - direct full-report PDF; public metadata is less searchable than the 2026 web report.
Dataset family
G2_BUYER_BEHAVIOR_2025
Provenance
Official G2 PDF search
Origin
master
Corrections applied
None
S204Decision: ContextReview: AbstractEvidence: Stealth ranking manipulationB2B: Low

StealthRank: LLM Ranking Manipulation via Stealthy Prompt Optimization

Yiming Tang et al.; Michigan State University and collaborators / 2025-04 / Academic research

Open direct source
Sample or setting

Optimized natural-language suffix attacks against open-model rankers

Key result

Short optimized suffixes could raise target ranks while preserving apparent fluency, exposing a trade-off between stealth and manipulation strength.

Limitation

Relevant security mechanism but not evidence that ordinary content earns production visibility.

Why this decision

Relevant security mechanism but not evidence that ordinary content earns production visibility.

Source classification and provenance
Study design
Not stated
Evidence stream
Stealth ranking manipulation
B2B relevance
Low
Source type
Workshop preprint
Source class
Academic
Transparency
High
Dataset family
StealthRank attack benchmark
Provenance
Martinez 2026 core matrix; arXiv; Olivier Martinez 45-study literature matrix.
Origin
master
Corrections applied
None
S205Decision: HoldReview: Landing-pageEvidence: AI activity versus traditional searchB2B: High market baseline; no B2B cut.

State of Search Q2 2026

Datos and Semrush / 2026 / Industry / practice research

Open direct source
Sample or setting

Large-scale clickstream across the US, EU and UK; desktop activity and search/AI-tool comparisons; full methods gated.

Key result

Public summaries cite AI-tool domains near 1.83% of US desktop activity and 1.91% in EU/UK, but exact denominator and tables need the primary download.

Limitation

Potentially central denominator, but no headline should rely on secondary excerpts until the gated primary report is obtained and extracted.

Why this decision

Potentially central denominator, but no headline should rely on secondary excerpts until the gated primary report is obtained and extracted.

Source classification and provenance
Study design
Not stated
Evidence stream
AI activity versus traditional search
B2B relevance
High market baseline; no B2B cut.
Source type
gated official report landing page
Source class
Industry / original-data report
Transparency
Low at screen - official landing page/full report not yet downloaded; exact measures remain unverified.
Dataset family
DATOS_STATE_OF_SEARCH_Q2_2026
Provenance
Existing research gap; stable official report URL checked
Origin
master
Corrections applied
RDC-0007
S206Decision: HoldReview: FullEvidence: AI traffic and conversion by industry/platformB2B: High potential; B2B subset not disclosed.

Study: The AI Search Landscape, Beyond the SEO vs GEO Hype

Seer Interactive / 2025 / Industry / practice research

Open direct source
Sample or setting

Agency client analytics across several industries; public article gives platform/vertical conversion comparisons but sample counts are not sufficiently disclosed.

Key result

AI platforms showed higher average conversion in several client verticals, but healthcare was an exception; Perplexity averaged 15.9% versus organic 9.77% in the reported set.

Limitation

Useful heterogeneity signal, but unnamed client count, event definitions and weighting prevent a defensible pooled benchmark.

Why this decision

Useful heterogeneity signal, but unnamed client count, event definitions and weighting prevent a defensible pooled benchmark.

Source classification and provenance
Study design
Not stated
Evidence stream
AI traffic and conversion by industry/platform
B2B relevance
High potential; B2B subset not disclosed.
Source type
multi-client agency research article
Source class
Industry / original-data report
Transparency
Low-moderate - full article but critical sample/method details are missing.
Dataset family
SEER_MULTI_CLIENT_AI_LANDSCAPE_2025
Provenance
Official Seer search
Origin
master
Corrections applied
None
S207Decision: HoldReview: AbstractEvidence: B2B SaaS traffic and visibilityB2B: High

AI Search Impact Assessment: The ASIA Framework

Avinash Tripathi; independent practitioner-researcher / 2026-05 / Industry / practice research

Open direct source
Sample or setting

Claimed application across 143 B2B SaaS client properties over June 2024-June 2025

Key result

The paper reports a citation paradox for 73% of cited companies, with mentions rising while organic traffic fell, but public detail is insufficient to verify property selection, controls or duplicated clients.

Limitation

Uniquely on-scope B2B SaaS evidence but methodology and provenance need author clarification before use.

