Three names, one question: how do you get an AI answer to use and cite your page? The names are generative engine optimization (GEO), answer engine optimization (AEO) and LLM SEO.
This page answers with measurements, not advice. Every study below is a dated entry on one of our trackers, with its method and its limits stated.
Our rule for this topic. We publish how AI answers choose sources.
We do not publish vendor "share of voice" scoreboards, and we name who paid for each study.
Every study we have covered, newest first
Each row is a dated entry on one of our trackers. The entry states who ran the study, what it measured and what it does not show.
| Date | Event | Tracker | What happened |
|---|---|---|---|
| 2026 | |||
| Audit of 15,942 citations: AI Overviews lean on commercial health sites and YouTube where ChatGPT leans on government pages and Wikipedia | AI Overviews | Harvard researchers audited ChatGPT, Perplexity and AI Overviews on twenty mental health questions. All three lean on ten domains, and AI Overviews gave no citation for 30% of Hindi queries. | |
| Study: the language of the question, not the user's location, decides whether ChatGPT names local suppliers | ChatGPT | In 234 runs, asking in the local language surfaced local suppliers nearly always, and English from Estonia or Turkey never did. The top pick also changed across identical runs on four of six prompts. | |
| ChatGPT grounds most answers on its own index, a 200-character snippet and a cache that ignores your headers | ChatGPT | A study of 1,200 ChatGPT answers, 88,000 search results and 26,900 pages finds three retrieval layers: an index, a shared page cache and live page opens. Free instant answers open no page at all 93% of the time. | |
| Testing a prompt on its own changes the answer in 44.7% of cases | Claude | A paired test on 180 conversations kept the final message and removed everything before it. Answers changed materially in 44.7% of cases, an error under every tool replaying lone prompts. | |
| Pinterest moves search relevance judging to vision-language models | Social Platforms | Pinterest researchers published how VLM judges now evaluate search experiments, extending the LLM relevance judge deployed in December to images. No platform documents its ranking machinery in public the way Pinterest's engineering team does. | |
| Top Stories sits inside the AI Overview on one in six news searches | AI Overviews | NewzDash measured the Top Stories carousel Google moved inside AI Overviews, which replaces the standalone box rather than adding to it. In the US it sat inside the AI Overview on 15.5% of searches with a carousel, in the UK 17.5%. | |
| ChatGPT is Google's sixth most-clicked destination, and buys more of those clicks than any other | ChatGPT | iPullRank measured 13.1 billion Google search events and found ChatGPT the sixth most-clicked destination from Google Search. About 4.75% of its Google traffic arrives through ads, the highest paid share of any top destination ranked. | |
| Productrise: AI Mode shows 95% fewer product listings than standard search | AI Mode | Productrise ran identical shopping queries on both surfaces for 21 days in July. AI Mode showed about 95% fewer product listings than standard search, sharing just 0.8% of products. | |
| Similarweb puts AI Overviews on 43% of US searches, nearly triple last year | AI Overviews | Similarweb's 2026 Generative AI Landscape report puts AI Overviews on 43% of US searches, up from 15% a year earlier. AI Mode visits reached 279 million a month against 126 million in June 2025, and Similarweb has published no methodology. | |
| Similarweb report: citations reach 6.8% of ChatGPT answers, ads reach 26% of chats | ChatGPT | Similarweb's 2026 Generative AI Landscape measured ChatGPT citing the web in 6.8% of answers in May 2026, up from about 1.3% a year earlier. It found 58.8% of referral traffic landing on homepages and ads in 26% of US desktop chats. | |
| AOP: Google referrals to eight UK publishing groups fell 7.1% in a quarter | AI Overviews | An AOP study of 10.8 billion pageviews across eight UK publishing groups found organic Google referrals down 7.1% quarter on quarter. Sustained, that rate halves Google's pageviews to those publishers by the third quarter of 2027. | |
