Nearly a third of AI Overview citations are not on page one

What the study measured, and over what sample

The paper is Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact by Haofei Xu, Umar Iqbal and Jacob M. Montgomery of Washington University in St. Louis, posted to arXiv as 2605.14021v1 under cs.CY on 13 May 2026. The collection window ran 40 days, from 13 March to 21 April 2026.

The method decides what the numbers are evidence of, so it is worth stating before any of them. The authors navigated the US localised Google Trends dashboard every 24 hours, exported the trending queries for 19 topical categories, then issued each query to Google in a fresh, stateless browser instance driven by a Puppeteer crawler on AWS Lambda. The crawler waited for the networkidle0 event, meaning no outstanding network requests for 500 milliseconds, before deciding whether an AI Overview was present on the rendered page. For each result it captured the overview, its embedded reference citations, the first page results shown alongside it, and the content and advertising of every cited page.

Two consequences follow, and the paper is direct about both. The sample is trending queries rather than a random sample of everything people search, so it carries whatever the public was interested in during those 40 days: Sports alone accounts for 51.4 percent of the corpus and Entertainment for another 14.9 percent. And the measurement is of a rendered results page as a browser receives it, which is the opposite end of the pipe from what an AI crawler access check looks at. The two answer different questions and neither substitutes for the other.

How often does an AI Overview actually appear?

Of the 55,393 queries, 7,583 returned an AI Overview. That is 13.7 percent, and it is the first figure to hold onto, because a great deal of writing about generative engine optimisation implies the surface is everywhere. On this sample, in these categories, during this window, it was on roughly one search in seven.

Google says something compatible in its own words. Its AI features documentation, carrying a last updated date of 10 December 2025, states that AI Overviews are only shown when its systems determine that it is additive to classic Search, and as such, often do not trigger. A measured 13.7 percent and a vendor sentence saying the feature often does not fire are the same claim approached from two directions, which is about as much corroboration as this field usually offers.

The aggregate hides the structure that is actually useful. The study classified a query as question form when its leading whole word is one of 15 interrogatives: who, what, where, when, why, how, which, is, are, was, can, do, does, did and has. Question form queries triggered an overview 64.7 percent of the time. Everything else triggered one 9.5 percent of the time. That is a 6.8 times difference within one corpus over one window.

By category the spread is wider still, a factor of 13. Hobbies and Leisure sat at 46.1 percent and Science at 39.9 percent; Beauty and Fashion at 3.5 percent and Climate at 7.4 percent. Politics came in at 7.5 percent and Law and Government at 9.6 percent, which the authors read as consistent with Google exercising caution on sensitive topics, while noting that Health at 26.6 percent shows the caution is not applied evenly across everything sensitive.

Be careful what you carry away from that. The phrasing that predicts an overview is the phrasing of the searcher's query, not the phrasing of your headings. Nothing in the paper says a page written as a question is more likely to be cited. What the split tells you is where the surface exists at all, which is a targeting question rather than an on-page one, and a reason to know which of your terms get asked as questions before spending anything on answer engine optimisation.

Where AI Overview citations come from, if not page one

Across the 7,583 overviews, Google cited 61,212 reference URLs drawn from 7,479 unique hostnames, a median of 8 references per overview with a range of 1 to 32. Those references are not a re-ranking of the results underneath them. Averaged across the overviews, 25.0 percent of reference domains overlapped with the top five first page results, 41.4 percent with the top ten, and 70.2 percent with the full first page. The remaining 29.8 percent of reference domains appeared nowhere on the first page for that query. At URL level the same gap reads 28.5 percent, 17,451 of the 61,206 references with a parseable host.

The authors' reading is that the AI Overview system draws on a source pool and a prioritisation distinct from Google's own ranking algorithm. Google's documentation describes a mechanism that fits without contradicting them: both AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics and data sources, which the same page says allows Google to display a wider and more diverse set of helpful links associated with the response than a classic web search shows. A fan-out across subtopics will surface pages that the original query never ranked.

The study also asked whether the two pools differ in quality, scoring domains with the continuous PC1 credibility measure from Lin and colleagues, published in 2023, which is the first principal component of expert and crowd ratings of news domain credibility on a 0 to 1 scale covering 11,520 domains. Across the 37,020 references that resolved to a score, mean credibility was 0.732, against 0.645 across 159,752 matched first page URLs, a gap of 0.087. The authors note this runs against earlier work finding that AI Overviews drew on lower quality sources, so it is a contested result rather than a settled one.

For anyone buying or selling AI visibility work, the shape of this is uncomfortable in a useful way. A rank tracker reports on the pool the overview partly ignores. Page one is not a prerequisite for the citation above it, and it is not a guarantee of one: a 70.2 percent overlap in one direction still leaves a great many first page results that were never cited. Two systems, two objectives, correlated but not identical, which is the distinction the reference page for Google AI Overviews exists to hold.

One claim in nine was not supported by the page it cited

The third measurement is the one to read slowly, because it concerns what the answer says rather than where it points. The authors decomposed each overview into atomic claims, one fact per claim, and checked every claim against the content of the pages that overview cited, using Grok 4.1 Fast Reasoning at temperature 0 and five labels: Clear, Vague, Ambiguous, Incorrect and Omitted.

