AEO and AI search statistics, with sources
Public studies on AEO, GEO and AI search, one finding per row, each with its sample, method, sponsor and a footnote to the primary source.
Footnotes LabExperimentsAI searchUpdated 37 studies
How rows are chosen
Most statistics pages on this subject collect vendor numbers without the method behind them. This one keeps a study only when it passes four checks.
- Verified primary source. We opened the study itself, and every number in a row appears there. Press coverage does not count, and studies we could not verify are left out.
- Method shown. Each row says how the data was collected and how large the sample was. A vendor number without a stated method stays out.
- Sponsor disclosed. Each row names the publisher and any sponsor, and marks whether the publisher sells an AEO, GEO, AI visibility or AI crawler product or service: 19 of 37 rows do.
- Stated as the source states it. Findings are not combined or recomputed, and each row says what the study does not show.
Studies 37
Showing 37 of 37 studies
| Finding | Sample | Method | Published | Publisher |
|---|---|---|---|---|
Of the 118 domains of the tools and agencies in the AEO Footnotes directory, 116 served a robots.txt and 64 served an llms.txt.37 Crawlers and llms.txt. Does not show: Sites in one directory, not a sample of the web; publishing llms.txt says nothing about whether AI systems read it. | 118 domains | Fetch of /robots.txt and /llms.txt on each domain; blocking judged by the RFC 9309 group that appliesEngines: 14 AI crawler user agents | Data: October 29, 2026 | AEO Footnotes (Footnotes Lab)Sponsor: None; AEO Footnotes runs the directory the sites come from and sells no AEO productSells an AEO product: no |
10.13% of nearly 300,000 domains had an llms.txt file, and removing llms.txt from a model of AI citation frequency made its predictions more accurate.33 Crawlers and llms.txt. Does not show: The citation sample and model are not published in detail; a weak predictor in one model is not proof of no effect. | 300,000 domains (nearly) | Check for llms.txt on each domain, then Spearman correlation, XGBoost regression and SHAP analysisEngines: LLM responses (models not named) | Data: Not stated on the page | SE RankingSponsor: SE Ranking (sells SEO tools and an AI Search Toolkit)Sells an AEO product: yes |
9.3% of the top 1,000 domains and 8.3% of the top 10,000 served a confirmed llms.txt or llms-full.txt file.34 Crawlers and llms.txt. Does not show: Unknown results count as not confirmed (474 of the top 1,000), so the shares are floors for the ranked list. | 10,000 domains from the Tranco list | Root file check of each domain, with blocked or inconclusive checks kept in the denominatorNo engine measured | Data: September 18, 2026 | RankabilitySponsor: Rankability (sells AI search visibility software for agencies)Sells an AEO product: yes |
11.72% of 5,167,831 sampled /llms.txt URLs returned a text body, and 68.27% of the 584,107 files collected were templated.35 Crawlers and llms.txt. Does not show: Hosts Common Crawl can already fetch, not the whole web; the post says the rate is not comparable to the Web Almanac figure. | 5,167,831 /llms.txt URLs from a random sample of hosts | Crawl experiment: /llms.txt added to the seed list for a random sample of hosts in the July 2026 crawlNo engine measured | Data: July 2026 crawl (CC-MAIN-2026-30) | Common Crawl FoundationSponsor: None stated (nonprofit that operates CCBot)Sells an AEO product: no |
28% of 137,210 domains published an llms.txt file, and 97% of those files received zero requests in May 2026.36 Crawlers and llms.txt. Does not show: Ahrefs Web Analytics customers skew technical, so the study treats the 28% adoption figure as an upper bound. | 137,210 domains in Ahrefs Web Analytics with traffic | Root file check plus bot and user-agent requests to the llms.txt pathEngines: AI bots and assistants that request llms.txt | Data: May 2026 | AhrefsSponsor: Ahrefs (sells SEO tools and Brand Radar for AI visibility)Sells an AEO product: yes |
