Google's AI Overviews are correct nine out of ten times, study finds

Study tested Google AI overviews with the SimpleQA benchmark on 4,326 Google searches.

In the October round Gemini 2 answers were 85% correct; in February Gemini 3 answers were 91% correct.

The study focused on accuracy and verifiability; Google displays a disclaimer that AI responses may include mistakes under AI-generated search results.

Verifiability declined: the share of correct but ungrounded answers rose from 37% (Gemini 2) to 56% (Gemini 3).

Out of 5,380 sources cited, Facebook and Reddit were among the most common; Facebook appeared in 5% of correct and 7% of incorrect answers. Examples of errors included a wrong Yo-Yo Ma induction record, a misread river name, and an incorrect opening year for the Bob Marley Museum.