# Study: Google AI Overviews cite sources that often don't support their health claims

Published: 2026-10-30T06:01:23.061Z · Source: arXiv (Agnoli, Reilly, Barna, Horvát) (https://arxiv.org/abs/2610.39339v1)
Source date: 2026-10-29
Flags: arXiv preprint — not peer reviewed
Entities: ai-overviews, google

A preprint posted to arXiv on October 29 compared Google's AI Overviews against the top five traditional results for 3,611 benchmark health questions. Overviews appeared for 96.4% of questions, but only 44.4% of the top five results were cited by the matching overview, which leaned toward commercial health publishers (WebMD and Healthline are the authors' examples) over academic and government sources such as Cleveland Clinic and the NHS. After correcting for classifier error, the authors estimate 56.7% of cited statements were supported by the page they cite, and a further 20.8% of all statements were factual claims with no citation.

Graded against physician-written rubrics, the overviews earned 38.2% of available points (accuracy 54.2%, completeness 21.6%); among questions whose rubric anticipated a specific inaccurate or potentially harmful claim, 22.2% of overviews contained it. The authors report that an overview's accuracy was unrelated to how well its statements were grounded in its sources. This is a first version of the paper and has not been peer reviewed.

Why it matters: Health is the query type where Google says it holds results to its strictest quality standards, so an outside measurement that the cited source often does not support the AI Overview's statement bears directly on what a citation proves.

## What this answers

**How accurate are Google AI Overviews for health questions?**

A preprint that graded AI Overviews against physician-written rubrics on 3,611 health questions found they earned 38.2% of available points, including 54.2% on accuracy and 21.6% on completeness.

**Do Google AI Overviews cite the same sources as regular search results for health queries?**

No. Only 44.4% of the top five traditional results were cited by the matching AI Overview, which leaned on commercial health publishers such as WebMD and Healthline over academic and government sources, per the authors.


Canonical: https://anythingengineoptimization.com/item/2026-10-30-study-google-ai-overviews-cite-sources-that-often-don-t-support-their/
From Anything Engine Optimization (AEO Wire) — https://anythingengineoptimization.com/ · Standards: https://anythingengineoptimization.com/standards/