A study posted to arXiv on October 30, 2026, found that Google’s AI Overviews for health questions often drew on different sources from the regular search results below them. The researchers compared Overviews with the top five traditional results for 3,611 benchmark questions, checking source quality, whether cited pages supported the answers, and how answers scored against physician-written rubrics.
An Overview appeared for 96.4% of the questions. But it cited only 44.4% of the top-five results for the same question. The study says Overviews relied more on commercial health publishers such as WebMD and Healthline, and less on academic and government sources, including the Cleveland Clinic and the UK’s National Health Service.
How well did the answers hold up?
The study assessed whether statements in an Overview were supported by their citations. After adjusting for errors in the tool used to classify support, the researchers estimated that 56.7% of cited statements were supported by the cited page. A further 20.8% of all statements made factual claims without any citation. A missing citation does not, by itself, show that a claim is false; it does mean a reader has no cited page to check for that claim.
Against physician-written rubrics, Overviews earned 38.2% of available points overall. They scored 54.2% on accuracy and 21.6% on completeness. In the subset of questions where a rubric anticipated a specific inaccurate or potentially harmful claim, 22.2% of the Overviews included that content. The abstract does not give the size of that subset, so the figure should not be read as the share of all health Overviews that contained harmful content.
What the study can—and can’t—tell us
These results describe performance on a benchmark set of questions, not every health search a person might make. They also do not establish that a particular source choice caused an answer to be wrong: the study found no relationship between citation grounding and rubric accuracy. That distinction matters because a cited statement can be supported by a page without being complete or accurate, while an uncited statement may still be correct. The authors recommend teaching users how to assess AI Overviews and holding them to at least Google’s own standards for ranking health information.
Why it matters
For people using Google to research a health concern, an AI Overview may cite different pages from the regular results and may make claims that its citations do not support. The study measured benchmark questions, not every search or a real-world health outcome, so it cannot tell users how reliable a particular answer is; checking the cited page and other medical sources remains important.
