A new peer-reviewed study of Google’s AI Overviews found that nearly 30% of the pages the AI feature cites do not appear anywhere on the first page of regular search results for the same query. The research, presented at the 2026 ACM Internet Measurement Conference, also found that only 41.9% of AI Overviews were fully supported by the sources they cited.
For anyone working on an AI Overviews citation study of their own, or just trying to figure out why a competitor with weaker rankings keeps showing up in Google’s AI answers, this research answers the core question directly: ranking on page one is not a requirement for getting cited. It helps, but it is not the whole story.
The paper, titled “Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact,” comes from Haofei Xu, Umar Iqbal, and Jacob M. Montgomery at Washington University in St. Louis. The researchers collected data between March 13 and April 21, 2026, and presented their findings at IMC ’26 in Karlsruhe, Germany, from October 12 to 16, 2026.
What the AI Overviews Citation Study Measured
The team ran 55,393 trending queries through Google Search and logged every case where an AI Overview appeared. They then compared the pages cited inside each AI Overview against the ten organic results shown on the same search results page.
Four things were measured: how often an AI Overview shows up at all (activation), how credible the cited sources are compared with regular results (source quality), how well the AI Overview’s claims hold up against the pages it cites (claim fidelity), and what happens to publishers whose pages get cited (publisher impact).
This is one of the larger independent audits of AI Overviews to date, and it is useful precisely because it comes from a university research team rather than an SEO tool vendor with content to sell.
Activation Rates Varied Heavily by Topic
An AI Overview appeared for about 13.7% of the trending queries tested, or 7,583 out of 55,393. That overall number hides a wide spread. Question-style queries triggered an AI Overview 64.7% of the time, compared with 9.5% for other phrasing.
Topic mattered just as much. Hobbies and Leisure queries triggered an AI Overview 46.1% of the time. Beauty and Fashion queries triggered one only 3.5% of the time. Politics sat lower still, at 7.5%. The researchers describe these gaps as evidence of undisclosed editorial decisions about when Google chooses to show an AI-generated answer at all.
Why AI Overviews Citations Often Skip Page One
This is the finding most relevant to anyone doing GEO work. The study found that 29.8% of the domains cited inside AI Overviews do not appear anywhere among the first page of regular organic results for that same query.
In other words, roughly three out of every ten citations come from pages that a rank tracker would never flag as a page-one competitor. The AI Overview is pulling from a different, broader pool of pages than the one shown in the traditional blue links.
User-generated content tells a related story in reverse. It made up 49.9% of first-page organic results but only 14.2% of AI Overview citations. Google’s AI layer is pulling proportionally less from forums and UGC platforms than classic search results do, even in queries where that kind of content ranks well normally.
Source credibility also came out higher for AI Overview citations. Using a continuous credibility score, cited sources averaged 0.732 versus 0.611 for standard first-page results. That gap narrowed, and in some comparisons reversed, when the researchers looked only at the very top-ranked organic results instead of the full first page. The practical reading: Google’s AI Overviews system appears to weigh source credibility on its own terms, not simply by borrowing the existing ranking order.
This lines up with what DevByteDaily covered when Preferred Sources usage jumped 7x inside AI answers earlier this month. Google has been building out a separate source-selection layer for AI features, and this new research gives the first independent, large-sample look at how that layer actually behaves.
How Accurate Are Google’s AI Overviews
The researchers broke every AI Overview down into individual factual claims, then checked each one against the pages Google cited as support. Out of 98,020 claims checked, 89.0% were consistent with their cited sources.
The remaining 11.0% split into three failure types. Omission was the largest, at 7.0% of claims, meaning the cited source never actually addressed the claim at all. Direct contradiction between a claim and its cited source accounted for 2.7%. The last 1.4% came from cases where multiple cited sources disagreed with each other.
Only 41.9% of AI Overviews were fully grounded, meaning every single claim inside them was fully supported by a cited source with no gaps. More than half contained at least one claim that wasn’t properly backed, even when the overall overview was mostly accurate.
Where AI Overviews Claims Break Down
The researchers flag that real-time categories like weather and school closures likely inflate the error count, since a page can change between when Google generates the AI Overview and when a researcher crawls the cited page to check it. They excluded 25 AI Overviews covering 248 claims in Jobs & Education, Technology, and Shopping for this reason, using a keyword-based filter they admit is imperfect.
Even after accounting for that, omission, not outright fabrication, is the dominant failure mode. Google’s AI Overviews are more likely to state something a cited page never mentions than to flatly contradict a source. That distinction matters for anyone fact-checking AI-generated content, a topic DevByteDaily has covered in detail around Google’s own guidance to manually fact-check AI content before publishing.
The Publisher Impact: Ads Still Follow the Citations
The study also looked at what happens to the publishers whose pages get cited. At least 50.6% of the 61,212 cited URLs displayed advertising, a figure the authors call a lower bound since some ad-serving methods are harder to detect through automated crawling.
Google’s own sponsored ads appeared on the search results page in 2.2% of cases where an AI Overview was present, and in 39 of those instances, a Google ad was placed above the AI Overview itself.
The researchers argue that AI Overviews likely reduce the clicks publishers receive while leaving Google’s own ad inventory largely intact, since the AI Overview sits above both the ads and the organic results on most queries. They are careful to note they did not directly measure traffic loss. That claim would need separate click-through data that this study did not collect.