Why this decision

Uniquely on-scope B2B SaaS evidence but methodology and provenance need author clarification before use.

Source classification and provenance
Study design
Not stated
Evidence stream
B2B SaaS traffic and visibility
B2B relevance
High
Source type
Working paper
Source class
Academic
Transparency
Low
Dataset family
ASIA B2B SaaS client panel
Provenance
SSRN search for B2B SaaS AI visibility
Origin
master
Corrections applied
None
S208Decision: HoldReview: FullEvidence: ecommerce AI traffic and conversionB2B: Medium - ecommerce only.

AI Discovery and Ecommerce Visibility Report 2026

Lebesgue / 2026 / Industry / practice research

Open direct source
Sample or setting

First-party observed ecommerce data across 35,000+ analyzed brands; landing page does not disclose session dates, geography, weighting or channel logic.

Key result

90% of brands had measurable AI traffic; reported AI conversion was 3.61% and 68% had an AI-attributed purchase.

Limitation

Very large claimed brand base but method opacity and strong commercial incentive make headline use unsafe pending full report extraction.

Why this decision

Very large claimed brand base but method opacity and strong commercial incentive make headline use unsafe pending full report extraction.

Source classification and provenance
Study design
Not stated
Evidence stream
ecommerce AI traffic and conversion
B2B relevance
Medium - ecommerce only.
Source type
official report landing page
Source class
Industry / original-data report
Transparency
Low - landing-page headline metrics; full report/method not inspected.
Dataset family
LEBESGUE_ECOM_35K_2026
Provenance
Official Lebesgue report search
Origin
master
Corrections applied
RDC-0009
S209Decision: HoldReview: FullEvidence: ecommerce referral conversionB2B: Medium-high as commerce transfer evidence.

The Conversion Advantage: Why ChatGPT Traffic Outperforms Google Search for E-commerce

ThoughtMetric / 2025-06-17 / Industry / practice research

Open direct source
Sample or setting

100 ecommerce stores in May 2025; ChatGPT referral conversion compared with Google Search.

Key result

Average ChatGPT conversion was 6.7% versus Google Search 3.9% across the sampled stores.

Limitation

Clear cross-store headline but store selection, session counts, weighting, uncertainty and attribution logic are not public.

Why this decision

Clear cross-store headline but store selection, session counts, weighting, uncertainty and attribution logic are not public.

Source classification and provenance
Study design
Not stated
Evidence stream
ecommerce referral conversion
B2B relevance
Medium-high as commerce transfer evidence.
Source type
full vendor cross-store analysis
Source class
Industry / original-data report
Transparency
Low-moderate - full article states n/month/rates; no deeper methodology.
Dataset family
THOUGHTMETRIC_100_STORE_ROLLING_2025
Provenance
Official ThoughtMetric search
Origin
master
Corrections applied
None
S210Decision: HoldReview: FullEvidence: prompt intent and buyer journeyB2B: High in concept; no B2B split.

AI Search Intent Study: What 50M+ ChatGPT Prompts Reveal

Profound / 2025-06-25 / Industry / practice research

Open direct source
Sample or setting

Sample from tens of millions of genuine ChatGPT prompts in Profound Prompt Volumes; real-user prompt data, not generated prompt lists.

Key result

The study classed 37.5% of behavior as generative intent and found conversations spanning awareness through consideration/conversion.

Limitation

Rare real-prompt source but the 37.5% construct, sample selection, categories and privacy pipeline are not sufficiently disclosed.

Why this decision

Rare real-prompt source but the 37.5% construct, sample selection, categories and privacy pipeline are not sufficiently disclosed.

Source classification and provenance
Study design
Not stated
Evidence stream
prompt intent and buyer journey
B2B relevance
High in concept; no B2B split.
Source type
full vendor real-prompt analysis
Source class
Industry / original-data report
Transparency
Low-moderate - full article states scale but methodology is too thin for the headline percentage.
Dataset family
PROFOUND_PROMPT_VOLUMES_50M_2025
Provenance
Official Profound research hub and direct page
Origin
master
Corrections applied
None
S211Decision: HoldReview: FullEvidence: referrals, platform share, CTR and visibilityB2B: High for enterprise channel planning.

AI Search Trend Report: What We're Seeing Across Hundreds of Brands

seoClarity / 2026 / Industry / practice research

Open direct source
Sample or setting

Traffic/behavior across monitored brands and top AI platforms January 2025-April 2026; 46M AI visits reported; AI Mode observations begin June 2025.