About 84% of links cited by ChatGPT, Claude and Gemini came from earned media, and about 27% were journalistic.24 Citations and accuracy. Does not show: Earned media here includes Wikipedia, social platforms and third-party corporate content; the prompt set is not published. | 25,000,000 links from AI responses (25,000,000+) | Prompt runs across a large consumer prompt set; every cited URL categorizedEngines: ChatGPT, Claude, Gemini | Data: 2026 | Muck RackSponsor: Muck Rack (sells PR software and Generative Pulse for AI visibility)Sells an AEO product: yes |
79% of top U.S. and UK news sites blocked at least one AI training bot in robots.txt, and 71% blocked at least one AI retrieval or live search bot.32 Crawlers and llms.txt. Does not show: Sites chosen by SimilarWeb traffic share; robots.txt is a request, and the post notes bots can ignore it. | 100 top news sites (50 U.S., 50 UK, deduplicated) | robots.txt disallow directives checked for 11 AI crawlersEngines: 11 AI crawlers, including GPTBot, ClaudeBot and PerplexityBot | Data: 2025 to 2026 | BuzzStreamSponsor: BuzzStream (sells outreach and digital PR software)Sells an AEO product: no |
Rerun with December data, the presence of an AI Overview correlated with a 58% lower average clickthrough rate for the top-ranking page.4 Clicks and CTR. Does not show: CTR for the baseline keywords without AI Overviews also fell (0.076 to 0.039), so the forecast absorbs other changes over two years. | 300,000 keywords (150,000 with an AI Overview, 150,000 informational without) | Same method as the April 2025 study: aggregated Google Search Console CTR, before and afterEngines: Google AI Overviews | Data: December 2023 and December 2025 | AhrefsSponsor: Ahrefs (sells SEO tools and Brand Radar for AI visibility)Sells an AEO product: yes |
For the same keywords measured before and after they began showing an AI Overview, the zero-click rate fell from 33.75% to 31.53%.7 Clicks and CTR. Does not show: A zero-click rate is not a clickthrough rate to any one page; the page does not give the size of the before and after keyword set. | 200,000 keywords with search volume over 100 (200K+) | Clickstream panel from Datos (a Semrush company), keywords that gained an AI Overview between May and OctoberEngines: Google AI Overviews | Data: January to October 2025 | SemrushSponsor: Semrush (sells SEO and AI visibility tools); data from Datos, a Semrush companySells an AEO product: yes |
AI Overviews appeared for 6.49% of keywords in January 2025, 24.61% in July and 15.69% in November.8 Adoption. Does not show: A share of keywords in one vendor's database, not a share of the searches people actually run. | 10,000,000 keywords (10M+) | SERP tracking across the Semrush keyword database, month by monthEngines: Google AI Overviews | Data: January to November 2025 | SemrushSponsor: Semrush (sells SEO and AI visibility tools)Sells an AEO product: yes |
YouTube mentions showed the strongest correlation with brand mentions in ChatGPT, AI Mode and AI Overviews (about 0.737), ahead of branded web mentions (0.66 to 0.71).19 Citations and accuracy. Does not show: Correlation, not causation, as the study says; both the signals and the AI mentions come from Ahrefs' own data. | 75,000 brands | Spearman correlation between brand signals in Ahrefs' indexes and mention counts in AI answersEngines: ChatGPT, Google AI Mode, Google AI Overviews | Data: Not stated on the page | AhrefsSponsor: Ahrefs (sells SEO tools and Brand Radar for AI visibility)Sells an AEO product: yes |
Organic CTR on informational queries that showed an AI Overview fell from 1.76% in June 2024 to 0.61% in September 2025, a 61% decline.5 Clicks and CTR. Does not show: Informational queries from Seer's clients only; organic CTR on queries without an AI Overview also fell 41% over the same period. | 3,119 search terms across 42 client organizations | Client organic and paid search data (25.1M organic and 1.1M paid impressions), split by AI Overview presence and citationEngines: Google AI Overviews | Data: June 2024 to September 2025 | Seer InteractiveSponsor: Seer Interactive (search agency that sells GEO services)Sells an AEO product: yes |