How This AI Overviews Citation Study Compares to Other Research
This is not the only recent attempt to measure AI Overviews accuracy. Earlier in 2026, an AI startup called Oumi ran a separate analysis for The New York Times, testing 4,326 searches against the SimpleQA benchmark. That analysis found accuracy around 85% when Google was using Gemini 2, rising to about 91% after Gemini 3 rolled out.
Google pushed back hard on the Oumi numbers, with a company spokesperson calling the study’s methodology flawed and arguing that SimpleQA questions don’t reflect how people actually search. Critics of the Oumi study also pointed out that it relied on its own AI model to grade correctness, which introduces its own error margin.
The Washington University research is a useful counterweight because it uses a different method entirely. Instead of grading whole answers against a benchmark, it breaks every AI Overview into atomic claims and checks each one against the specific page Google cited as support. Its 89.0% claim-level consistency rate lands close to the Oumi study’s answer-level accuracy range, which is a meaningful point of agreement between two independent teams using different techniques.
Where the two studies diverge is scope. Oumi focused narrowly on factual accuracy. This AI Overviews citation study goes further, covering which pages get cited in the first place, how those citations compare with regular rankings, and what happens to publishers afterward. That broader scope is what makes the page-one finding possible, since the Oumi work did not examine ranking position at all.
What This Means for Your GEO Strategy
This section is interpretation, not the study’s own findings.
In my experience advising on GEO work, the single biggest misconception teams carry over from traditional SEO is that AI Overviews citation is just a downstream reward for ranking well. This research is some of the best independent evidence yet that this assumption is wrong roughly three times out of ten.
If AI Overviews draw from a wider source pool than page-one rankings, then a page can be well-structured, clearly sourced, and topically precise without ever cracking the top 10, and still get pulled into an AI answer. That changes how you should prioritize content work. Chasing a top-3 ranking on a competitive head term may matter less for AI visibility than making sure a page states its claims in a clear, directly citable way, which is the same principle behind the 7-step content strategy framework for AI search DevByteDaily published earlier.
The claim fidelity numbers point to a second priority: clarity over cleverness. If omission is the main way AI Overviews claims fail, then pages that bury their key facts in dense paragraphs, behind interactive widgets, or several scrolls down are more likely to get paraphrased incorrectly or skipped as a source entirely, even when they contain the right information.
How to Improve Your Odds of an AI Overviews Citation
Based on what this study measured, a few practical adjustments follow directly from the data, without requiring guesswork:
- State the specific fact or number you want cited in a short, standalone sentence near the top of the relevant section, not buried in a long paragraph.
- Avoid splitting one factual claim across two sources on the same page, since the study found multi-source contradiction as a distinct failure mode.
- Don’t assume a page needs a top-10 ranking before it’s worth optimizing for AI citation. The credibility and clarity of the content itself appears to carry independent weight.
- Keep time-sensitive facts, like prices, availability, or hours, clearly dated or marked as current, since the study flagged these as the most error-prone category.
- Review your own published claims the way this study’s researchers did: pick a handful of your pages, check whether each factual statement is supported in a way an automated system could verify without inference.
None of this replaces fundamentals like accurate content and clean structured data. It does suggest that ranking position and AI Overviews citation are more loosely coupled than most GEO advice currently assumes, which is worth testing against your own site’s AI Overviews appearances if you’re tracking them. If you want a more structured way to work through these priorities, DevByteDaily’s free GEO Playbook eBook covers visibility tactics for both AI Overviews and AI Mode in more depth.
Citations inside Google’s AI Overviews increasingly come from pages outside the traditional top-10 rankings, according to the new AI Overviews citation study.
FAQ: AI Overviews Citation Study
Does my page need to rank in the top 10 to appear in AI Overviews?
No. The study found that 29.8% of AI Overviews citations came from domains that did not appear anywhere on the first page of organic results for the same query, so top-10 ranking is not a requirement.
How accurate are Google’s AI Overviews, according to this research?
Researchers found 89.0% of individual claims inside AI Overviews were consistent with their cited sources, while 11.0% showed some form of inconsistency, most often an omission where the cited page never addressed the claim.
Who conducted this AI Overviews study?
Haofei Xu, Umar Iqbal, and Jacob M. Montgomery of Washington University in St. Louis conducted the research and presented it at the 2026 ACM Internet Measurement Conference (IMC ’26) in Karlsruhe, Germany.
Do AI Overviews hurt publisher traffic, based on this study?
The study measured that at least 50.6% of cited pages carry advertising and that Google’s own ads remain largely intact alongside AI Overviews, but the researchers did not directly measure click-through or traffic loss, so that specific claim remains unverified by this paper.
Why This Matters
This is the first large, independent, peer-reviewed look at how AI Overviews actually select and cite sources, rather than a vendor study or a single company’s self-reported numbers. The headline figure to remember: ranking position and AI Overviews citation are correlated, not identical.
What to do next:
- Audit a sample of your own pages for standalone, clearly stated factual claims rather than assuming structure alone will carry the day.
- Stop treating AI Overviews visibility as purely a function of your existing rank tracker data.
- Revisit Google’s own guidance on AI features and your website for how Google says it selects and attributes sources, and compare it against what this independent study actually measured.
Leave a response