Research
Most statistics about AI search are summaries: a vendor runs some prompts and publishes a percentage, and nobody outside the company can check it. We do it the other way round. Every study below has its per-run data published in full, so you can open the file and tell us we read it wrong.
The studies, and where their data is
Nine studies. Four were run by hand: together 129 logged engine runs and 38 pages measured. One ran 150 scripted requests through the Claude API, one put 50 questions to five AI engines and checked 3,465 figures, one asked six AI engines about 20 brands and checked 3,640 facts, and two are case studies of this site. A run is one question put to one engine once.
What AI tells the world about Türkiye's brands
Six AI engines got 93.2% of basic brand facts right; most traceable errors were on a cited page
3,391 of 3,640 checkable claims (93.2%) matched the official source, and in 110 of 164 traceable errors (67%) a page the answer cited stated the same wrong value.
DOI 10.5281/zenodo.24040978 · Data and code · Turkish edition at /tr/yapay-zeka-turk-markalari
Read the study →AI Answer Evidence Index
By engine, the median answer has 57% to 92% of its figures on a cited page
On the 41 questions all five engines answered with a figure, the median share of an answer's figures found on a cited page runs from 57% to 92%.
DOI 10.5281/zenodo.23238429 · Preregistered for three engines (10.5281/zenodo.23145556) · Data and code
Read the study →Cross-Engine Citation Study
Four engines agree on the brands, not on the sources
The four engines agreed on 32% of the brands they recommended, and 1% of the sources they cited.
DOI 10.5281/zenodo.23427370 (rounds one and two) · Per-run CSV and JSON · Round three tables and JSON
Read the study →Global Citation Case Study
Cited in 7 countries in four months
Between June and September 2026, 15 publishers in seven countries cited or covered the site. Ten cite a finding, a method or a guide, and eight of those ten use a figure we measured or a method we built.
15 citing pages, each linked · Open data
Read the case study →AI Visibility Case Study
89K AI citations in three months
In the 91 days to 7 September, Bing counted about 89,000 citations of the site in AI answers from Microsoft Copilot and its partners. Four guides earned 94% of them, and one guide alone earned 42,092.
DOI 10.5281/zenodo.23427102 · Crawler data and code · Full-text PDF on ResearchGate
Read the case study →Claude Retrieval and Citation Study
When the query names your brand, its own site is usually not what the search returns
The brand's own site was among the retrieved domains in 36 of 108 run-brand pairs, 33.3%. Twelve of twenty-five named brands were never retrieved in any run.
DOI 10.5281/zenodo.22651233 · Data, code and protocol
Read the study →AI Search Evidence Index
Two numbers in three have no link to a source
Two numbers in three on AI-visibility pages have no link to a source a reader can reach.
DOI 10.5281/zenodo.22257563 · Data, code and instrument history
Read the study →AI Overview Industry Study
Every industry is answered from a different set of sources
Review and authority media appeared in 24 of 27 AI Overviews, but which outlets changed completely by industry.
DOI 10.5281/zenodo.23427380 · Per-search CSV and JSON · Method notes
Read the study →AI Overview Cited-Domains Study
Reddit is the source Google's AI Overview reaches for first
Reddit was cited in 13 of the 21 AI Overviews, more than twice the next site.
DOI 10.5281/zenodo.23427378 · Per-search CSV and JSON · Method notes
Read the study →Generative Engine Optimization Statistics
The most-recycled GEO stats do not say what they are quoted as saying
The most-recycled figures do not say what they are quoted as saying: a 525% revenue jump is one publisher's own number, and 71.5% describes surveyed marketers, not consumers.
Read the study →Collection conditions, limits and raw data
What AI tells the world about Türkiye's brands
Collected 15 to 18 October 2026 from a connection in Türkiye. ChatGPT, Gemini, Google AI Overviews, Bing Copilot, Perplexity and Claude were asked about 11 brands from Türkiye and 9 European control brands, in Turkish and in English, in two runs. Every checkable claim was compared with a fact card built from the companies' official sources, and the pages each wrong answer cited were read to see where the wrong value came from.
Limit. Rates are pooled counts on 20 brands, five template questions and two languages, through free consumer interfaces on the collection dates. The study accuses no brand or company.
AI Answer Evidence Index
Round two of the Evidence Index series, collected 5 to 7 October 2026. 50 price and cost questions were put to Google AI Overviews, ChatGPT, Gemini, Perplexity and Claude (API), and 3,465 figures in 241 answers were looked for on the pages those answers cite. The method was preregistered for the first three engines before the first answer was collected; Perplexity and Claude (API) were added afterwards.