Across 7,491 verifiable overviews the pipeline returned 98,020 claim level judgments. Of those, 87,204, or 88.97 percent, were consistent: 84.61 percent Clear and 4.36 percent Vague. The remaining 11.03 percent split into Omitted at 6.98 percent, where no cited source mentions the claim at all, Incorrect at 2.66 percent, where a cited source explicitly contradicts the claim it is cited for, and Ambiguous at 1.39 percent, where the cited sources disagree with each other. Omission is roughly 2.6 times more common than contradiction. The median overview had 93.33 percent of its claims grounded, but the tail is real: 205 overviews, 2.74 percent, had fewer than half their claims grounded, and 64, or 0.85 percent, had none at all.

The paper bounds its own error, and that bound should travel with the figure wherever it goes. The pipeline does not extract content from social and video platforms, so a claim supported only by an uncrawled forum or video page scores as unsupported. The authors quantify the exposure: 59.9 percent of inconsistent claims, 5,658 of 9,449, appear in overviews citing at least one such source, and they flag their Climate category number as a measurement artefact of the same gap rather than an AI Overview failure. Publishing the number that weakens your own headline is the part worth copying, and it is the same discipline behind what the evidence actually says about llms.txt, which reports a finding inconvenient for a tool this site ships.

There is a direct consequence for a site that does get cited. A claim can carry your domain as its source while saying something your page does not say, and a reader who sees only the summary has no way to detect that. It is not a crawl problem, and no crawler side check will ever surface it. It is a reason to read the overviews that cite you, which needs no tool at all.

Who pays for the pages an AI Overview reads

The fourth measurement is economic, and the authors are careful about what it is not. They state that they do not measure traffic loss directly. What they measure is the monetisation of the pages the overviews draw on. Of the 61,212 cited reference URLs, 30,994, or 50.63 percent, displayed visible ads. They call that a conservative lower bound, because a further 14.2 percent of references point at social and video platforms their pipeline does not scrape for advertising markup.

Set against that, the overviews barely displace Google's own inventory. Of the 7,583 results pages carrying an overview, 164, or 2.16 percent, also carried at least one Google sponsored search ad, and 39, or 0.51 percent, placed a sponsored slot above the overview block. No sponsored slot appeared inside the overview container. Where both appear they occupy separate regions of the page, so on this corpus the overview is additive to Google's ad inventory rather than a substitute for it. The authors attribute the low 2.16 percent to a query sample dominated by informational and trending searches rather than commercial ones, which is the right caveat to attach.

One further contrast from the same data set is worth carrying, because it cuts against a common assumption about AI answers. User generated content made up 14.2 percent of AI Overview references, 8,688 of 61,212, against 41.4 percent of first page result URLs, 127,814 of 308,407, on the authors' definition of platforms without strong editorial moderation, which excludes Wikipedia. On this corpus the overview cited proportionally less forum and social content than the results page did, not more.

None of that is a claim about your traffic and this post will not turn it into one. It is a description of the pages that get cited: more often than not, on this sample, they are funded by the page view the answer above them can absorb. If that is your business model, the figure you need comes from your own analytics and from Search Console, not from an external scan, and why no external scan can substitute is an argument this blog has already had to make about a different Google control.

What can you actually check on your own site?

Everything above was measured on Google's results pages by somebody else. The honest division of what a scanner adds to it is short, and stating it plainly is more useful than implying the gap is not there.

Start with the eligibility condition Google publishes, because it is first party and it is the shortest path. The AI features documentation says that to be eligible to appear as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet, and that there are no additional requirements beyond that. Indexed and snippet eligible is a Googlebot question and a robots meta question, and it is answered inside Search Console rather than by any third party tool, this one included.

What an external scan does answer is whether your text is in the response at all. That is prose parity: the share of your rendered visible text that was already present in the raw HTML. A page that only assembles its content in the browser is a page whose sentences may not exist for any retrieval system that does not run JavaScript, and the failure is silent rather than noisy, which is what makes it worth measuring rather than assuming. What GPTBot sees renders that comparison directly for a single URL, and six real storefront captures show two stores on the same platform producing opposite outcomes on it.

Then check access per crawler rather than as one verdict, because a robots.txt file that admits one fleet and refuses another is the ordinary case rather than an exotic one. The AI crawler checker tests each token separately, the robots.txt tester answers the same question for a specific path and agent, and the bot reference lists which tokens those are. None of them is Googlebot, and that limit is stated here rather than worked around.

After that comes the requirement every answer surface shares, whatever its retrieval looks like: whether a machine can work out what the page is about and who published it. Structured data is what carries that, and our methodology sets out what it is worth in the composite score and why it is worth that rather than more.

Last, the discipline this post has been trying to demonstrate. A grounded model answer is evidence about how models describe your category, not about Google Search, which is the argument in a Gemini citation tool is not an AI Overview tool. A per crawler figure counts requests that claimed to be a crawler, which is the argument in a user agent is a claim, not an identity. Named crawlers with published user agents are the reason the ChatGPT reference page and the Perplexity page can be more concrete than the Google one, and the research page publishes what our own sample supports with the sample size attached. If a vendor tells you it measures your AI Overview presence, ask what it requested and what came back. Anybody who knows their own pipeline can answer that in one sentence.