When a brand was cited in an AI Overview it earned 35% more organic clicks and 91% more paid clicks than when it was not cited.6 Citations and accuracy. Does not show: Seer states it cannot show that citation causes the higher CTR; brands with stronger baseline CTR may simply be cited more. | 3,119 search terms across 42 client organizations | Client search data split by whether the brand was cited or mentioned in the AI OverviewEngines: Google AI Overviews | Data: June 2024 to September 2025 | Seer InteractiveSponsor: Seer Interactive (search agency that sells GEO services)Sells an AEO product: yes |
45% of AI assistant answers to news questions had at least one significant issue, and 31% showed serious sourcing problems.17 Citations and accuracy. Does not show: News questions only, graded by journalists at the broadcasters whose content was being tested. | 3,000 responses (more than 3,000) | Journalists from 22 public service media organizations in 18 countries and 14 languages graded answersEngines: ChatGPT, Copilot, Gemini, Perplexity | Data: 2025 | European Broadcasting Union and BBCSponsor: None statedSells an AEO product: no |
82% of articles cited by ChatGPT and Perplexity were human-written and 18% AI-generated, as classified by an AI detector.23 Citations and accuracy. Does not show: Detector-based, with a 4.2% false positive rate Graphite reports; AI drafts edited by people were not evaluated. | 31,493 keywords across 10 categories (100 per category asked of answer engines) | AI detector (Surfer) run on 500-word chunks of ranking and cited articlesEngines: ChatGPT, Perplexity, Google Search | Data: June 2025 | GraphiteSponsor: Graphite (growth agency that sells SEO and AEO work)Sells an AEO product: yes |
For ranking questions about well-known brands, 93.5% of the sources ChatGPT cited were earned media, against 67.4% for Perplexity and 63.4% for Gemini.21 Citations and accuracy. Does not show: Domain labels rely partly on GPT-assisted classification, and API answers can differ from the consumer apps. | ranking-style brand queries in two sets (size not stated for this experiment) | Prompt runs through each engine's API; cited domains classified as brand, earned or socialEngines: ChatGPT, Claude, Gemini, Perplexity | Data: 2025 | Chen, Wang, Chen and Koudas, University of Toronto (arXiv 2509.08919)Sponsor: None statedSells an AEO product: no |
Over the 12 months to August 2025, 80% of AI crawling was for training, against 18% for search and 2% for user actions.29 Crawlers and llms.txt. Does not show: Purpose comes from what operators declare about each crawler, not from how the data is used. | requests across Cloudflare's network | Crawler traffic classified by purpose from operator disclosures and industry sourcesEngines: AI crawlers seen by Cloudflare | Data: 12 months before August 29, 2025 | Cloudflare (Radar)Sponsor: Cloudflare (sells AI Crawl Control and pay per crawl)Sells an AEO product: yes |
Google users who saw an AI summary clicked a traditional search result on 8% of visits, against 15% of visits without one, and clicked a link inside the summary on 1% of visits.1 Clicks and CTR. Does not show: Google and U.S. adults only, over one month; longer and question-style searches trigger summaries more often, so query type and summary presence are mixed. | 900 U.S. adults (68,879 Google searches) | Metered browsing panel (KnowledgePanel Digital): every Google search on tracked devices and what the user did nextEngines: Google AI Overviews | Data: March 1 to 31, 2025 | Pew Research CenterSponsor: None statedSells an AEO product: no |
18% of the Google searches in Pew's panel produced an AI summary, and 88% of those summaries cited three or more sources.2 Adoption. Does not show: A one-month snapshot from one panel; Google has changed how often it shows AI summaries since. | 68,879 Google searches by 900 U.S. adults | Metered browsing panel (KnowledgePanel Digital), all Google searches in the monthEngines: Google AI Overviews | Data: March 1 to 31, 2025 | Pew Research CenterSponsor: None statedSells an AEO product: no |
76.10% of pages cited in AI Overviews ranked in Google's top 10 for the query, and 14.40% did not rank in the top 100.18 Citations and accuracy. Does not show: Only the top three citations per answer, ranked for the tracked keyword rather than the fan-out queries behind the answer. | 1,900,000 citations from 1M AI Overviews | Rank lookup for the three most visible citations in each AI OverviewEngines: Google AI Overviews | Data: Not stated on the page | AhrefsSponsor: Ahrefs (sells SEO tools and Brand Radar for AI visibility)Sells an AEO product: yes |
