AI Overviews: 11% of Claims Unsupported and 5 Essential Fixes

Washington University researchers checked 98,020 claims in Google AI Overviews and found about 11% unsupported or contradicted by their own citations. The findings and five fixes for your content.

This article summarises a peer-reviewed measurement study from Washington University in St. Louis and related reporting. I did not run the study, and figures are attributed to the paper, the university’s release and coverage of it.

Google’s AI Overviews cite good sources and still get the details wrong about one time in nine. Researchers at Washington University in St. Louis ran 55,393 trending searches over 40 days, broke the resulting summaries into 98,020 individual claims, and found that about 11% were either not supported by the pages Google cited or directly contradicted them.

That number cuts two ways. For readers, it means the citations under AI Overviews are not proof that the summary is right. For publishers and SEOs, it points to something more practical: the most common failure was omission, the summary stating a fact the source never said. That is partly a content problem you can fix. Here is what the study found and five things to change on your site.

Table of Contents

AI Overviews accuracy checked with a magnifying glass
Good citations are not proof. In the study, the quality of a cited source did not predict whether the AI Overviews claim was right.

The Study Behind the 11% Figure

The research comes from Haofei Xu, Umar Iqbal and Jacob M. Montgomery at Washington University in St. Louis. Their paper, posted as a preprint in May 2026 and publicised by the university in late September, was accepted for the ACM Internet Measurement Conference in October 2026.

The method was simple in concept and large in scale. Over 40 days, from 13 March to 21 April 2026, the team issued 55,393 trending queries across 19 topic categories using a distributed automated browser. For each search they captured the AI Overview, its cited sources and the regular results shown on the same page.

They then broke every overview into atomic claims, 98,020 in total, and checked each claim against the 61,212 citations Google provided. That claim-by-claim approach is what makes the study useful. It measures not whether an overview “looks right” but whether each statement is actually backed by the source next to it.

How Often AI Overviews Appear

AI Overviews appeared for 13.7% of all queries in the sample. For queries phrased as questions, that jumped to 64.7%, according to MarketingProfs’ summary of the findings.

That gap is the most useful number in the study for content strategy. If your audience asks questions, and most informational audiences do, roughly two thirds of their searches now return an AI summary above the regular results. Short keyword-style searches mostly do not.

The sample was US trending queries, so it skews towards news, entertainment and current events. Commercial and niche B2B searches may behave differently, as one agency analysis of the paper pointed out.

How Accurate AI Overviews Are

About 89% of claims were consistent with their sources, either clearly stated or slightly vague but accurate. The remaining 11% broke down into two kinds of failure.

Claim type Share of 98,020 claims What it means
Consistent with source About 89% Clearly or broadly supported by the cited page
Omitted from source About 7% The claim was not found in the cited text the team could collect
Contradicted by source About 4% The cited page said something different

A stricter cut is less flattering. Only 41.9% of claims were fully grounded in the referenced text, according to a summary of the paper by HyperAI. The rest of the “consistent” group was broadly accurate but not stated explicitly by the source.

The authors are careful to call 11% an upper bound. Their pipeline could not fully capture user-generated content such as Reddit threads, forum posts and YouTube videos, so some claims marked unsupported may in fact appear in those sources. Even so, at Google’s scale, a single-digit error rate still means a very large number of wrong statements every day.

Good Sources, Shaky Claims

The finding that surprised me most: the domains AI Overviews cite often scored higher on credibility than the regular search results shown on the same page. Google is, broadly, picking reputable sources.

But the paper found that source quality and claim fidelity are largely independent. A trustworthy citation does not make the summary’s claim more likely to be correct. The university’s release put it plainly: readers still need to check what they are told, even when the citations look trustworthy.

For publishers, that is a double-edged result. Being cited is good for visibility. It does not protect you from being misquoted, and a misquote under your logo can do more harm than not appearing at all.

Which Topics AI Overviews Get Wrong Most

Accuracy varied by subject. PsyPost’s coverage reports that health and political summaries were more accurate than those about entertainment or technology.

That fits what you would expect if Google applies extra caution to sensitive categories. It also means technology publishers, including sites like this one, are in the less accurate group. If you write about products, software or AI, assume your pages are more likely to be summarised loosely, and write accordingly.

How This Compares With Other AI Overviews Research

This is not the only study. Earlier in 2026, the New York Times worked with AI company Oumi to test AI Overviews against OpenAI’s SimpleQA benchmark. That work found overviews wrong in about one search in ten, and at the individual-claim level only about 67% of claims were supported by their sources, according to later analyses. Google said that study was not representative of what people actually search for.

The two studies used different methods and different queries, so their numbers should not be substituted for each other. What they share is the direction: AI Overviews usually get the headline answer right and often add details the sources do not support.

That pattern, a correct main answer with an unsupported side claim, is the one to design your content around. It is also the classic weakness of any retrieval-based system, which we explain in our guide to retrieval-augmented generation.

Why Summaries Add Facts the Source Never Said

It helps to understand why omission is the most common failure, because the fixes follow from it.

A system like this works in two steps. First it retrieves a handful of pages that look relevant. Then a language model writes a short answer using those pages as context. The model is good at producing fluent, complete-sounding text, and it has absorbed a huge amount of general knowledge during training. When the retrieved pages leave a gap, the model often fills it from that background knowledge rather than leaving it blank.

Most of the time the filler is right, which is why the overall accuracy is high. But the citation is attached to the paragraph, not to each fact, so a reader cannot tell which sentence came from the source and which came from the model’s memory. That is exactly the kind of claim the researchers flagged as “not found in the cited text”.