| DataDome: Claude referrals more than double, while Meta's crawlers take over the AI web | Claude | Claude referral visits rose 111% to 876,000, the fastest-growing referrer DataDome tracks. Meta's crawlers made 9.1 billion of 17.7 billion second-quarter AI agent requests and sent almost nothing back. | |
| Ozone: publisher ad supply fell up to 40% in Q2 as AI search shrank the open web | AI Overviews | Ozone's benchmarking put second-quarter UK publisher ad requests down 39% to 41% and US requests down 32% to 37%. Its cohort includes the Guardian, News UK and the Wall Street Journal, and it blames zero-click AI search broadly, not AI Overviews. | |
| Common Sense Media rates Google AI search "Unacceptable" for children | AI Overviews | Common Sense Media's Youth AI Safety Institute gave AI Overviews and AI Mode its lowest rating, "Unacceptable Risk", after 2,624 test interactions. AI Overviews routed only 58% of crisis prompts to appropriate resources, and Google called the methodology flawed. | |
| Common Sense Media rates AI Mode "Unacceptable" for children | AI Mode | Common Sense Media's Youth AI Safety Institute rated AI Mode and AI Overviews "Unacceptable Risk" over 2,624 tests. AI Mode did better but failed all five severe-harm "Red Lines" and did 100% of homework assignments. | |
| First randomized experiment puts the click loss at 39.8%, and finds the lost clicks were not lower quality | AI Overviews | Two academics hid AI Overviews from randomly chosen users and measured what happened. Outbound clicks rose from 0.37 to 0.62 per search without the summary, a 39.8% loss attributed to AI Overviews. | |
| When an assistant recommends brands, it mostly cites other companies' pages | Claude | Across 102,025 responses, 102 brands and five assistants, 75.2% of citations point at other companies' pages. Only 2.9% hit the brand's own domain; "best-X" listicles take 21.0% of all citations. | |
| LLMs pick the brand they recognise in every tie, and a fabricated authority claim breaks it | ChatGPT | Across GPT-4o-mini, Claude Sonnet and Gemini 3 Flash, a known brand won 100% of trials when all ten products had identical specs. Any real quality difference cut that to about 4%, invented authority claims did almost as well, and universal use erased the advantage. | |
| SparkToro: 68% of US Google searches now end without a click | AI Overviews | SparkToro measured 68.01% of US Google searches ending without a click in the first four months of 2026, up 7.56 points on 2024. The same panel found only 0.34% of searches moved into AI Mode, the first independent figure for that tab's share. | |
| A natural experiment finds most AEO growth was the platform, not the optimization | ChatGPT | Glasp optimized part of its site for answer engines in January 2026 and used the rest as a control. Referrals grew 5.7 times but untreated pages grew 3.5 times, leaving a level effect of about 1.8 to 2.3 times, which its strictest test calls suggestive. | |
| SE Ranking: Claude referrals grew 386% in four months, from a tiny base | Claude | Across 101,574 Google Analytics sites, Claude's referral share rose from 0.0029% in January 2026 to 0.0141% in April. That is still just 1.40% of AI-referred traffic, against ChatGPT's 78%. | |
| Cursor data from 846,000 sessions: an AI Overview slows searchers down and flattens intent | AI Overviews | ClickStream Solutions sampled cursor and scroll behaviour in about 846,000 US Google sessions. With an AI Overview present, searchers pause more and scroll back up more, and search intent stops predicting dwell time. | |
| A 252,000-trial experiment on what wins the first citation | Claude | A peer-reviewed SIGIR 2026 paper found four gatekeepers for the first citation. They are topic fit, a stated price, a recent date and list position; formatting moved nothing. | |
| SISTRIX: an AI Overview sits on 72% of health searches for UK newsbrands | AI Overviews | SISTRIX data given to Press Gazette puts AI Overview exposure at 72% for UK newsbrands' health sections and 5% for football. Tech, travel, lifestyle and money all run above a third, making the vertical a bigger variable than most publishers assume. | |
