Research on AI search, with the data attached
Original measurement rather than recycled advice. Every study here ships with its methodology, its limitations, and wherever possible the raw data — so you can check the numbers instead of taking them on trust.
Featured
The state of AI search: September 2026
A snapshot of where AI-search research actually stands as this site's research program launches. Five findings, five unanswered questions, and an honest inventory.
AI search CTR statistics
How click-through behavior changes when an AI Overview appears. The specific population each figure describes stays attached, not stripped away.
Everything, grouped by what it's about
All 76 pieces of research, under the three areas this site works in. The pre-registered study protocols have their own index.
GEO / AI Search
Getting cited by ChatGPT, Perplexity, Gemini & AI Overviews.
The AI citation dataset strategy
Every dataset this site plans to publish, in one place: what it contains, when it ships, how it is licensed. Raw files, not just summary reports.
What is AI SEO?
The umbrella term for optimizing a site's visibility across the whole AI-search landscape: generative answers, AI crawlers, and the technical infrastructure behind both.
What is GEO?
Optimizing content so AI systems like ChatGPT, Perplexity and Gemini cite it when generating an answer. That is Generative Engine Optimization.
AI search glossary
Definitions for the terms used throughout this site's research. Kept tight, each linked to the fuller page where the concept is explored.
AI SEO statistics
Every AI-search statistic on this site, in one table, each with a verification grade and a link to full context.
AEO statistics
Answer Engine Optimization statistics. Genuinely fewer exist under this specific label than under GEO, and this page says so rather than padding itself.
GEO statistics
The most commonly cited Generative Engine Optimization statistics, each with a verification status attached rather than uniform, unearned confidence.
The Null-Results Registry
Every pre-registered prediction on this site that is, or is predicted to be, a null result, tracked in one place. Almost nobody publishes negative findings.
WebMCP and the agent-readable web
Google and Microsoft shipped an early preview of a structured agent-interaction protocol for websites in Feb 2026. What it is, and why site owners should care now.
MCP servers as a visibility channel
The Model Context Protocol lets AI agents query a business data directly. Does publishing one affect AI-search visibility? Untested, but genuinely early.
The AI Visibility Measurement Standard
'AI visibility' currently means whatever a vendor's dashboard measures. A proposed open specification so numbers from different sources become comparable.
GEO vs AEO vs LLMO: which term is winning
Three acronyms describe overlapping ideas about optimizing for AI-generated answers. The industry hasn't settled on one. What each means, where they overlap, and which is gaining ground.
The GEO Tactic Evidence Scoreboard
Every tactic the industry recommends for getting cited by AI search, graded by the actual evidence behind it. Not by how often it is repeated.
Bing Copilot SEO: the easiest AI engine to crack
ChatGPT search runs on Bing's index. Getting into Bing is the fastest path to your first AI citation. Here's the full playbook.
How to get cited by ChatGPT: 9 tactics, with data
The Princeton-backed levers - statistics, quotes, structure, freshness - that actually move AI citations.
Technical SEO
Schema, llms.txt, crawlers, IndexNow - the plumbing AI reads.
Open-source SEO agent tooling: what to build
The citation-collection harness, log-analysis pipeline and audit agents behind this site's own research are more credible open than closed.
Claude Code SEO failure modes
Where agentic SEO work actually goes wrong. Almost all existing coverage is promotional. This is the first-party failure catalogue instead.
Claude Code for SEO: the complete guide
Using Claude Code as an execution agent for technical SEO: audits, schema generation, structural fixes, log analysis. With a worked example and safety practices.
AI crawler statistics
How much of the web's bot traffic is AI. Which crawlers take the most relative to what they give back, and how that's shifted year over year. Verification status on every figure.
The AI Bot User-Agent Registry
Every AI crawler that matters for a website owner in one place: what it's for, how to verify it, and what to put in robots.txt. Maintained, not a one-off listicle.
