Study Finds 11% of Claims in Google AI Overviews Lack Support from Cited Sources - General science

An analysis of more than 55,000 Google searches found that AI Overviews frequently cite credible websites but still include unsupported information in roughly one out of every nine factual claims. The study also found that the feature draws on sources outside Google’s traditional first-page rankings, potentially reducing traffic and advertising revenue for publishers.

Google’s AI Overviews often rely on credible websites, but a new study found that 11 percent of the factual claims in its generated summaries were not supported by the cited pages. The research, published as a preprint on arXiv, examined how the feature selects sources, summarizes information and affects the websites supplying that material.

AI Overviews appear at the top of some Google results, where they combine information from multiple web pages into a single answer. Researchers from Washington University in St. Louis evaluated the system over 40 days in early 2026 by running 55,393 trending queries across 19 subject areas.

The feature appeared for 13.7 percent of all searches in the study. It was much more common when users entered direct questions, activating for 64.7 percent of those queries, compared with 9.5 percent of conventional keyword searches. Activation also differed by topic: AI Overviews appeared more often for hobbies and science searches and less often for politically sensitive subjects.

The researchers found that the sources cited by AI Overviews generally had higher credibility scores than the websites shown in the standard first-page results. The system also relied less on unvetted user-generated content, including social media and video platforms. However, nearly 30 percent of its cited websites did not appear on the first page of regular results, indicating that it uses a different process to select information.

To assess accuracy, the team divided the summaries into 98,020 factual claims and used an automated verification system to compare them with the cited pages. About 11 percent of the claims lacked support. Most involved information that was absent from the linked source, while a smaller share directly conflicted with it.

Using more credible sources did not eliminate the problem. The study found no statistical relationship between a website’s credibility and whether its content was accurately summarized. Accuracy also varied across subjects, with health and political summaries performing better than those covering entertainment and technology.

The analysis also examined the possible effect on publishers. More than half of the pages cited in AI Overviews contained display advertising. Because users can receive an answer without visiting the original site, publishers may lose traffic and the advertising revenue associated with those visits. Google, meanwhile, continues to show sponsored advertisements alongside the generated summaries, and in some cases places ads above them.

The researchers identified limitations in their automated approach. Live information, such as weather data, could change between the creation of a summary and the researchers’ review of its cited page, producing false accuracy failures. The team also excluded sites that required authentication or blocked automated access, meaning some claims could have been supported by material the researchers were unable to examine.

The study’s authors—Haofei Xu, Umar Iqbal and Jacob M. Montgomery—said further research is needed to determine how users respond to AI Overviews, whether they notice citation errors and how the summaries influence beliefs about important subjects. The paper is titled “Measuring Google AI Overviews: Activation, Source Quality, Claim Fidelity, and Publisher Impact.”

Sources: PsyPost.