| A causal study finds AI Overviews raised Reddit engagement, and AI Mode erased the gain | AI Overviews | Researchers used Google's ban on citing adult subreddits as a control group. AI Overviews raised daily Reddit comments 12.0% and commenting users 12.4%, and AI Mode's arrival largely wiped the gain out. | |
| A causal study finds AI Mode erased the Reddit engagement gain AI Overviews created | AI Mode | AI Overviews raised Reddit's comments 12.0%; AI Mode then cut that gain 59% for commenting users. A conversational interface substitutes for the discussion a summary only pointed at, researchers say. | |
| A 55,000-query audit finds 30% of AI Overview sources appear nowhere on the first page | AI Overviews | Washington University researchers ran 55,393 trending queries over 40 days and captured 7,583 AI Overviews. Nearly 30% of cited domains were absent from the first page, and 11% of claims were unsupported by the cited sources. | |
| SISTRIX: AI Overviews reuse the same sources for months | AI Overviews | SISTRIX tracked 82,619 prompts weekly for 17 weeks and found no cited source changed in 53% of AI Overviews prompts. AI Mode replaced 56% of its domains weekly, and the two surfaces cited different domains 83% of the time. | |
| SISTRIX: AI Mode replaces 56% of its cited sources every week | AI Mode | SISTRIX tracked 82,619 prompts weekly for 17 weeks. AI Mode cites 14 to 16 domains per answer and swaps 56% weekly; AI Overviews barely move. | |
| A peer-reviewed study finds sites that block Google-Extended cited less often in AI Overviews | AI Overviews | A SIGIR 2026 paper compared Google Search, AI Overviews and Gemini across 14,212 queries in December 2025. Sites blocking Google-Extended were cited less often by AI Overviews, despite Google saying the token has no effect there. | |
| Seer: first partial CTR recovery on AI Overview queries | AI Overviews | Seer Interactive's 5.47 million query dataset shows AI Overview CTR recovering, and cited brands earning about 120% more clicks per impression. It rose from 1.3% in December 2025 to 2.4% in February 2026. | |
| Chartbeat: search referrals to small publishers are down 60% in two years | AI Overviews | Chartbeat data shows search referrals down 60% for small publishers over two years, 47% for mid-sized and 22% for large. Chatbot referrals grew more than 200% but remain under 1% of publisher page views. | |
| BrightEdge: AI Overviews now appear on about 48% of tracked queries | AI Overviews | Across nine industries, BrightEdge measured AI Overviews on roughly 48% of tracked queries, up 58% year over year. Healthcare sits at 88% and education at 83%. | |
| SISTRIX: AI Overviews cost the German market 265 million clicks a month | AI Overviews | Across more than 100 million German keywords, SISTRIX found top-result click-through falls from 27% to 11% when an AI Overview is present. That is a drop of almost 60%, and it estimates the German market loses 265 million organic clicks a month. | |
| The first measurement of the February core update: more topics, fewer publishers | Google Discover | NewzDash measured the February 2026 core update: more topics through fewer publishers, with regional titles and X.com posts gaining. On Google's anti-clickbait goal its verdict was "directional" rather than proven. | |
| Ahrefs: AI Overviews now cut top-result CTR by 58% | AI Overviews | Ahrefs' December 2025 data puts the top result's click loss under an AI Overview at 58%, from 34.5% a year earlier. The same day Alphabet reaffirmed 2 billion plus monthly AIO users. | |
| Chartbeat data: Discover referrals down 21% globally, 29% in the US | Google Discover | The Reuters Institute's 2026 trends report used Chartbeat data from more than 2,500 publisher sites. Discover referrals fell 21% in the year to November 2025, the US fell 29%, and search fell further. | |
| 2025 | |||
| Marfeel: AI Summaries are 51% of the Discover feed, and their clicks default to YouTube | Google Discover | Marfeel measured AI Summaries at 51% of Discover feed positions in the US, Brazil and Mexico, and 82.7% beyond position twenty. In the US, 77% of AI Summary cards default to an inline YouTube play, not a publisher link. | |