GPTBot vs OAI-SearchBot: the AI crawler guide
The most-confused pair in AI SEO. One is a training crawler, one gets you cited. Plus every other AI bot and exactly what to allow or block.
The 50-point technical GEO audit
A complete technical checklist to make a site under 500 pages fully legible to Google, Bing and the AI answer engines.
Data & Studies
Original experiments and numbers you can't get elsewhere.
The state of AI search: September 2026
A snapshot of where AI-search research actually stands as this site's research program launches. Five findings, five unanswered questions, and an honest inventory.
AI search CTR statistics
How click-through behavior changes when an AI Overview appears. The specific population each figure describes stays attached, not stripped away.
AI search and YMYL content: what is actually known
Health, finance, legal and safety queries carry higher stakes. Do AI engines cite differently for them? That is an open, under-studied question.
Local AI search: what is actually known
Local intent is a large share of all search behavior. Rigorous AI-search research on it is almost entirely absent. Stated plainly, not filled with a borrowed number.
AI shopping and agentic commerce statistics
AI agents are beginning to complete purchases on a shopper's behalf. What is actually measured, and how much of the coverage is still forecast rather than fact.
Agentic browsers: what do they actually fetch?
ChatGPT Atlas, Perplexity Comet and Gemini in Chrome are shipping now. What they actually request from a page has barely been studied. Pre-registered.
Do AI crawlers render JavaScript?
Retrieval bots may not execute client-side JavaScript. If not, any content that depends on it is invisible to them entirely. Pre-registered experiment.
AI Mode vs. AI Overviews: source delta
Two distinct Google AI surfaces, routinely treated as interchangeable. Do they actually cite the same sources for the same query? Pre-registered study.
The Reddit dependency: what Perplexity cites when Reddit has nothing
Perplexity's citations skew heavily to Reddit. For B2B and enterprise topics with thin discussion, what fills the gap? Pre-registered study.
The Wikipedia dependency: what ChatGPT cites when Wikipedia has nothing
ChatGPT's citations skew heavily to Wikipedia. For topics it doesn't cover well, what fills the gap? Pre-registered study.
Brand mentions vs. backlinks: which predicts AI citation?
Reported correlations put brand mentions at roughly 3x the predictive strength of backlinks for AI Overview visibility. A confound nobody has controlled for.
Does an author bio affect AI citation?
Author credentials and E-E-A-T signals are recommended constantly. No controlled test isolates whether they change AI citation rate. A pre-registered RCT.
Does content freshness increase AI citations?
'Keep content updated' is universal GEO advice. No controlled study has tested it against citation rate. A pre-registered RCT.
Crawl-to-citation latency
How many days pass between an AI crawler first fetching a new page and that page appearing in a citation? Pre-registered panel study.
The Query Fan-Out Corpus
Everyone repeats that Google AI Mode fans a query into '8-16 sub-queries.' Nobody has published the data behind that number. A pre-registered corpus design.
Cross-platform citation concordance
How much do ChatGPT, Perplexity, Gemini, Claude and Google agree on what to cite? A reported ~11% two-engine overlap suggests: not much. Pre-registered study.
AI citation half-life: how long do citations last?
Once a URL is cited by an AI engine, how long does it stay cited? Nobody has measured this. A 12-month tracking study, pre-registered.
Does schema markup increase AI citations?
The most contested tactic in GEO. This page registers a randomized test to settle it. Pre-registered before a single page is enrolled.
AI search conversion benchmarks
Published claims about AI traffic conversion range from 4.4x to 23x organic. Different orders of magnitude. Here is why, and the honest version of the statistic.
AI referral traffic statistics
How much traffic AI platforms actually send to websites, how it splits between engines, and why standard analytics likely undercount it.
Zero-click search statistics
The most commonly quoted zero-click figures, each attached to the specific population it actually measures. Collapsing them into one number is where most content goes wrong.