9% of the more than 366,000 citations in AI Search Arena responses referenced news sources, concentrated among a small number of outlets.22 Citations and accuracy. Does not show: Arena users and their prompts are not a representative sample of people searching. | 366,000 citations (over 366,000) in 65,000 responses | Analysis of conversations from the AI Search Arena, a head-to-head evaluation platformEngines: OpenAI models, Perplexity models, Google models | Data: 2025 | Kai-Cheng Yang, Northeastern University (arXiv 2507.05301)Sponsor: None statedSells an AEO product: no |
For June 19 to 26, 2025, crawl-to-refer ratios ranged from 70,900 HTML page crawls per referral for Anthropic down to 0.1 for Mistral.28 Crawlers and llms.txt. Does not show: Native apps such as Claude's send no Referer header, so their referrals are undercounted, as Cloudflare notes. | requests across Cloudflare's network | HTML requests from each platform's crawlers divided by HTML requests whose Referer is that platformEngines: Anthropic, OpenAI, Perplexity, Google, Mistral, and others on Radar | Data: June 19 to 26, 2025 | Cloudflare (Radar)Sponsor: Cloudflare (sells AI Crawl Control and pay per crawl)Sells an AEO product: yes |
Of 3,816 top 10,000 domains with a robots.txt, 546 (about 14%) had allow or disallow directives aimed at AI bots.30 Crawlers and llms.txt. Does not show: Following robots.txt is voluntary; a directive shows what a site asks, not what crawlers do. | 3,816 domains with a robots.txt, from the top 10,000 | robots.txt files fetched by Cloudflare Radar, matched against a list of AI crawler namesEngines: AI crawlers (GPTBot most named) | Data: June 6, 2025 | Cloudflare (Radar)Sponsor: Cloudflare (sells AI Crawl Control and pay per crawl)Sells an AEO product: yes |
34% of U.S. adults say they have ever used ChatGPT, roughly double the share in 2023, including 58% of adults under 30.12 Adoption. Does not show: Self-reported ever-use; it does not measure how often people use ChatGPT or whether they use it to search. | 5,123 U.S. adults | Survey of Pew's American Trends Panel, recruited by random address sampling and weightedEngines: ChatGPT | Data: February 24 to March 2, 2025 | Pew Research CenterSponsor: None statedSells an AEO product: no |
Of 54 cases tested, only 3 showed statistically significant ranking gains from existing conversational SEO methods for a single adopter.26 GEO experiments. Does not show: A simulated conversational search engine over fixed documents, not live AI search products. | 54 method, task and domain cases | Benchmark: C-SEO methods applied to documents, rank change measured in LLM answers for two tasks and six domainsEngines: gpt-4o-mini, Claude 3.5 Haiku, o3, o4-mini | Data: 2025 | Puerto et al., Parameter Lab and TU Darmstadt (arXiv 2506.11097, NeurIPS 2025 Datasets and Benchmarks)Sponsor: None statedSells an AEO product: no |
In Profound's tracked data, Wikipedia was 7.8% of ChatGPT citations, and Reddit was 6.6% of Perplexity citations and 2.2% of Google AI Overviews citations.20 Citations and accuracy. Does not show: Shares of a vendor's tracked prompt set, which is not published; not a sample of what users ask. | 680,000,000 citations | Citation counts from the prompts Profound tracks across three platformsEngines: ChatGPT, Google AI Overviews, Perplexity | Data: August 2024 to June 2025 | ProfoundSponsor: Profound (sells an AI visibility platform)Sells an AEO product: yes |
The presence of an AI Overview correlated with a 34.5% lower average clickthrough rate for the top-ranking page, comparing March 2024 with March 2025.3 Clicks and CTR. Does not show: Desktop CTR only, from Search Console data Ahrefs can see; the effect is a forecast-based correlation, not a controlled test. | 300,000 keywords (150,000 with an AI Overview, 150,000 informational without) | Before and after CTR from aggregated Google Search Console data, with informational keywords as the baselineEngines: Google AI Overviews | Data: March 2024 and March 2025 | AhrefsSponsor: Ahrefs (sells SEO tools and Brand Radar for AI visibility)Sells an AEO product: yes |
