What the evidence says about how AI search finds, cites and recommends brands. Every number sourced, dated, and given a note on how far I'd trust it. Updated as new research lands.
Last updated 13 October 2026
Most research on AI search is young, small, or published by a company selling a product. That doesn't make it worthless, but it does mean you should know who ran it and what they measured. I read it so you don't have to, and I say plainly when I think a finding is solid and when it isn't.
GEO, AEO and AI search mean roughly the same thing here: how brands appear in answers from ChatGPT, Google's AI Overviews and AI Mode, Gemini, Claude, Perplexity and the rest. New to it? Start with What is GEO?
The numbers worth knowing
The figures I'd quote in a meeting right now, with the catch for each.
Figure
What it says
The catch
Source
84%
of AI citations are earned media, across 25m+ links in ChatGPT, Claude and Gemini.
Broad definition. Journalism alone is 25–27%. Other studies sort the data differently and get much lower figures.
Muck Rack, May 2026
30.5%
of searches where Google's AI Mode and AI Overviews both cited sources, but shared none.
400 US searches, one run, by a GEO agency; method published and corrected.
Underneath, Oct 2026
51%
of US adults say they never use AI chatbots.
Many still see Google's AI answers without thinking of it as AI.
Pew Research, Oct 2026
1 in 4
recommended brands shared between answers for different personas, vs 2 in 5 for the same persona.
Vendor research; three consumer categories.
Profound, Sep 2026
88%
of citations at the B2B evaluation stage point to content brands control.
One agency's client base; evaluation-stage prompts only.
Ten Speed, Sep 2026
1.7%
of question-and-website pairs where ChatGPT, Claude, Gemini and Perplexity all cited the same site.
72 prompts, three B2B brands; directional.
Orbit Media, Sep 2026
+25%
brand visibility from keyword-style prompts; ranking-style prompts +20%.
Vendor study; middle-of-funnel prompts most sensitive.
Peec AI, Jun 2026
<1 in 100
chance ChatGPT or Google's AI gives the same list of brands twice for the same prompt.
12 prompts, ~3,000 runs. Co-authored by a firm that sells AI tracking.
SparkToro / Gumshoe, 2026
38%
of Google AI Overview citations came from top-ten pages, down from ~76% in mid-2025.
Google only. 31% came from pages outside the top 100.
Ahrefs, Mar 2026
88%
of Google AI Mode citations aren't in the organic top ten for the same query.
One Google product; US and UK queries.
Moz, Feb 2026
30%
of brands stay visible from one AI answer to the next. 20% hold across five runs.
Vendor research; methodology only partly published.
AirOps, 2026
68%
of brand mentions in AI search appear in only one AI tool.
As above.
AirOps, Mar 2026
~2x
more likely to lose AI citations if a page hasn't been updated in a year.
Correlation, not proof of cause.
AirOps, 2026
7 days
for half of newly published pages that get cited to be cited. 90% within 37 days.
Small dataset; ChatGPT and Claude only. Measures the whole pipeline, not just content.
Profound, May 2026
68%
of US Google searches ended without a click, up from 60% in 2024.
US data. About 60% fewer clicks when an AI Overview shows.
SparkToro, 2026
~4%
of people who get news from AI chatbots click through to the source.
News use specifically, not brand research.
Reuters Institute, Jun 2026
51%
of B2B software buyers start their research in an AI chatbot rather than Google.
Survey of software buyers; G2 sells to that market.
G2, 2026
900m
weekly ChatGPT users.
OpenAI's own figure. Weekly users aren't all asking about brands.
OpenAI, Feb 2026
No lift
in AI citations from adding JSON-LD schema to 1,885 pages, against 4,000 controls.
Pages that were already being cited.
Ahrefs, May 2026
Full references sit with each entry below.
Latest research, newest first
Added October 2026 · Google
Google has two AI answers, and they don't agree
30.5%of searches with sources in both had no cited URL in common
0.14average URL overlap between AI Mode and AI Overviews (0 = none, 1 = identical)
51%of US adults say they never use an AI chatbot
Underneath ran the same 400 US searches through Google's AI Overviews and AI Mode on the same day. Where both cited sources (236 searches), 30.5% shared no URL at all, and they led with the same URL only about a fifth of the time. AI Overviews leaned more on forums, social and video. Separately, Pew found 51% of US adults say they never use AI chatbots, even though AI Overviews now sit on a large share of Google searches and lead into AI Mode with a follow-up question.