Limit. The check matches numbers. A figure that is not found is not thereby false, a found figure is not thereby supported by its page, and engines whose intervals overlap are not ranked.
Cross-Engine Citation Study
Round three (October 2026). A different question set: 50 price and cost questions, put to Google AI Overviews, ChatGPT, Gemini, Perplexity and Claude (API). On the 40 questions where all five cite a source, all five cite the same page in 2 and at least two of them in all 40. This round counts sources only, not brands, and is not pooled with the two earlier rounds.
Round two (August 2026). All 48 runs: 189 recommended brands and 322 cited sources, one row per engine-question pair. Brands are listed in the order the engine named them; every source carries a type code (vendor site, review media, independent listicle, Reddit, forum, YouTube, reference). The JSON carries the same rows plus the run conditions, the source-type key and the stated limits, so the file stands on its own if it travels without this page.
Round one (19 June 2026). Smaller and less even than August: Perplexity was run on all twelve questions, the other three engines on fewer. Several Perplexity source panels were truncated behind a “show more” control we could not fully expand, so the recorded list is what was visible and written down at the time. The churn figures quoted across this site (80 sources in June, 42 still cited in August, 48% turnover) are computed from this file and the August one.
Global Citation Case Study
One row per publisher. A page is counted only after it has been opened and the passage that names or links the site has been read. Each page is dated by its own publication date. The country is where the publisher is based. Our own profiles, directory entries and our other sites are left out.
AI Visibility Case Study
Bing. The AI Performance report in Bing Webmaster Tools for ismybrandinai.com, read on 9 September 2026: the 91-day total, the daily series, the page view and the grouped phrases. The report displays daily values of 1,000 or more rounded to the hundred, and we use them as displayed. Page and phrase figures come from the report's own export. Bing describes the data as a sample of overall citation activity, designed for trend analysis rather than precise accounting.
Cloudflare. The GraphQL Analytics API, dataset httpRequestsAdaptiveGroups, for 3 to 9 September 2026, read on 10 September. We track 20 AI and search crawler names. A request counts as verified when Cloudflare places it in a verified bot category. A request counts as asking for a hidden file when its path has a segment that starts with a dot, other than the public /.well-known/, or matches a short list of well-known secret files.
Claude Retrieval and Citation Study
Fifty B2B-SaaS buyer questions, three runs each, 150 requests against claude-sonnet-5 with the web search tool called directly, so the search results come back alongside the answer and the retrieved list can be recorded separately from the cited one. Twenty-three of the fifty queries name a brand in the query text, which is where the headline figure comes from: which brand is at issue needs no interpretation.
Limit. The surface measured is the Claude API's web search tool, not claude.ai. Whether the two use the same search provider is not documented by Anthropic. In the first pass, twenty runs' searches failed silently and were re-run about twenty minutes later; that defect is published with the data rather than quietly corrected.
AI Search Evidence Index
Round one, dataset frozen 30 August 2026. 38 pages that rank for AI-visibility queries were sampled, 32 of them retrieved inside one 84-second window, and the 13 that carried enough numeric claims were scored. Every credited block on the scored pages was opened and read one at a time.
The dataset is deposited at Zenodo with a DOI, and the instrument's full revision history is published with it, so a later round can be compared against this one.
AI Overview Industry Study
30 searches across six industries, collected by hand on 11–12 June 2026. Twenty-seven produced an AI Overview. Each row carries the source count, whether Reddit appeared, and the site types recorded for that answer.
Limit. Site type was assigned by hand from each source's title and brand, not by reading the page. That classification is our judgement, and the rows are published so you can disagree with any single one of them.
AI Overview Cited-Domains Study
24 buyer searches, collected by hand on 10 June 2026. Twenty-one of the twenty-four produced an AI Overview; the other three returned a Shopping grid. Each row carries the domains the source panel listed, which is how the most-cited table on that page is counted. The rows themselves were not downloadable until now.
Limit. Site type was assigned by hand from each source's title and brand, not by reading the page. That classification is our judgement, and the rows are published so you can disagree with any single one of them.
Generative Engine Optimization Statistics
This page is not one of our measurements. It reads other people's published figures back against their sources and records, for each one, what the source actually says, the year, and the sample it rests on. It is listed here because it belongs to the same discipline as the studies, not because the numbers in it are ours.
Cited by
Publishers in fourteen countries use these studies and guides in their own work. See who cites them, and where.
Everything here is CC BY 4.0. Eight studies sit behind this page, so cite the one your number came from rather than the page as a whole; each study page carries its own attribution and its own conditions. The shared method notes are on the methodology page, and journalists or researchers who want the working notes can get in touch.