Traffic to U.S. retail sites from generative AI sources rose 1,200% in February 2025 compared with July 2024.11 Adoption. Does not show: Growth from a small base; Adobe says generative AI traffic remains modest next to channels such as paid search or email. | 1,000,000,000,000 visits to U.S. retail sites (more than 1 trillion) | Adobe Analytics visit and transaction data, plus a survey of 5,000 U.S. consumersEngines: Generative AI sources (not broken out) | Data: July 2024 to February 2025 | AdobeSponsor: Adobe (sells Adobe Analytics and Adobe Brand Visibility, formerly LLM Optimizer)Sells an AEO product: yes |
Eight AI search tools gave incorrect answers to more than 60% of 1,600 queries asking them to identify a news article's headline, publisher, date and URL.13 Citations and accuracy. Does not show: One narrow task (finding the source of an excerpt); it does not measure answer quality for ordinary questions. | 1,600 queries (20 publishers, 10 articles each, 8 tools) | Prompt runs with direct article excerpts, each answer graded by hand for article, publisher and URLEngines: ChatGPT Search, Perplexity, Perplexity Pro, DeepSeek Search, Copilot, Grok 2, Grok 3, Gemini | Data: February 2025 | Tow Center for Digital Journalism, Columbia Journalism ReviewSponsor: None statedSells an AEO product: no |
51% of AI assistant answers to questions about the news were judged to have significant issues, and 19% of answers citing BBC content introduced factual errors.16 Citations and accuracy. Does not show: One publisher's content, rated by its own journalists; news questions only. | 362 AI responses reviewed by 45 BBC journalists | Four assistants given access to the BBC website answered news questions; expert journalists rated each answerEngines: ChatGPT, Copilot, Gemini, Perplexity | Data: December 2024 | BBCSponsor: None statedSells an AEO product: no |
AI chatbots sent 0.17% of the average site's visitors, and 63% of sites received at least one visit from an AI chatbot.9 Clicks and CTR. Does not show: Sites that use Ahrefs Web Analytics, not a random sample; visits that arrive without an AI referrer are not counted as AI traffic. | 3,000 websites | Referral data from an anonymized sample of sites running Ahrefs Web Analytics, seven chatbotsEngines: ChatGPT, Perplexity, Gemini, Claude, Copilot, Mistral, Jasper | Data: Not stated on the page | AhrefsSponsor: Ahrefs (sells SEO tools and Brand Radar for AI visibility)Sells an AEO product: yes |
ChatGPT search returned partially or entirely incorrect sources for 153 of 200 publisher quotes and acknowledged it could not answer only 7 times.14 Citations and accuracy. Does not show: An early test of a new product; OpenAI called it an atypical test, and 40 of the quotes came from publishers that blocked its crawler. | 200 quotes from 20 publications | Prompt runs: block quotes from 10 articles per publisher, answers graded for publisher, date and URLEngines: ChatGPT Search | Data: November 2024 | Tow Center for Digital Journalism, Columbia Journalism ReviewSponsor: None statedSells an AEO product: no |
For every 1,000 Google searches in the U.S., 360 clicks went to the open web, and just under 60% of searches ended without a click.10 Clicks and CTR. Does not show: Collected before AI Overviews reached most U.S. searches; a baseline for zero-click behavior, not a measurement of AI answers. | devices in a multi-million device clickstream panel | Clickstream panel (Datos), desktop and mobile searches weighted by device shareEngines: Google Search | Data: January to May 2024 (mobile data from January 2024) | SparkToroSponsor: Data from Datos, a Semrush companySells an AEO product: no |
Adding an optimized strategic text sequence to a product page raised the product's rank in about 40% of 200 randomized LLM evaluations and left it unchanged in about 60%.27 GEO experiments. Does not show: A fictitious catalog and an open model, not a live search product; it measures manipulation, not a recommended practice. | 200 LLM evaluations with randomized product order | Adversarial text optimization (GCG) on a catalog of fictitious coffee machinesEngines: Llama-2 (open model) | Data: 2024 | Kumar and Lakkaraju, Harvard University (arXiv 2404.07981)Sponsor: None statedSells an AEO product: no |
