Research

The evidence base for AI search visibility.

A curated index of the primary studies, large-scale data analyses, and peer-reviewed research on how AI engines retrieve, cite, and recommend brands. Organised by research question. Updated as new evidence emerges.

A note on evidence quality

The AI search visibility space has a significant noise problem. Many widely cited statistics come from vendors with a commercial interest in the findings, thin sample sizes, or secondary reporting that has drifted from the original source. This index distinguishes between primary research, vendor-commissioned studies, and claims that lack reliable sourcing.

The strongest evidence in this space comes from a small number of large-scale studies with clear methodology and no direct vendor incentive. The Princeton KDD paper, the Ahrefs brand correlation studies, and the Seer Interactive citation analysis are the three most methodologically rigorous sources available as of mid-2026.

Evidence key
Primary research — peer-reviewed studies, large-scale controlled analyses, or data from platforms with no direct incentive to bias the finding.
Vendor research — primary data collection by a company with a commercial relationship to the finding. Directionally useful but should be read with the incentive structure in mind.
Weak or unverified — widely circulated claims that lack primary sourcing, have thin sample sizes, or originate from secondary reporting of secondary reporting.
Research question 01
What signals predict whether an AI engine cites or recommends a brand?

The most important question in the field. The answer reshapes how visibility strategy should be built — and it differs significantly from what traditional SEO optimises for.

Ahrefs
August 2025 — December 2025
AI Overview Brand Visibility Factors — 75,000 Brands
Branded web mentions correlate with AI Overview visibility at 0.664, versus 0.218 for backlinks. The December follow-up found YouTube mentions correlate at 0.737 — the strongest single signal measured across both studies. The top three correlating factors are all off-site brand signals.
The most comprehensive brand-level correlation study available. The 3:1 ratio between mentions and backlinks is the single finding that most clearly separates AI search strategy from traditional SEO.
75,000 brands · Primary research
View study →
Seer Interactive / Trustpilot
March 2026
How Review Profiles Shape Brand Presence in AI Search — 804,491 Responses
Brands with no active review profile appear in just 1% of AI answers. Brands with actively managed profiles of 80 or more reviews appear in 75.3%. Review and trust sites account for 14% of all AI citations, rising to 24% at the buying intent stage.
The largest study on review signals and AI citation rates conducted to date. Note: commissioned by Trustpilot, a review platform with a commercial interest in the finding. The sample size and methodology are nonetheless robust.
804,491 responses · 1,926 brands · Vendor research
View study →
Muck Rack
May 2026 — 3rd edition
What Is AI Reading? Generative Pulse
Earned media accounts for 84% of all AI citations across ChatGPT, Claude, and Gemini. Non-paid sources account for 94%. Paid and advertorial content accounts for 0.3%. Consistent across three separate editions of the same research.
The consistency across three editions is what makes this finding credible despite the vendor context. Muck Rack is a PR platform with an interest in earned media appearing valuable — but the directional finding has been replicated independently.
25M+ citations · 17 industries · Vendor research
View study →
Research question 02
What content changes move AI citation rates?

The only controlled experiment available in this space. Everything else is correlational. The Princeton findings are the closest thing to causal evidence for on-page optimisation.

Princeton University · IIT Delhi · Georgia Tech · Allen Institute for AI
Published KDD 2024 · Preprint November 2023
GEO: Generative Engine Optimization
Adding statistics improves AI citation probability by +41%. Expert quotations add +28%. Citing sources inline adds +30–40%. Keyword stuffing performs −10% below baseline — actively harmful. Combining statistics with fluency optimisation produces the strongest combined effect.
The foundational controlled experiment in the field. The only study that isolates specific content changes and measures their causal effect on AI citation rates. Every other content recommendation in this space is either derivative of this study or correlational.
~10,000 queries · GEO-bench · Peer-reviewed
View study →
SE Ranking
November 2025 — updated March 2026
What Drives ChatGPT Citations — 129,000 Domain Study
Pages with FAQ schema markup averaged 3.6 citations versus 4.2 for pages without. Optimal section length is 120–180 words, producing approximately 70% more citations than longer unbroken sections. High domain trust scores correlate with approximately 4x more citations.
One of the largest domain-level citation studies conducted. The FAQ schema finding is counterintuitive and significant — structured data does not reliably increase AI citation rates for this format.
129,000 domains · 100,000 prompts · Vendor research
View study →
Cyrus Shepard · Zyppy Signal
May 2026
54-Study Meta-Analysis of AI Citation Factors
URL accessibility scores 9.5/10 as the top AI citation factor. Traditional search rank scores 9.4/10. LLMs.txt scores 2.0/10 — the lowest of 23 factors tested. The meta-analysis covers ChatGPT, Gemini, and Perplexity across 54 experiments, patents, and case studies.
The most useful synthesis of the available evidence weighted by repeatability and methodological rigour. The scoring system makes it easier to prioritise actions than reading individual studies in isolation.
54 studies synthesised · Meta-analysis
View study →
Research question 03
How volatile are AI recommendations and what does that mean for measurement?

Understanding volatility is essential before measuring anything. Position tracking is largely meaningless in a probabilistic system. These studies establish the baseline for why frequency-based measurement is necessary.

