What the researchers did
The study is the broadest independent audit of Google's AI Overviews published so far, and it is unusually transparent: the authors released their code and data. They took trending queries from Google Trends every day across 19 topic categories and ran each one in a fresh, signed-out Chrome browser from a single US location, Amazon's data centre region in northern Virginia. For every overview they expanded the full text, recorded each citation and the first page of ordinary results shown beside it, and then fetched the cited pages.
To check accuracy they broke each overview into single factual claims and asked a language model, xAI's Grok 4.1 Fast Reasoning at temperature zero, to label each claim against the full text of every cited page: clearly supported, vaguely supported, ambiguous (cited pages disagree), incorrect (contradicted) or omitted (not mentioned). They then had two people re-label a stratified sample of 100 verdicts. The humans agreed with each other almost perfectly and with the model on 98 of 100; both disagreements were cases where the model had been too generous, not too harsh.
The queries are trending searches, which skew towards sport, entertainment and news events. That matters for what the numbers can and cannot say about searches for a company or a product, which we return to below.
When an overview appears
Across the 40 days, 13.7 per cent of searches produced an overview. The form of the query mattered far more than its subject: 64.7 per cent of question-form searches got one, against 9.5 per cent of everything else. Among searches that were not questions, the university reports that the rate rose from 9.9 per cent for one-word queries to 38.7 per cent for six words or more.
Topic mattered too. Hobby and leisure searches triggered an overview 46.1 per cent of the time and science searches 39.9 per cent, against 7.5 per cent for politics and 9.6 per cent for law and government. Google says it holds back overviews on sensitive or fast-moving topics but publishes no criteria, and the authors read the pattern as undisclosed editorial discretion. The practical point is that the longer and more question-like the search, the more likely the searcher reads a summary first, which is also the advice in our guide to AI Overviews and clicks.
How often the summary matches the source
Most claims held up. Of 98,020 checked, 84.61 per cent were clearly supported and 4.36 per cent vaguely supported, 88.97 per cent in all. The failures were dominated by omission rather than contradiction.
Share of 98,020 claims from 7,491 AI Overviews, by verification label
Source: Xu, Iqbal and Montgomery, Measuring Google AI Overviews, arXiv:2605.14021v2, section 4.3, read in full 5 October 2026.
Labels assigned by a language model and validated against human review on a sample of 100. Social and video pages could not be read, so some claims marked omitted may be supported by them; the authors put the floor for unsupported claims at about 5.3%.
At the level of whole overviews the picture is less tidy. The median overview had 93.33 per cent of its claims supported, but only 41.9 per cent had every single claim supported. A reader who sees only the summary has no way of telling which sentence is the unsupported one, and an “incorrect” label here means a page that Google itself cited as support says something different.
Two corrections make the headline rate smaller, and the authors make both. First, they could not read social and video pages, so a claim drawn from a cited YouTube video or Reddit thread would be marked as omitted. 59.9 per cent of the unsupported claims came from overviews that cited at least one such page; if all of those were in fact supported, the unsupported rate would fall to about 5.3 per cent. Second, live data distorts some categories. Climate scored 48.2 per cent, almost entirely because weather and pollen figures had changed by the time the cited pages were fetched. Excluding real-time queries, 18 of 19 categories fell between 85.9 and 94.8 per cent. Health scored highest.
The study also found that source quality and claim accuracy were largely independent: a credible cited page did not make an accurate summary of it more likely. Earlier audits of AI search tools found far worse figures, including the citation tests described in how often AI search tools cite the wrong source, but those measured different things with different methods and are not directly comparable.
Which sources get cited
The 7,583 overviews carried 61,212 citations to 7,479 different hostnames, a median of eight per overview. The most cited were youtube.com at 5.49 per cent of citations, en.wikipedia.org at 4.39 per cent, facebook.com at 3.68 per cent, instagram.com at 3.65 per cent and usatoday.com at 2.80 per cent. The spread was wide: the 100 most cited hostnames accounted for 57.1 per cent of citations, against 72.3 per cent of links on the first page of results.
The finding with the most practical weight is how far the cited sources diverge from the ranking. On average only 31.5 per cent of an overview's cited domains also appeared in the top five results, 48.0 per cent in the top ten and 70.2 per cent anywhere on the first page. 29.8 per cent were not on the first page at all, and those off-page sources scored slightly higher for credibility than the on-page ones.
Average share of an overview's cited domains that also appear in the ordinary results for the same search
Source: Xu, Iqbal and Montgomery, Measuring Google AI Overviews, arXiv:2605.14021v2, section 4.2, averaged across 7,562 overviews, read in full 5 October 2026.