My take. "Are we in Google's AI answer?" is now two questions. And many of the people reading those answers don't think they're using AI at all, which is exactly why it matters for reputation. More in Cited, edition 8.
84% vs 88%earned media overall, vs brand-controlled content at the B2B evaluation stage
1 in 4brands shared between answers for different personas
1.7%of cases where four AI tools cited the same site for the same question
A run of studies from the summer of 2026 shows how much AI answers depend on how you ask. Across all kinds of questions, most citations point to earned media. But at the stage where B2B buyers are comparing options, Ten Speed found 88% of citations pointed to content brands control, and BrightEdge found brand-owned pages took 42–79% of consideration-stage citations across eight industries. Profound found that changing the persona attached to a question cut the overlap in recommended brands from about two in five to one in four. Peec AI found the format of a question shifted brand visibility by 20–25%, with consideration-stage questions the most sensitive. And Orbit Media found four AI tools agreed on the same cited site for the same question in just 1.7% of cases.
My take. This is the strongest argument against "type your brand name in and see what happens". Different stages of a decision, different audiences and different tools draw on different sources. Research only tells you something useful if the questions are designed around your audiences, your buying journey and your objective. For PR teams, it also means earned coverage and owned content do different jobs at different moments.
Earned media dominates AI citations. How much depends on who's counting.
84%earned media share of AI citations (Muck Rack, 25m+ links)
25–27%journalism's share, steady across three editions
0.3%paid and advertorial content
Muck Rack's third What Is AI Reading? study analysed more than 25 million links cited by ChatGPT, Claude and Gemini. Earned media accounted for 84%, and has stayed between 82% and 89% since July 2025. The type of question asked changed what got cited.
The catch is the definition. Muck Rack counts journalism, academic research, government sources, encyclopaedias and third-party company content as earned. Meltwater, analysing 5.35 million citations in April 2026, reportedly put earned and news media at 39.5% by sorting domains differently. Both can be right.
My take. Whichever number you prefer, most of what AI says about a brand comes from somewhere other than the brand. That's the strongest argument yet that GEO belongs with PR and comms. Just don't quote 84% as "84% of citations are press coverage". It isn't.
Ask the same question twice, get a different list of brands
SparkToro's Rand Fishkin and Gumshoe.ai's Patrick O'Donnell had 600 volunteers run 12 prompts through ChatGPT, Claude and Google's AI nearly 3,000 times. The chance of getting the same list of brands twice was under 1 in 100. The same list in the same order was closer to 1 in 1,000. Even list length varied, from two names to more than ten.
The useful part most people skipped: how often a brand appeared across many runs was much more stable. For a given intent, the same leading names kept coming back.
My take. This is why I don't give scores or "rankings" in AI. A single answer is a sample, not a verdict. What's worth knowing is which brands and sources keep turning up across lots of answers, and why.
Ranking on Google and being cited by Google's AI are drifting apart
76% → 38%AI Overview citations from top-ten pages, mid-2025 to early 2026
31%from pages outside the top 100
88%AI Mode citations not in the organic top ten
Ahrefs analysed 863,000 keywords and 4 million AI Overview URLs. Moz looked at nearly 40,000 AI Mode queries in the US and UK. Both point the same way: Google's AI runs several related searches behind the scenes and builds one answer from all of them, so it cites far beyond the pages that rank for the original query.
My take. Good SEO still helps you get found. But if your AI strategy is "rank higher", you're aiming at a shrinking slice. Being covered well, in several places, across the related questions people ask, matters more.
Schema and llms.txt aren't the shortcut they're sold as
Ahrefs tracked 1,885 pages that added JSON-LD schema against 4,000 matched pages that didn't. Citations moved by +2.2% in ChatGPT and +2.4% in AI Mode, both statistically indistinguishable from zero, and fell 4.6% in AI Overviews. Separately, Ahrefs found AI crawlers almost never requested llms.txt files on the domains it checked.