By the end of 2023, 48% of the most widely used news websites in ten countries blocked OpenAI's crawlers and 24% blocked Google's AI crawler.31 Crawlers and llms.txt. Does not show: An end-of-2023 snapshot from before most publisher licensing deals; blocking rules do not show what crawlers obey. | 15 most used news sites in each of 10 countries | Archived robots.txt files from the Wayback Machine for every available day of 2023Engines: GPTBot, ChatGPT-User, Google-Extended | Data: 2023 | Reuters Institute for the Study of JournalismSponsor: None statedSells an AEO product: no |
Content edits such as citing sources, adding quotations and adding statistics raised source visibility in generative engine answers by up to 40% on the GEO-bench benchmark.25 GEO experiments. Does not show: Relative gains in a lab engine and on Perplexity (up to 37% there); the paper did not test effects on search rankings. | 10,000 queries (GEO-bench) | Controlled experiment: edits applied to one source, visibility measured in the engine's answer against an unedited baselineEngines: Research generative engine, Perplexity.ai | Data: 2023 to 2024 | Aggarwal et al., KDD 2024 (arXiv 2311.09735)Sponsor: None statedSells an AEO product: no |
Across four generative search engines, 51.5% of generated sentences were fully supported by their citations and 74.5% of citations supported their sentence.15 Citations and accuracy. Does not show: 2023 systems, some since discontinued; today's engines were not measured. | 1,450 queries | Human evaluation of each sentence and citation in the engines' answersEngines: Bing Chat, NeevaAI, perplexity.ai, YouChat | Data: 2023 | Liu, Zhang and Liang, Stanford University (arXiv 2304.09848)Sponsor: None statedSells an AEO product: no |
Footnotes 37
- 1Pew Research Center: Google users are less likely to click on links when an AI summary appears in the results
pewresearch.org
Metered browsing panel (KnowledgePanel Digital): every Google search on tracked devices and what the user did next. Sample: 900 U.S. adults (68,879 Google searches).
Back to text - 2Pew Research Center: Google users are less likely to click on links when an AI summary appears in the results
pewresearch.org
Metered browsing panel (KnowledgePanel Digital), all Google searches in the month. Sample: 68,879 Google searches by 900 U.S. adults.
Back to text - 3Ahrefs: AI Overviews Reduce Clicks by 34.5%
ahrefs.com
Before and after CTR from aggregated Google Search Console data, with informational keywords as the baseline. Sample: 300,000 keywords (150,000 with an AI Overview, 150,000 informational without).
Back to text - 4Ahrefs: Update: AI Overviews Reduce Clicks by 58%
ahrefs.com
Same method as the April 2025 study: aggregated Google Search Console CTR, before and after. Sample: 300,000 keywords (150,000 with an AI Overview, 150,000 informational without).
Back to text - 5Seer Interactive: AIO Impact on Google CTR: September 2025 Update
seerinteractive.com
Client organic and paid search data (25.1M organic and 1.1M paid impressions), split by AI Overview presence and citation. Sample: 3,119 search terms across 42 client organizations.
Back to text - 6Seer Interactive: AIO Impact on Google CTR: September 2025 Update
seerinteractive.com
Client search data split by whether the brand was cited or mentioned in the AI Overview. Sample: 3,119 search terms across 42 client organizations.
Back to text - 7Semrush: Semrush AI Overviews Study: What 2025 SEO Data Tells Us About Google's Search Shift
semrush.com
Clickstream panel from Datos (a Semrush company), keywords that gained an AI Overview between May and October. Sample: 200,000 keywords with search volume over 100 (200K+).
Back to text - 8Semrush: Semrush AI Overviews Study: What 2025 SEO Data Tells Us About Google's Search Shift
semrush.com
SERP tracking across the Semrush keyword database, month by month. Sample: 10,000,000 keywords (10M+).
Back to text - 9Ahrefs: 63% of Websites Receive AI Traffic (New Study of 3,000 Sites)
ahrefs.com
Referral data from an anonymized sample of sites running Ahrefs Web Analytics, seven chatbots. Sample: 3,000 websites.