| Semrush: AIO prevalence peaked at 24.61% in July, settled near 16% | AI Overviews | Semrush's refreshed 10 million keyword study charts AI Overview prevalence from 6.49% in January 2025 to a 24.61% July peak. It then eased to 15.69% by November 2025. | |
| Ahrefs: AI Overviews and AI Mode agree on the answer and disagree on the sources | AI Overviews | Across 540,000 query pairs, Ahrefs found AI Overviews and AI Mode cited the same URL only 13.7% of the time. Answers matched in meaning 86% of the time, yet AI Overviews cited nothing in 11% of responses, against 3% for AI Mode. | |
| Ahrefs: AI Mode cites more, and it barely cites what AI Overviews cites | AI Mode | Ahrefs compared 540,000 AI Mode and AI Overview pairs: 13.7% shared a URL, 86% matched in meaning. AI Mode cited sources in 97% of answers, naming 2.5 times as many entities. | |
| Cloudflare's crawl-to-refer series: Anthropic crawls tens of thousands of pages per referral | Claude | Cloudflare put Anthropic's crawl-to-refer ratio at roughly 71,000:1 in June 2025 and 38,066:1 in July. Early August read roughly 50,000:1, the highest of any AI operator; Claude app referrals carry no referrer header. | |
| Authoritas: UK news publishers lose 47.5% of desktop clicks to AI Overviews | AI Overviews | Authoritas found UK news publisher clickthrough falling 47.5% on desktop and 37.7% on mobile when an AI Overview appears. It covered 3,500 search terms, went to the CMA as evidence, and drew a Google rebuttal calling it inaccurate. | |
| Pew: users click a result on 8% of visits with an AI summary vs 15% without | AI Overviews | Pew found Google users clicked a traditional result on 8% of visits with an AI summary, versus 15% without one. They clicked a source inside the summary on just 1% of those visits. | |
| Amsive: AI Overviews cut average CTR by 15.49%, branded queries gain 18.68% | AI Overviews | Amsive's study of 700,000 keywords measured an average 15.49% CTR drop on queries with an AI Overview. It worsened to 37.04% when a featured snippet also appeared, while branded queries gained 18.68%. | |
| Ahrefs: an AI Overview means a 34.5% lower CTR for the top result | AI Overviews | Ahrefs compared 300,000 keywords and found the top-ranking page's average CTR was 34.5% lower when an AI Overview was present. It was the first large-scale quantification of AIO click loss. | |
Where the terms come from
GEO has an academic origin. AEO and LLM SEO are trade coinages for the same problem.
"Generative engine optimization" was coined in a paper by Pranjal Aggarwal and colleagues at Princeton, Georgia Tech, the Allen Institute and IIT Delhi. It was first posted on 16th Nov 2023 and published at KDD 2024.
The paper introduced GEO as "the first novel paradigm to aid content creators in improving their content visibility in generative engine responses". Its headline: "GEO can boost visibility by up to 40% in generative engine responses."
That 40% is a lab result on the authors' own benchmark and visibility metric. It is not a traffic measurement.
Microsoft adopted the term in 2026. Bing Webmaster Tools called its AI Performance report "an early step toward Generative Engine Optimization (GEO) tooling", as our Copilot tracker records.
"Answer engine optimization" and "LLM SEO" have no single origin. They describe the same goal for assistants that answer rather than list.
How AI answers find pages
Before a page can be cited it has to be retrieved. Two studies looked inside that step.
A study of 1,200 ChatGPT answers found three retrieval layers: an index, a shared page cache and live page opens. Free instant answers opened no page at all 93% of the time, per our 17th Aug 2026 entry.
The language of the question decides the market. In 234 runs, asking in the local language surfaced local suppliers nearly always.
English from Estonia or Turkey never did, per our 30th Aug 2026 entry.
Conversation history matters too. Removing everything before the final message changed the answer materially in 44.7% of 180 conversations.
