Google AI Overviews reach over 2 billion monthly users. When the overview says something, most people take it as the answer. A peer-reviewed study by researchers at Washington University in St. Louis, presented at the October 2026 ACM Internet Measurement Conference, checked whether those answers are actually supported by the sources they claim to cite.
The short version: about 11% of the time, they are not.
The study
The researchers ran 55,393 Google searches on trending topics over 40 days, from March 13 through April 21, 2026. For each search that triggered an AI Overview, they collected the overview text, the embedded citation links, the first-page organic results, and the full content of every cited webpage.
They decomposed each AI Overview into individual atomic claims, then checked each claim against the cited source page. A claim like "the average adult should sleep 7-9 hours per night" was verified by searching the cited page for supporting text.
The total: 98,020 verifiable claims across 61,212 citations.
The numbers
- 89% of claims were clearly or broadly supported by the cited source.
- 7% were not addressed in the cited text. The source page was fetched, but the claim could not be found in it.
- 4% directly contradicted the cited source. The overview said one thing, the source said another.
- Only 41.9% of overviews were fully grounded, meaning every verifiable claim was supported by the available cited text.
The researchers cautioned that the 11% error rate should be treated as an upper limit. Their pipeline could not fully capture content from Reddit threads, forum posts, and YouTube videos, which are among the most-cited domains in AI Overviews (see the Ahrefs ranking). Rapidly changing information like weather forecasts and school closings may also have been updated between the search and the page capture.
But even as an upper limit, 11% is significant. One in nine claims in an AI Overview may not be supported by the source it points to.
What "not addressed" means
The 7% of claims that were "not addressed" is the more interesting category. These are not cases where the AI hallucinated a citation. The citation exists, the page exists, but the specific claim cannot be found in the page content.
This can happen for several reasons:
Fan-out retrieval. Google splits the original query into sub-queries and retrieves pages for each. A claim in the overview might be synthesized from multiple sub-query results, but the citation points to only one of them. The claim is not wrong, but the cited source does not contain it by itself.
Content extraction failure. The cited page contains the information, but it is inside JavaScript-rendered content, embedded media, or a format the researchers' pipeline could not parse. The claim is supported, but the extraction missed it.
Synthesis beyond the source. The AI generated a claim that is a reasonable inference from the source but not explicitly stated. "The study suggests that..." becomes "The study confirms that..." in the overview. The source suggests, it does not confirm.
What "contradicted" means
The 4% of claims that directly contradicted the cited source is the more troubling category. The overview says X, the source says not-X.
The researchers did not publish a breakdown of what kinds of claims were contradicted, but the study focused on "Your Money or Your Life" queries, which cover health, finance, and safety. A contradiction in a health query is not a minor error.
The publisher revenue angle
The study also examined the financial impact. More than half of AIO-cited pages carry display advertising. When an AI Overview answers a question directly, the user has less reason to click through to the source page. The publisher loses the visit and the ad revenue, while Google's own sponsored ads continue to appear on the same page, sometimes above the overview itself.
The researchers found that 2.2% of search pages containing an AI Overview also displayed a Google-sponsored ad, including 39 pages where an ad appeared above the overview. Google captures ad revenue from the page while the publisher whose content informed the answer loses the click.
As Montgomery, one of the researchers, put it: "Google's AI Overviews are built on content that publishers pay to produce, and many of those sites depend on advertising revenue. If people get their answer from the overview and never click through, those publishers lose the visit and the revenue."
What this means for readers
The study's most direct quote:
"Google's summaries generally draw on credible websites, which is good news. But quality sources are not enough. In a meaningful share of cases, we could not find support for a claim in the pages Google cited. Readers still need to check what they're told, even when the citations look trustworthy."
The citation is not a guarantee. It is a pointer. The presence of a blue link next to a claim creates an impression of grounding that the underlying data does not always support.
What this means for publishers
If you produce content that gets cited in AI Overviews, your content is being used to generate answers that may or may not accurately represent what you wrote. You have no control over how your words are synthesized, no way to flag a misrepresentation, and no compensation for the traffic you lose when the overview answers the question for you.
The best defense is the same one the data supports for getting cited in the first place: produce content that is specific, original, and non-commodity. Generic content is easier to misrepresent because it is easier to synthesize. Specific content with original data and explicit claims is harder to twist, and when it is twisted, the contradiction is easier to spot.
The worst case for a publisher is not being uncited. It is being cited incorrectly on a sensitive topic, with no way to correct the record.
Top comments (0)