Rand Fishkin · SparkToro · Patrick O'Donnell · Gumshoe.ai
January 2026
AI Recommendation Consistency Study
Only ~30% of brands appearing in an AI response appear again on the next run of the same prompt. Run the same prompt five times consecutively and just 20% of brands persist across all five. However the consideration set — the pool of brands that could appear — is stable, with top brands appearing in 55–77% of responses.
The clearest evidence that tracking position in any single AI response is meaningless. The consideration set stability finding is equally important — it explains why frequency-based measurement across many runs produces stable, actionable data.
2,961 prompt tests · 600 volunteers · 4 engines
View study →
Kevin Indig · Growth Memo
2026
The Ghost Citation Problem — 3,981 Domains
74.9% of domains are cited by AI engines but only 38.3% are mentioned by name. 61.7% of citations are ghost citations — AI links to a domain without naming the brand. Only 13.2% of cited brands receive both a link and a name mention. Gemini names brands in 83.7% of citations but links in only 21.4%.
The ghost citation finding has significant implications for how brand visibility is measured. Citation tracking that only counts links misses the majority of the competitive picture.
3,981 domains · 115 prompts · 4 engines
View study →
Profound
2026
AI Citation Volatility Analysis
40–60% of cited domains change month to month for identical queries across ChatGPT. Turn 1 of a conversation is 2.5x more likely to trigger citations than Turn 10. Analysis drawn from approximately 240 million ChatGPT citations and 730,000 conversations.
The month-to-month volatility finding establishes why the AI Appearance Rate must be tracked monthly rather than treated as a fixed score. Model updates can significantly shift citation patterns even without changes to the brand's own signals.
240M citations · 730,000 conversations · Vendor research
View study →
Research question 04
How does content freshness affect AI citation rates?

Freshness preference varies significantly between engines. Perplexity is the most recency-biased. Google AI Overviews are the exception — they favour slightly older, more authoritative content.

Ahrefs
2026
Do AI Assistants Prefer to Cite Fresh Content? — 16.975M URLs
AI-cited content is 25.7% fresher than content appearing in the organic Google top-10 for the same queries. The average AI-cited URL is 1,064 days old versus 1,432 days for organic top-10 results. ChatGPT shows the strongest freshness preference. Google AI Overviews is the exception — it cites slightly older content.
The largest freshness study available. The per-engine variation is the most actionable finding — strategy should differ based on which engine a brand is prioritising.
16.975M cited URLs · 7 platforms · Primary research
View study →
GrowByData · ConvertMate
2026
Content Freshness and ChatGPT Citation Rate
Content updated within 30 days earns approximately 3.2x more ChatGPT citations than content not recently updated. 76.4% of ChatGPT's most-cited pages were updated within 30 days. Content 90 or more days without update sees 40–60% citation rate drops.
Corroborated by multiple independent vendor analyses. The 30-day threshold is the most actionable finding for content refresh strategy, though the causal mechanism is update recency rather than cosmetic date changes.
Multiple datasets · Vendor research
View study →
Ahrefs
2026
Schema Markup and AI Citation Rates — Controlled Study
Tracking 1,885 pages that added JSON-LD schema against 4,000 matched controls over seven months found no meaningful citation uplift on any AI platform. Google AI Overviews actually showed a −4.6% decline — the only statistically significant result.
The most important negative finding in the field. Schema markup is useful infrastructure for entity clarity and rich results but is not a reliable AI citation lever. This study directly contradicts a significant amount of GEO advice circulating in the industry.
1,885 pages · 4,000 controls · Primary research
View study →
Research question 05
What is the scale and commercial impact of AI search visibility?

Understanding the size of the opportunity and the downstream commercial effects of AI visibility is essential context for any brand investment in this area.

Victorious
Q1 2026
Q1 2026 Quarterly Search Report — AI Mention Rate
Of 177 brands tested across 8 AI platforms, 89.8% had zero AI mentions in Q1 2026. Only 18 brands in the dataset had any AI mention rate above zero.
Small sample but the direction is consistent with other visibility studies. The 89.8% figure is the most-cited baseline for AI search invisibility and is directionally reliable despite the limited sample.
177 brands · 8 platforms · Vendor research
View study →
Semrush
June 2025
Impact of AI Search on SEO Traffic
The average AI search visitor is 4.4x as valuable as the average traditional organic visitor across 500 high-value topics. Brands cited in AI Overviews earn approximately 120% more organic clicks per impression than uncited brands on the same queries.
Vendor research from a company selling AI visibility tools. The 4.4x figure is frequently cited but should be treated as directional rather than precise given the commercial context.
500+ topics · Vendor research
View study →
SparkToro · Similarweb
June 2026
Zero-Click Search and the Open Web — 2026
In the first four months of 2026, 68.01% of Google searches ended without a click to any website — up from 60.45% in 2024. Only 276 of every 1,000 Google searches reach the open web. AI search is accelerating the structural shift away from click-based discovery.
The clearest evidence of the structural shift driving AI search relevance. The zero-click trend is the macro context within which all AI visibility strategy operates.
US Google searches · Jan–Apr 2026 · Primary research
View study →
Claims to treat with caution

Widely circulated statistics in this space that lack reliable primary sourcing.

"77% of brands are invisible to AI search"
Originates from Loamly and PRWeb, February 2026. No methodology published, no sample size disclosed, no primary data accessible. Circulates widely in trade press despite thin sourcing.
"llms.txt improves AI search visibility"
Google confirmed in June 2026 that Google Search does not use llms.txt. Ahrefs telemetry found 97% of llms.txt files received zero AI requests in May 2026. The file has legitimate uses for coding agents but is not an SEO or GEO lever.
"FAQ schema boosts AI citation rates"
SE Ranking's 129,000-domain study found pages with FAQ schema averaged 3.6 citations versus 4.2 without. Google deprecated FAQ rich results in May 2026. Schema remains useful infrastructure but is not a reliable citation lever.
Apply the research

See how this research applies to your brand's AI visibility.

The free scanner applies the signal framework from this research to your specific brand and category — returning your AI Appearance Rate, technical signal audit, and a view of which competitors are appearing instead.

Run Your Free AI Visibility Scan → Read the AAR Methodology