Contradictions are a slightly different problem. They tend to appear when sources disagree with each other, when a page is out of date, or when the model compresses a nuanced statement into something simpler and loses a qualifier along the way. “Usually” becomes “always”. “Up to” becomes a fixed number.

Both failure types share a cause that publishers can influence: ambiguity on the page. The clearer and more complete your own statement of a fact, the less room the model has to improvise. That is the thread running through all five fixes below.

There is also a limit to what any publisher can do. If the model draws on its background knowledge regardless, even a perfectly written page will sometimes be paraphrased loosely. The aim is not perfection. It is to make the accurate version the easiest one to write.

Fix 1: State Facts Plainly in Text

Omission was the dominant failure mode: the overview stated something the cited page did not say. Some of that is the model filling gaps. You can reduce the gaps.

Write key facts as complete, self-contained sentences. “The plan costs $20 a month and includes 50 GB of storage” is far harder to misread than a pricing table with icons, a footnote and a “see details” link. If a fact matters, say it in words, in one sentence, on the page.

Fix 2: Answer the Question Near the Top

With question-style queries triggering AI Overviews about two thirds of the time, your page will often be summarised in response to a specific question. Put the direct answer high on the page, then expand.

This is the same advice that works for featured snippets and for AI assistants generally, and it is covered in depth in our generative engine optimization guide. The difference now is that you have hard data on how often it matters.

Fix 3: Keep Facts Out of Images and Video Alone

The study’s pipeline struggled with content it could not read as text. Google’s systems have the same limit in many cases. Specifications in an image, prices in a video or comparisons inside an embedded widget are invisible or hard to verify.

Keep the images, but repeat the important facts in plain text nearby. Add descriptive alt text. It helps accessibility, it helps search, and it gives any summarising system something concrete to quote.

Fix 4: Date and Update Your Numbers

Contradicted claims often come from stale or conflicting information. If an old page says one price and a new page says another, an overview may blend them.

Put a visible “last updated” date on pages with facts that change, and update or redirect old pages rather than leaving outdated versions online. When a number changes, change it everywhere on your site, not just on the main page.

Fix 5: Monitor What AI Overviews Say About You

You cannot fix a misquote you do not know about. Make a short list of the questions your customers ask most, search them regularly, and record what the AI Overview says and which pages it cites.

If it gets your facts wrong, check whether your own page states them clearly. Often the fix is on your side. Google also offers feedback options on AI Overviews, which are worth using for clear errors. Tools that help with on-page basics, such as those in our roundup of AI tools for SEO meta titles, can speed up the clean-up.

What AI Overviews Accuracy Means for Readers

Most of this article is aimed at people who publish content. But the study has a simpler message for everyone who searches.

Treat the summary as a map, not the territory. It is usually pointing in the right direction, and it often saves you a click. When a detail matters, such as a dose, a deadline, a price or a legal requirement, open the cited page and check that it actually says what the summary claims. Eleven claims in a hundred is a small share until one of them is the one you act on.

Notice the topic too. The study found technology and entertainment summaries less accurate than health and politics. If you are researching software, gadgets or AI tools, be a little more sceptical than usual.

And remember the summary is built from what it could read. If the best answer to your question lives in a forum thread or a video, the overview may not reflect it at all. Scrolling past the summary to the regular results is still a perfectly good way to search.

Who Should Act on AI Overviews Accuracy

If you are Start with Why Link
A publisher of how-to or product content Rewrite key facts as plain sentences Omission was the most common failure Read the paper
An SEO lead Track question queries in your niche They trigger AI Overviews about 65% of the time GEO guide
A brand or product manager Monitor what AI Overviews say about you monthly Good citations do not guarantee accurate claims WashU summary
A reader Click through before relying on a detail About one claim in nine may not match its source PsyPost explainer

Related on this site: for more context, see Siri AI Two Months In.

FAQ About AI Overviews

How accurate are Google AI Overviews?
In Washington University’s study of 98,020 claims, about 89% were consistent with their sources and about 11% were unsupported or contradicted. The authors treat 11% as an upper bound.

How often do AI Overviews appear in search?
In the study’s sample of trending US queries, AI Overviews appeared for 13.7% of all searches and 64.7% of searches phrased as questions.

Do AI Overviews cite reliable sources?
Generally, yes. Cited domains often had higher credibility scores than the regular results on the same page. But source quality did not predict whether the claim itself was accurate.

What is the most common AI Overviews error?
Omission: the summary states a fact that does not appear in the cited page. About 7% of claims fell into that group, against about 4% that contradicted the source.

Which topics have the most errors?
Coverage of the study reports higher accuracy for health and political topics and lower accuracy for entertainment and technology.

Can publishers reduce errors about their content?
Partly. Stating facts plainly in text, answering questions early, avoiding facts locked in images and keeping numbers current all make your pages easier to summarise correctly.

Verdict

AI Overviews are mostly right and sometimes confidently wrong, and the citations underneath do not tell you which. With about 11% of claims unsupported or contradicted in the largest independent audit so far, readers should treat them as a starting point, and publishers should assume their pages will occasionally be misquoted.

The caveat is that the study covered US trending queries over 40 days in early 2026, and the authors call the error rate an upper bound. Google’s systems change constantly, and your niche may perform better or worse than the average.

Your next step: pick the ten questions your customers search most, check today what AI Overviews say for each, and rewrite any page where your key fact is not stated plainly in a single sentence.

Sources and Further Reading

On this site, read our generative engine optimization guide for the wider strategy, retrieval-augmented generation to understand why summaries drift from sources, and how to write a WordPress post with AI for content that summarises cleanly.