Key result

ChatGPT supplied about three-quarters of AI referrals; Gemini doubled in a quarter and Claude reached 8% share; report estimates roughly 20 background searches per click.

Limitation

Useful broad enterprise trend line but brand count, denominator, 20:1 model and weighting are not sufficiently documented publicly.

Why this decision

Useful broad enterprise trend line but brand count, denominator, 20:1 model and weighting are not sufficiently documented publicly.

Source classification and provenance
Study design
Not stated
Evidence stream
referrals, platform share, CTR and visibility
B2B relevance
High for enterprise channel planning.
Source type
official vendor research report page
Source class
Industry / original-data report
Transparency
Low-moderate - full report page with period/visit count; critical sample and estimation methods are opaque.
Dataset family
SEOCLARITY_AI_SEARCH_TRENDS_2025_2026
Provenance
Official seoClarity research search
Origin
master
Corrections applied
None
S212Decision: HoldReview: Landing-pageEvidence: Visibility/citationB2B: High

AI Answer Engine Citation Behavior: An Empirical Analysis of the GEO16 Framework

Arlen Kumar and Leanid Palkhouski / 2025-09-13 / Academic research

Open direct source
Sample or setting

70 product-intent prompts, 1,702 citations, 1,100 unique URLs, three engines and 16 English-language B2B SaaS verticals.

Key result

The authors report a strong association between a constructed GEO score and citation probability, including an odds ratio around 4.2.

Limitation

Direct B2B SaaS relevance, but observational design, possible target leakage in the constructed score and unavailable underlying data prevent claim use pending replication.

Why this decision

Direct B2B SaaS relevance, but observational design, possible target leakage in the constructed score and unavailable underlying data prevent claim use pending replication.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
preprint
Source class
Prior-review mined
Transparency
Partial
Dataset family
GEO16-B2B-SAAS-2025
Provenance
Cyrus Shepard 54-source evidence sheet; citation-chain candidate not in Martinez matrix.
Origin
master
Corrections applied
None
S213Decision: HoldReview: Landing-pageEvidence: Visibility/citationB2B: High

The Science of How AI Picks Its Sources

Kevin Indig / Growth Memo / 2026-03-23 / Industry / practice research

Open direct source
Sample or setting

Public summary reports more than 21,000 citations; full methodology is inside the premium report.

Key result

The analysis examines source-selection patterns and the relationship between retrieval and final citations.

Limitation

Potentially valuable Kevin source, but this pass could verify only the public landing page; download the premium report before extraction or claim use.

Why this decision

Potentially valuable Kevin source, but this pass could verify only the public landing page; download the premium report before extraction or claim use.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
premium industry analysis
Source class
Prior-review mined
Transparency
Partial
Dataset family
KEVIN-SOURCE-SELECTION-2026
Provenance
Cyrus Shepard 54-source evidence sheet; Kevin Indig premium synthesis trail.
Origin
master
Corrections applied
RDC-0008
S214Decision: IncludeReview: FullEvidence: Visibility/citationB2B: High

The Science of How AI Pays Attention

Kevin Indig / Growth Memo with Gauge / 2026-02 / Industry / practice research

Open direct source
Sample or setting

Reported 1.2 million AI responses or search results and about 30 million citations through a Gauge partnership.

Key result

Citations were disproportionately drawn from earlier page sections, while selected passages were often taken from within paragraphs.

Limitation

High-potential scale and direct content relevance, but full method and dataset definitions require the premium report before claim use.

Why this decision

High-potential scale and direct content relevance, but full method and dataset definitions require the premium report before claim use. JY source 10; Kevin portal item.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
industry data analysis
Source class
Prior-review mined
Transparency
Partial
Dataset family
GAUGE-KEVIN-ATTENTION-2026
Provenance
JY Scauri review: retained source 10 of 34; Kevin Indig synthesis trail.
Origin
master
Corrections applied
APPENDIX-DELTA-S214
S215Decision: HoldReview: Landing-pageEvidence: Visibility/citationB2B: High

How Do Different AI Models Cite Sources?

Superlines / 2026-01 / Industry / practice research

Open direct source
Sample or setting

Current public page reports 189,439 AI responses across ten major LLM platforms.

Key result

Citation rates differed by roughly two orders of magnitude across platforms, so a single visibility score masks platform behavior.