AI search market share statistics
How traffic is distributed across AI search platforms, how that's shifted over the past year, and how AI referral engagement compares to classic organic.
Gemini citation statistics
Gemini, AI Overviews, and AI Mode are three related but distinct Google surfaces, routinely conflated in coverage. What is actually known about Gemini specifically.
Perplexity citation statistics
Perplexity is the most transparent of the major AI search engines about its sources. Its citation pattern looks meaningfully different from ChatGPT.
Claude citation statistics
What is actually known about what Claude cites, and an honest accounting of how much isn't known. Claude is the least-studied major AI search surface.
ChatGPT citation statistics
What ChatGPT actually cites, how it retrieves information, and how weakly Google ranking predicts showing up in a ChatGPT Search answer.
Most-cited domains in AI search
Which kinds of sources dominate citations across ChatGPT, Perplexity and Google AI Overviews. And how little the major engines actually agree with each other.
Most-cited domains in ChatGPT search
Two public ChatGPT domain lists, kept apart. Profound's share of every citation, through June 2025, is led by Wikipedia at 7.8%. Ahrefs' US mention share of the top sources, September 2026, is led by Reddit at 16.8%. Those are not the same ranking.
Most-cited domains in Perplexity
Reddit leads the large US Perplexity samples, and the famous 46.7% is its share of the top 10, not of all citations. The same Profound row is 6.6% of every Perplexity citation. By September 2026, Ahrefs' US top-source list had Reddit at 21.6% and YouTube at 20.8%.
Most-cited domains in the Gemini app
The Gemini app has its own US list, and it is not AI Overviews. In Ahrefs' September 2026 snapshot, Reddit is 28.5% of the top-source pool, YouTube 14.6%, and English Wikipedia 8.8%. google.com is not in that top 50.
Most-cited domains in Google AI Overviews
On Ahrefs' US snapshot updated 2 September 2026, YouTube is the leading AI Overview source at 22.9% of the top-source pool, then Reddit, Facebook, and google.com. Individual news domains sit under 1%. This is a platform list, and it is not the Gemini app.
Most-cited domains in Google AI Mode
AI Mode is not AI Overviews. In Ahrefs' US snapshot updated 2 September 2026, Reddit is 17.9% of the top-source pool and YouTube is 17.8%. google.com is third at 12.5%. Other 2026 studies, counting every citation, put google.com first and disagree about what those Google links are.
Most-cited domains in Claude
There is no public large-sample ranking of domains Claude cites. What has been measured is the mix: brand domains about 64 to 66 percent, news or media about 15 to 24 percent, and encyclopedia plus social in the low single digits. A 75-prompt name list is a probe, not a leaderboard.
Most-cited domains in Microsoft Copilot
Copilot's public US list is a shopping index. Ahrefs, September 2026, over 3 million queries: Amazon is 16.8% of the top-50 pool, Walmart 12.6%, English Wikipedia 7.6%. Reddit is not in that top 50. Profound had ranked Reddit 31st on Copilot in November 2025.
Domains cited by every AI engine, and domains cited by only one
On prompts where ChatGPT, Gemini, Perplexity, and AI Overviews all cited something, Writesonic found 3.8% of domains on all four engines and 72 to 73 percent on exactly one. Profound's two-engine pool, ChatGPT and Perplexity, put the shared slice at 11%. Same-day agreement is rarer than agreement across a month.
Most-cited news domains in AI search
There is no public ranking whose only job is the top news domains in AI search. Forbes is the news brand that keeps appearing near the top of ChatGPT's mixed list. A widely quoted 12% is the rate at which a retrieved news URL becomes a citation, not news's share of citations.
Most-cited documentation domains in AI search
No opened study ranks documentation domains, or even documentation's share, for each engine on one sample. On ChatGPT, Promptwatch watched documentation rise from 3.3% to 8.2% of classified citations during August 2026. Product pages were the large format, near 29%.