My take. Do the technical basics once, properly. Then stop paying for them. The work that moves AI answers is mostly elsewhere.
What dominates the information environment shapes what AI says
A peer-reviewed paper in Nature (Waight et al.) found that large language models show a more pro-government slant in the languages of countries with less media freedom. The mechanism appears to be training data: in one large multilingual training corpus, 3.1 million Chinese-language documents (1.64%) closely matched state-coordinated media. The effect is shown in depth for China and correlationally across many languages.
My take. This is the serious foundation under GEO, and its dark mirror. What circulates shapes what AI says. Improving how accurately AI represents an organisation is legitimate work. Flooding the information environment to bend AI's view is the propaganda playbook this paper documents. Anyone advising on AI visibility should be able to say where that line is. Two practical points for comms teams: check what AI says about you in every language that matters, because the English answer may not be what other markets hear; and don't paste AI summaries of contested or international subjects into client work as if they were neutral.
Source: Waight et al., Nature, 2026. DOI 10.1038/s41586-026-10506-7. Trust: high, within its stated limits
June 2026 · Search behaviour
Two thirds of Google searches now end without a click
68%of US Google searches ended with no click (60% in 2024)
~60%drop in click-through when an AI Overview appears
SparkToro's zero-click analysis covers January to April 2026. Over a decade, zero-click searches have climbed from roughly 45% to 68%. Where AI Overviews appear, they keep people on the results page.
My take. If you only measure website traffic, you only see what's left over. More and more of the first impression now happens in the answer itself.
Source: SparkToro zero-click study, 2026. Trust: solid, US data
May 2026 · Timing
The first proper clock on AI citation
Josh Blyskal at Profound measured how long ChatGPT and Claude take to start citing newly published pages. Half were cited within 7 days, 90% within 37 days, and the typical time to first citation was 6.81 days. Commenters on the original post made two good points: this measures the whole pipeline (indexing, then retrieval), not just content quality; and new topics get cited faster than crowded ones.
My take. That's how quickly the door opens. Whether it stays open is a different question, and given how much answers vary, it's often the more useful one. I've written more in AI cites you in 6.81 days.
Source: Profound (Josh Blyskal), LinkedIn, 11 May 2026. Trust: useful, small sample
March–June 2026 · Volatility, freshness, sources
AirOps: four numbers worth keeping, and a playbook to read carefully
30%of brands stay visible run to run; 20% across five runs
~2xmore likely to lose citations if not updated in a year
6.5xmore likely to be cited via someone else's page than your own
~50%of citations from community sites; Reddit in over a fifth of answers
AirOps, a US content platform, published several studies and a 17-page playbook built on about 15 million data points across 21,000 brands. Other useful figures: 68% of brand mentions appear in only one AI tool; 85% of brand mentions come from third-party sources; around 90% of third-party citations come from listicles, comparisons and review sites; and content under three months old was about three times more likely to be cited.
My take. The data is worth reading. The frameworks are sales material: every featured case study is an AirOps customer, and the prompt sets, model versions and markets aren't published. Quote the numbers as "AirOps found", not as independent benchmarks. And the community figure is the one most likely to be misused: Reddit is huge in aggregate and absent in plenty of categories. Check yours. More in Four AirOps numbers worth keeping.
Sources: AirOps, 2026 State of AI Search; The Complete AI Search Playbook for Marketers, March 2026. Trust: medium, vendor data
June 2026 · Measurement
Google's AI mentions report arrives in Search Console
Google started rolling out a report in Search Console that separates a site's appearances in AI Overviews and AI Mode: impressions, pages, countries, devices and dates. It shows no clicks and no queries, covers Google only, and went to a subset of UK site owners first.
My take. Switch it on and keep a baseline; it's free and first-party. Leave the "block from AI" toggle alone unless you're a publisher with a licensing strategy. It's one tile, not the picture. My full read is in Switch on Google's AI mentions report.
Source: Search Engine Land coverage, 3 June 2026. Trust: first-party, limited
June 2026 · Publishers and regulation
Publishers draw the line: the CMA opt-out and "brazen theft"
The UK's Competition and Markets Authority used new powers to make Google let sites opt out of its AI summaries without being penalised in normal search, something the BBC and other UK publishers had pushed for. In the same month, New York Times publisher AG Sulzberger told the world's news publishers that AI scraping is "brazen theft".