Limitation

JY described a 62-brand March study, but the resolvable primary page now reports a larger rolling response dataset; lineage equivalence is unconfirmed. Do not use JY's 62-brand statistic without the archived study.

Why this decision

JY described a 62-brand March study, but the resolvable primary page now reports a larger rolling response dataset; lineage equivalence is unconfirmed. Do not use JY's 62-brand statistic without the archived study.

Source classification and provenance
Study design
Not stated
Evidence stream
Visibility/citation
B2B relevance
High
Source type
industry platform comparison
Source class
Prior-review mined
Transparency
Partial
Dataset family
SUPERLINES-CITATION-VARIANCE-ROLLING
Provenance
JY Scauri review: retained source 30 of 34; resolved to current Superlines primary research page, not the exact cited snapshot.
Origin
master
Corrections applied
None
S216Decision: IncludeReview: FullEvidence: Observed AI Mode behaviour and zero-click evaluationB2B: See detail below

What Our AI Mode User Behavior Study Reveals about the Future of Search

Kevin Indig and Amanda Johnson / Growth Memo / 2025-10-06 / Industry / practice research

Open direct source
Sample or setting

37 English-speaking US adults and more than 250 valid desktop task records across seven task types, including one B2B Ramp versus Brex comparison.

Key result

77.6% of sessions had no external visit, the median number of external clicks per task was zero and approximately 88% of first meaningful interactions were with AI Mode text.

Limitation

Thirty-seven US desktop participants, repeated designed tasks and two instructions that deliberately required an external purchase action; not a B2B population benchmark.

Why this decision

Directly observed user behaviour across more than 250 designed task records; relevant to the mechanism by which evaluation occurs without a referral.

Source classification and provenance
Study design
Remote, unmoderated mixed-methods usability study with screen recordings, cursor and scroll events, audio and think-aloud comments.
Evidence stream
Observed AI Mode behaviour and zero-click evaluation
B2B relevance
Indirect mechanism evidence. One task was a B2B comparison, but the participants were not a B2B buyer population.
Dataset ID
DS-KI-AIMODE-UX-2025
Provenance
Tomas Growth Memo Premium access; full article and supplied study deck inspected 2026-08-10.
Origin
delta
Corrections applied
None
S217Decision: IncludeReview: FullEvidence: High-consideration shortlist and choice behaviourB2B: See detail below

How Consumers Navigate High-Stakes Purchases in AI Mode

Kevin Indig / Growth Memo / 2026-04-07 / Industry / practice research

Open direct source
Sample or setting

48 US participants and 185 high-consideration consumer purchase tasks: 149 in AI Mode and 36 in standard search.

Key result

64% of AI Mode tasks had no click. External visits occurred in 23% of AI Mode tasks versus 67% of standard-search tasks. Published shortlist denominators remain inconsistent.

Limitation

Forty-eight US participants and consumer categories; related 88%, 74%, 117/147 and 149-task figures require denominator clarification before headline use.

Why this decision

Rare directly observed comparison of AI Mode and standard search in 185 high-consideration purchase tasks; supports shortlist-compression mechanism.

Source classification and provenance
Study design
Remote, unmoderated, within-subject usability comparison of Google AI Mode and standard Google Search with screen recordings and think-aloud audio.
Evidence stream
High-consideration shortlist and choice behaviour
B2B relevance
Indirect high-consideration mechanism evidence from consumer categories, not a B2B software buyer benchmark.
Dataset ID
DS-KI-HIGHSTAKES-2026
Provenance
Tomas Growth Memo Premium access; full article inspected 2026-08-10.
Origin
delta
Corrections applied
None
S218Decision: IncludeReview: FullEvidence: Cross-engine citation overlapB2B: See detail below

The Consensus Gap

Kevin Indig / Growth Memo with Omnia / 2026-05-11 / Industry / practice research

Open direct source
Sample or setting

3.7 million URL citations, with a random 20,000-prompt headline sample and three additional 5,000-prompt cohorts from a Europe-weighted customer pool.

Key result

2.37% of cited URLs appeared in all three engines for the same prompt, while 91.07% appeared in only one engine.

Limitation

Customer-selected, Europe- and Spain-heavy prompt pool; rule-based classifications and unpinned model versions; citation overlap is not recommendation or buyer exposure.

Why this decision

Large cross-engine citation analysis with a random 20,000-prompt headline sample; supports platform-fragmentation and measurement-design conclusions.