On credibility the authors used a published index that scores about 11,500 news domains from 0 to 1. Cited sources averaged 0.732 against 0.611 for the first page as a whole, and leaned far less on social platforms, 14.2 per cent of citations against 49.9 per cent of first-page links. The comparison flips at the very top: the first result alone averaged 0.777 and the top three 0.753, both slightly above the overview's sources. Only about 60 per cent of cited URLs had a score, because the index does not cover most social platforms and brand-owned sites. An earlier audit using a different measure had found overview sources less credible than search results; of the three alternative measures the authors tried, two confirmed their result and one showed a small, non-significant difference the other way.
What this means for a company that wants to be cited
Ranking and citation are related, not the same. Nearly a third of cited domains were not on the first page, so a page can be used in the summary without ranking near the top, and a top-ranked page can be passed over. That is consistent with what we set out in ranking in Google versus being cited by an AI.
Being cited is not the same as being represented accurately. The commonest failure was a claim that none of the cited pages contained, sitting next to citations that look like support. For a company, the risk is a sentence about it that its own page, or the coverage cited, never said. The check is to read the summary against the cited pages, not to assume the citation proves the sentence.
The query sample is not about companies. Trending searches are dominated by sport, entertainment and news events. Commercial searches, where ads also appear, are under-represented: only 2.16 per cent of pages with an overview carried a sponsored ad, and in 39 cases the ad sat above the overview. Nothing in the study measures searches for a brand, a product or an executive, so the accuracy rate for those is unknown.
The click is still the casualty. At least 50.6 per cent of cited pages carried display advertising, a lower bound because social pages were counted as ad-free. The study did not measure lost traffic, and says so.
Methodology and limitations
This data report summarises Xu, Iqbal and Montgomery, “Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact”, arXiv:2605.14021, version 2 of 1 October 2026, accepted to the ACM Internet Measurement Conference 2026. We read the full paper on 5 October 2026 and the Washington University news release of 29 September 2026. The figures for activation by query length and by topic, and the 41.9 per cent fully grounded share, are quoted from the release; all other figures are from the paper. We did not re-run the analysis or inspect the released data. The authors' limits apply: a single US vantage point, signed-out browsers, trending rather than commercial queries, social and video pages excluded from claim checking, a credibility index that covers about 60 per cent of cited URLs, and claim labels assigned by a language model and validated on a sample of 100. The claim-label percentages in the chart are computed from the paper's counts and percentages and may differ in the second decimal place from its table.
Conclusion
On a large and well-documented sample, Google's AI Overviews mostly say what their sources say, and their sources are on average more credible than the page of results around them. The exceptions are not rare: about one claim in nine, and more than half of all overviews, contained something the cited pages did not support, though the true figure is probably lower once unreadable social pages are allowed for. The more useful finding for anyone hoping to be cited is that the summary draws on its own pool of sources, nearly a third of them absent from the first page. Ranking well is not a guarantee of being cited, and being cited is not a guarantee of being summarised correctly.
Frequently Asked Questions
Who ran the AI Overviews audit, and when?
Haofei Xu, Umar Iqbal and Jacob Montgomery of Washington University in St. Louis. They ran 55,393 trending US searches from 13 March to 21 April 2026, and the paper is accepted at the ACM Internet Measurement Conference in October 2026.
How often do AI Overviews appear?
On 13.7 per cent of the searches in the study, rising to 64.7 per cent for searches phrased as questions and falling to 7.5 per cent for political searches.
How accurate are AI Overviews?
88.97 per cent of 98,020 claims were supported by a cited page. 6.98 per cent were found in no cited page, 2.66 per cent were contradicted by one and 1.39 per cent drew conflicting support. The authors treat the 11 per cent failure rate as an upper bound.
Do AI Overviews cite the pages that rank first?
Often not. Only 31.5 per cent of cited domains also appeared in the top five results, and 29.8 per cent did not appear on the first page at all.
Are AI Overview sources credible?
On average more credible than the first page of results as a whole, and much less reliant on social platforms. The very top results scored slightly higher than the overview's sources.
Does this tell me how AI Overviews describe my company?
No. The study used trending searches dominated by sport, entertainment and news, not searches for companies or products. It shows what kinds of error occur, not how often they occur in searches about a particular business.
Sources
- academic Haofei Xu, Umar Iqbal and Jacob M. Montgomery, Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact, arXiv:2605.14021v2, 1 October 2026; accepted to IMC 2026. Read in full 5 October 2026.
- official Sara Savat, Washington University in St. Louis, WashU research finds gaps in Google AI Overview citations, claims, 29 September 2026. Activation by query length and topic; fully grounded share. Read 5 October 2026.
- our research Media Morpheus Research, Do AI Overviews Reduce Clicks?, 23 September 2026.
- our research Media Morpheus Research, How Often Do AI Search Tools Cite the Wrong Source?, 23 September 2026.