My take. This is a bargaining position for licensing deals more than a retreat from AI; publishers need the reach. For almost every other brand, the lesson runs the other way: blocking yourself from AI answers is self-harm.
Trust: news, not research
May–June 2026 · What to measure
Is being cited even the point? Burson and AMEC weigh in
Burson's report The Credibility Paradox argued that being mentioned by AI is necessary but not enough: a mention is worthless if the reader doesn't believe it. AMEC, the global body for communications measurement, launched its GEO Principles in May 2026 and warned that "AI visibility alone is the next AVE": a precise-looking number that proves nothing on its own.
My take. Both are right, and it's one reason I've stopped producing scores. Being cited is the starting line. Being described accurately and believed is what comes next, and it comes from consistency, good coverage and good content.
Seer queried six AI tools across 231,347 responses about every category connected to the 2026 Winter Olympics. Brands strong on three kinds of signal (a clear identity, third-party validation, and people talking about them) appeared about 7.8 times more often than brands weak on them. Wikipedia was the most-cited source overall, and roughly one in five mentions used stale or out-of-date descriptions.
My take. The frame is useful; the specific sources aren't universal. The Olympics happens to lean on Wikipedia and Reddit. Many sectors live on trade press, industry bodies, analyst reports or LinkedIn instead. Map which sources AI uses in your category before deciding where to invest. And Wikipedia isn't a tactic: an entry follows independent coverage, it can't be written into existence.
Source: Seer Interactive, GEO Olympics study, February 2026. Trust: rigorous, one sector
May 2026 · Market
State of Martech 2026: the category gets a name
Scott Brinker's annual report renamed the SEO category of his marketing technology map to SEO/AEO/GEO. It found 63.1% of marketers were publishing AI-optimised content, but only 13.6% were measuring whether it was included in AI answers or drove anything.
My take. Lots of doing, not much checking. The gap is less about tools than about someone asking the right questions of the results.
Source: Brinker and Riemersma, State of Martech 2026, chiefmartec. Trust: survey
May 2026 · Communities
Reddit's CEO: AI "would not exist as we know it" without Reddit
Steve Huffman called Reddit's human conversation "modern oil", and has licensing deals with Google and OpenAI to match.
My take. Reddit matters a lot in some categories and not at all in others. Check before you invest, and never astroturf.
Trust: a CEO talking his book
What I think we know so far
Putting the research together with what I've seen in practice. Labelled by how confident I am.
HighOther people's pages do most of the work. Coverage, reviews, reference sites and commentary drive far more AI citations than brands' own sites.
HighAnswers vary constantly. One check tells you little. Patterns across many answers tell you a lot.
HighDifferent AI tools cite different sources. Doing well in one says little about the others.
HighRanking and being cited are related but separate. Especially on Google, the overlap has shrunk.
MediumFreshness matters. Recently updated pages are cited more, though cause and effect isn't fully established.
MediumConsistency helps. Brands that tell the same story everywhere seem to be described more clearly and specifically.
MediumSources are sector-specific. Which outlets, communities and reference sites matter varies a lot by market.
MediumLanguage and country matter. The same brand can be visible in one market and absent in another.
LowTechnical tweaks are overrated. Schema and llms.txt show little effect so far, but the evidence is young.
What I'm still watching
How directly PR coverage turns into AI citations, and how quickly.
Whether AI agents that browse and buy on people's behalf change which sources matter.
How AI answers differ by country and language for the same brand.
Whether the PR measurement platforms publish their methods, so their numbers can be compared.
What advertising inside AI answers does to trust in them.
How I put this page together
I'm one consultant, not a research institute. I read research from universities, industry bodies, platforms and practitioners, and include what's useful to PR and comms people. I link to the original wherever I can, say who ran each study and whether they sell something related, and mark how far I'd trust it. When newer evidence replaces something here, I update the page rather than leave old numbers standing. If you spot an error, tell me.
Get in touch
Tell me the problem.
A short email is fine: what's going on, what you've been asked, and when you need it. I'll give you an honest answer, including if I'm not the right person.