Source classification and provenance
Study design
Repeated cross-engine citation comparison for the same prompts, localized by country and run concurrently against ChatGPT, Perplexity and Google AI answer surfaces.
Evidence stream
Cross-engine citation overlap
B2B relevance
Mixed industries including SaaS and B2B services. Customer-selected prompts are not population weighted.
Dataset ID
DS-KI-OMNIA-CONSENSUS-2026
Provenance
Tomas Growth Memo Premium access; full article inspected 2026-08-10.
Origin
delta
Corrections applied
None
S219Decision: IncludeReview: FullEvidence: Reasoning mode, retrieval and source selectionB2B: See detail below

Reasoning Lift: What Happens to AI Visibility When AI Thinks Harder

Kevin Indig / Growth Memo with Semrush / 2026-05-18 / Industry / practice research

Open direct source
Sample or setting

200 responses across 20 constructed buyer journeys in B2B SaaS, finance, consumer technology and health or lifestyle.

Key result

Citation rate rose from 50% to 68%, average sources per cited response rose from 2.6 to 4.5 and only 25.6% of cited domains overlapped between reasoning settings.

Limitation

Small researcher-designed prompt set, one GPT-5.2 snapshot and no observation of buyer adoption or commercial outcomes.

Why this decision

Controlled 100-prompt paired comparison with designed B2B SaaS journeys; shows that reasoning mode changes retrieval and cited-source pools.

Source classification and provenance
Study design
Controlled comparison of 100 prompts run twice through GPT-5.2, once with minimal reasoning and once with high reasoning.
Evidence stream
Reasoning mode, retrieval and source selection
B2B relevance
Directly includes B2B SaaS as one of four designed categories. It measures model retrieval, not observed buyer adoption.
Dataset ID
DS-KI-SEMRUSH-REASONING-2026
Provenance
Tomas Growth Memo Premium access; full article inspected 2026-08-10.
Origin
delta
Corrections applied
None
S220Decision: IncludeReview: FullEvidence: AI Overview and AI Mode search-session behaviourB2B: See detail below

Users Behave Differently in AI Overviews vs. AI Mode

Kevin Indig / Growth Memo with Surfer and Clickstream Solutions / 2026-05-25 / Industry / practice research

Open direct source
Sample or setting

Approximately 846,000 US Google search sessions collected in February and March 2026 from tens of thousands of users.

Key result

Ten of 15 sectors showed longer sessions with an AI Overview. Among sessions with reverse scrolling, median upward scroll share was 47.5% with an AI Overview versus 27% without one.

Limitation

Observational, incomplete public weighting and model detail, and cursor/scroll proxies rather than purchase outcomes.

Why this decision

Large observational analysis of approximately 846,000 US Google sessions; supports the changing validation/evaluation role of search.

Source classification and provenance
Study design
Observational comparison of US Google sessions with and without an AI Overview.
Evidence stream
AI Overview and AI Mode search-session behaviour
B2B relevance
Mixed-sector behavioral evidence. It supports an evaluation mechanism, not a B2B pipeline benchmark.
Dataset ID
DS-KI-SURFER-CLICKSTREAM-2026
Provenance
Tomas Growth Memo Premium access; full article inspected 2026-08-10.
Origin
delta
Corrections applied
None
S221Decision: IncludeReview: FullEvidence: AI Overview search-session interpretationB2B: See detail below

What to Do Now That AIOs Turned Search into Reading Sessions

Kevin Indig / Growth Memo with Surfer and Clickstream Solutions / 2026-06-01 / Industry / practice research

Open direct source
Sample or setting

The same approximately 846,000 US Google sessions collected in February and March 2026 that underpin S220.

Key result

AI Overview sessions behaved more like reading and validation sessions across several intents. This publication adds interpretation, not an independent evidentiary vote.

Limitation

Same underlying dataset as S220; must contribute no additional independent evidentiary vote.

Why this decision

Decision-relevant follow-up analysis and interpretation of the same 846,000-session dataset; retained as a publication but not an independent dataset.

Source classification and provenance
Study design
Follow-up interpretation of the same observational Google clickstream pool used by S220.
Evidence stream
AI Overview search-session interpretation
B2B relevance
Mixed-sector behavioral evidence from the same pool as S220. It is not a B2B pipeline benchmark.
Dataset ID
DS-KI-SURFER-CLICKSTREAM-2026
Provenance
Tomas Growth Memo Premium access; full article inspected 2026-08-10.
Origin
delta
Corrections applied
None