That undercuts tools that replay lone prompts, per our 3rd Aug 2026 entry.
On Google, a peer-reviewed SIGIR 2026 study found sites blocking Google-Extended were cited less often in AI Overviews. Google's documentation does not predict that, and our 30th Apr 2026 entry has the details.
What wins a citation
The controlled experiments point at topic fit, recency, a stated price and being on a list. Formatting moved nothing.
A SIGIR 2026 paper ran 252,000 trials on what wins the first citation. Four gatekeepers emerged: topic fit, a stated price, a recent date and list position.
Formatting moved nothing, per our 25th May 2026 entry.
Most citations are not the brand's own pages. Across 102,025 responses, 75.2% of citations pointed at other companies' pages and only 2.9% at the brand's own.
"Best-X" listicles took 21.0%, per our 18th Jun 2026 entry. Ranqo is a visibility vendor, so read the direction of its interest.
Recognition beats specs. With ten identical products, a known brand won 100% of trials across three models.
Any real quality difference cut that to about 4%. Invented authority claims did almost as well, per our 16th Jun 2026 entry.
AI Overviews draw from outside the first page. A 55,393-query audit found nearly 30% of cited domains absent from page one.
The same audit found 11% of claims unsupported by their own sources, per our 13th May 2026 entry.
Sources churn at different rates by surface. AI Mode swapped 56% of its cited domains each week over 17 weeks.
AI Overviews barely moved, per our 1st May 2026 entry.
Does optimizing actually move traffic?
One natural experiment separated the platform's growth from the optimization. Most of the lift was the platform.
Glasp optimized part of its site for answer engines in January 2026 and used the rest as a control. Referrals to treated pages grew 5.7 times, and untreated pages grew 3.5 times.
The difference, about 1.8 to 2.3 times, is the effect the optimization can claim. The study's strictest test calls it suggestive, per our 3rd Jun 2026 entry.
The supply side is changing too. A detector study flagged 8.90% of pages behind generative answers as written for generative engines.
Among pages modified in 2026 the share was 16.36%, per our 17th Aug 2026 entry.
Scale the expectations. ChatGPT cited the web in 6.8% of answers in May 2026, per Similarweb.
It sent 78% of AI referrals in March 2026, per Statcounter. Both figures are on our AI search statistics page.
What the evidence does not show
Vendor rankings, hidden weights and one-prompt snapshots are the three things to distrust.
No study here shows a ranking formula. The controlled experiments identify factors that move citation odds in a test setting, not weights inside a model.
No vendor scoreboard appears here. "Share of voice" numbers depend on the prompt set, the day and the account.
The conversation-history study shows why a lone prompt misleads.
No study here measures your site. The Glasp experiment is one site, and the citation experiments used synthetic products.
Measure your own referrals before and after any change.
Our methodology explains why we cover the mechanism and skip the leaderboard.
Sources
Frequently asked questions
What is generative engine optimization?
The practice of making content more likely to be used and cited in AI-generated answers. The term comes from a 2023 Princeton-led paper published at KDD 2024.
That paper claimed visibility gains of "up to 40%" on its own benchmark.
Is GEO different from SEO?
The retrieval step overlaps: AI answers pull from search indexes and live pages. The citation step differs.
Controlled tests found topic fit, recency, a stated price and list position moved citations, while formatting did not.
Does answer engine optimization work?
The one natural experiment we have covered found most referral growth came from the platform, not the optimization. The residual lift was about 1.8 to 2.3 times, which the study itself calls suggestive.
Why does this page have no AI visibility rankings?
Because they measure a prompt set on a day, not a market. Our policy is to publish how AI answers choose sources and to name who ran each study.
Compiled by Claude from primary sources, checked by Akash Mansukhani before publication. See how we source and correct this page.Keywords Everywhere, "AI search optimization: what the studies say", last updated 6th Sep 2026, https://keywordseverywhere.com/news/ai-search-optimization/