Health: Out of 3,611 questions, Google's AI Overviews score only 38% of the points in a grid written by doctors

A study shows that Google's AI summaries favor commercial publishers, and that nearly one in two quoted claims is not supported by its source.

AI Overviews, the summaries generated by Gemini that top Google's results, appear for 96.4% of the 3,611 health questions tested by a team of four researchers. But they don't draw from the same sources as traditional search: only 44.4% of the traditional top five results are cited. Summaries overrepresent commercial publishers like WebMD or Healthline, to the detriment of academic or public sources like the Cleveland Clinic or the NHS, and cited pages score lower on four of the usual six credibility criteria.

Quotes that prove nothing

After correcting for classifier errors, about 56.7% of statements accompanied by a quote are actually supported by the page cited, and 20.8% of factual statements have no source. Scored on grids written by doctors, summaries receive 38.2% of the points (54.2% in accuracy, 21.6% in completeness). When the grid anticipated a specific false or dangerous statement, 22.2% of summaries contained it.

Notably, the accuracy of the summaries has no link with the quality of their citations: a reference displayed does not guarantee anything. The authors argue for training the public to read these summaries and ask that Google at least apply its own criteria for ranking health content to them.