Back to text - 10SparkToro: 2024 Zero-Click Search Study: For every 1,000 EU Google Searches, only 374 clicks go to the Open Web. In the US, it's 360.
sparktoro.com
Clickstream panel (Datos), desktop and mobile searches weighted by device share. Sample: devices in a multi-million device clickstream panel.
Back to text - 11Adobe: Adobe Analytics: Traffic to U.S. retail websites from Generative AI sources jumps 1,200 percent
blog.adobe.com
Adobe Analytics visit and transaction data, plus a survey of 5,000 U.S. consumers. Sample: 1,000,000,000,000 visits to U.S. retail sites (more than 1 trillion).
Back to text - 12Pew Research Center: 34% of U.S. adults have used ChatGPT, about double the share in 2023
pewresearch.org
Survey of Pew's American Trends Panel, recruited by random address sampling and weighted. Sample: 5,123 U.S. adults.
Back to text - 13Tow Center for Digital Journalism, Columbia Journalism Review: We compared eight AI search engines. They're all bad at citing news.
cjr.org
Prompt runs with direct article excerpts, each answer graded by hand for article, publisher and URL. Sample: 1,600 queries (20 publishers, 10 articles each, 8 tools).
Back to text - 14Tow Center for Digital Journalism, Columbia Journalism Review: How ChatGPT Search (Mis)represents Publisher Content
cjr.org
Prompt runs: block quotes from 10 articles per publisher, answers graded for publisher, date and URL. Sample: 200 quotes from 20 publications.
Back to text - 15Liu, Zhang and Liang, Stanford University (arXiv 2304.09848): Evaluating Verifiability in Generative Search Engines
arxiv.org
Human evaluation of each sentence and citation in the engines' answers. Sample: 1,450 queries.
Back to text - 16BBC: Representation of BBC News content in AI Assistants
bbc.co.uk
Four assistants given access to the BBC website answered news questions; expert journalists rated each answer. Sample: 362 AI responses reviewed by 45 BBC journalists.
Back to text - 17European Broadcasting Union and BBC: Largest study of its kind shows AI assistants misrepresent news content 45% of the time
bbc.co.uk
Journalists from 22 public service media organizations in 18 countries and 14 languages graded answers. Sample: 3,000 responses (more than 3,000).
Back to text - 18Ahrefs: 76% of AI Overview Citations Pull From the Top 10
ahrefs.com
Rank lookup for the three most visible citations in each AI Overview. Sample: 1,900,000 citations from 1M AI Overviews.
Back to text - 19Ahrefs: Top Brand Visibility Factors in ChatGPT, AI Mode, and AI Overviews (75k Brands Studied)
ahrefs.com
Spearman correlation between brand signals in Ahrefs' indexes and mention counts in AI answers. Sample: 75,000 brands.
Back to text - 20Profound: AI Platform Citation Patterns: How ChatGPT, Google AI Overviews, and Perplexity Source Information
tryprofound.com
Citation counts from the prompts Profound tracks across three platforms. Sample: 680,000,000 citations.
Back to text - 21Chen, Wang, Chen and Koudas, University of Toronto (arXiv 2509.08919): Generative Engine Optimization: How to Dominate AI Search
arxiv.org
Prompt runs through each engine's API; cited domains classified as brand, earned or social. Sample: ranking-style brand queries in two sets (size not stated for this experiment).
Back to text - 22Kai-Cheng Yang, Northeastern University (arXiv 2507.05301): News Source Citing Patterns in AI Search Systems
arxiv.org
Analysis of conversations from the AI Search Arena, a head-to-head evaluation platform. Sample: 366,000 citations (over 366,000) in 65,000 responses.
Back to text - 23Graphite: How Does AI-Generated Content Perform in Search and Answer Engines?
graphite.io
AI detector (Surfer) run on 500-word chunks of ranking and cited articles. Sample: 31,493 keywords across 10 categories (100 per category asked of answer engines).
Back to text - 24Muck Rack: What Is AI Reading? May 2026
media.muckrack.com
Prompt runs across a large consumer prompt set; every cited URL categorized. Sample: 25,000,000 links from AI responses (25,000,000+).