Most-cited review domains in AI search
G2 and Capterra received zero citations in one 40-category ChatGPT study, and low single-digit rates in wider counts. G2's often-quoted 84% is a share inside review platforms, not a share of all citations. The review-site list is short because the shares are small.
Most-cited community domains in AI search
Reddit is the community domain the studies actually quantify. Quora's large shares belong to Google AI Overviews and AI Mode, not to Perplexity. Stack Exchange does not appear in the published Perplexity top 50. Social is not a synonym for forums.
Most-cited video domains in AI search
YouTube is the video domain, and its share depends on the engine. In one US vendor's September 2026 top-source lists it runs from 22.9% on AI Overviews to 1.9% on Copilot. TikTok is a smaller, real row on the Google surfaces. A 73% timestamp figure is a split between two Google products, not a share of citations.
Most-cited reference domains in AI search
Wikipedia is the reference domain the studies measure, and its share is not one number. Dictionaries and a few medical .gov sites are the rest of the visible layer. No opened study ranks what gets cited when Wikipedia has no article.
Most-cited SaaS vendor domains on commercial prompts
There is no public top 10 of SaaS domains with market-wide citation shares. The measured split is whether the prompt names the company. Its own site is about 29% of sources when it does, and about 6% when the prompt only names the need.
Most-cited pages in AI search, not domains
No one has published a top 10 of URLs with citation shares. The closest public cut is the type of page inside ChatGPT's 1,000 most-cited pages in September 2025. Wikipedia pages were 29.7% of that thousand. The URLs were not printed.
The long tail of AI citations
Most cited domains appear once. In one six-engine panel, 46.9% of 23,280 domains were observed a single time. The claim that 15 websites own 68% of citations does not survive the wide counts, where a single ChatGPT domain stays under 4%.
Which most-cited domains block AI crawlers
Nobody has joined an Ahrefs or Profound top 50 to robots.txt. The only opened join is ten Google AI Mode news queries. Of the cited domains inside the crawl sample, 51.9% disallow at least one AI crawler, against 15% of the sample.
Which most-cited domains publish original data
No one has audited the most-cited domains for original data. On marketing and SEO queries, the cited URL classified as original research is about 0.6% of citations. The cited page is usually a how-to, a product page, or a listicle.
How the most-cited domain list changes from quarter to quarter
A full quarter-to-quarter league table has not been published. The repeated series that moves is Reddit's share of ChatGPT citations: 6.11% in May 2026, then 0.52% in mid-August, a drop the tracker calls provisional. Ahrefs still had Reddit first on 2 September, on a different denominator.
Google AI Mode statistics and query fan-out, explained
How AI Mode actually retrieves information: the query fan-out mechanism, its documented patent basis, and what is confirmed versus estimated.
Google AI Overview statistics
Trigger rates, citation-to-ranking overlap, and click-through impact for AI Overviews - with a verification status attached to every figure.
The robots.txt AI-Blocking Census
How many of the web's most important sites block AI crawlers, measured directly from robots.txt, not quoted from a report with no visible sample. Pre-registration.
Where AI SEO statistics actually come from
We traced twelve of the most-quoted numbers in AI search back to their origin. Five hold up. Three dead-end. The most-repeated figure in the field is three studies wearing one number.
The AI Citation Index
A quarterly, open measurement of what seven AI search engines actually cite. Query set, raw data and methodology published in full. This is the pre-registration.
Zero to Cited: how a new site climbs into AI search
The evidence-backed timeline and playbook for taking a brand-new site from launch to its first citations in Bing, Perplexity, ChatGPT and AI Overviews.
Does llms.txt actually work? A 90-day log test
I deployed llms.txt and read the server logs for 90 days. Here is what the data, and the largest studies, actually say.
One experiment. Every week.
The field notes in your inbox - one thing I tested, the raw numbers behind it, and what it means for getting cited by AI.