Back to text - 25Aggarwal et al., KDD 2024 (arXiv 2311.09735): GEO: Generative Engine Optimization
arxiv.org
Controlled experiment: edits applied to one source, visibility measured in the engine's answer against an unedited baseline. Sample: 10,000 queries (GEO-bench).
Back to text - 26Puerto et al., Parameter Lab and TU Darmstadt (arXiv 2506.11097, NeurIPS 2025 Datasets and Benchmarks): C-SEO Bench: Does Conversational SEO Work?
arxiv.org
Benchmark: C-SEO methods applied to documents, rank change measured in LLM answers for two tasks and six domains. Sample: 54 method, task and domain cases.
Back to text - 27Kumar and Lakkaraju, Harvard University (arXiv 2404.07981): Manipulating Large Language Models to Increase Product Visibility
arxiv.org
Adversarial text optimization (GCG) on a catalog of fictitious coffee machines. Sample: 200 LLM evaluations with randomized product order.
Back to text - 28Cloudflare (Radar): The crawl before the fall... of referrals: understanding AI's impact on content providers
blog.cloudflare.com
HTML requests from each platform's crawlers divided by HTML requests whose Referer is that platform. Sample: requests across Cloudflare's network.
Back to text - 29Cloudflare (Radar): The crawl-to-click gap: Cloudflare data on AI bots, training, and referrals
blog.cloudflare.com
Crawler traffic classified by purpose from operator disclosures and industry sources. Sample: requests across Cloudflare's network.
Back to text - 30Cloudflare (Radar): From Googlebot to GPTBot: who's crawling your site in 2025
blog.cloudflare.com
robots.txt files fetched by Cloudflare Radar, matched against a list of AI crawler names. Sample: 3,816 domains with a robots.txt, from the top 10,000.
Back to text - 31Reuters Institute for the Study of Journalism: How many news websites block AI crawlers?
reutersinstitute.politics.ox.ac.uk
Archived robots.txt files from the Wayback Machine for every available day of 2023. Sample: 15 most used news sites in each of 10 countries.
Back to text - 32BuzzStream: Which News Sites Block AI Crawlers in 2025? [New Data]
buzzstream.com
robots.txt disallow directives checked for 11 AI crawlers. Sample: 100 top news sites (50 U.S., 50 UK, deduplicated).
Back to text - 33SE Ranking: LLMs.txt: Why Brands Rely On It and Why It Doesn't Work
seranking.com
Check for llms.txt on each domain, then Spearman correlation, XGBoost regression and SHAP analysis. Sample: 300,000 domains (nearly).
Back to text - 34Rankability: llms.txt Adoption Research Report
rankability.com
Root file check of each domain, with blocked or inconclusive checks kept in the denominator. Sample: 10,000 domains from the Tranco list.
Back to text - 35Common Crawl Foundation: A Content Analysis of llms.txt Files from the July 2026 Crawl Archive
commoncrawl.org
Crawl experiment: /llms.txt added to the seed list for a random sample of hosts in the July 2026 crawl. Sample: 5,167,831 /llms.txt URLs from a random sample of hosts.
Back to text - 36Ahrefs: We Analyzed 137K Sites: 97% of llms.txt Files Never Get Read
ahrefs.com
Root file check plus bot and user-agent requests to the llms.txt path. Sample: 137,210 domains in Ahrefs Web Analytics with traffic.
Back to text - 37AEO Footnotes (Footnotes Lab): robots.txt and llms.txt across 118 AEO tools and agency sites: what we found
aeofootnotes.com
Fetch of /robots.txt and /llms.txt on each domain; blocking judged by the RFC 9309 group that applies. Sample: 118 domains.
Back to text
Cite this page
Footnotes Lab, AEO Footnotes. "AEO and AI search statistics, with sources." https://aeofootnotes.com/lab/aeo-statistics. Updated 2026-10-30. CC BY 4.0.
The rows, summaries and caveats are licensed under Creative Commons Attribution 4.0: reuse them with credit to AEO Footnotes and a link to this page. The studies themselves belong to their publishers. Spotted a study we should add, or a number that no longer matches its source? Write to hello@aeofootnotes.com.
Also seerobots.txt and llms.txt across 118 AEO tools and agency sites: what we foundAll Footnotes Lab pieces