Google's AI Overviews get their facts wrong more often than you'd hope - and it's happening at a scale few online products ever reach.
Researchers tracked Google AI Overviews, the AI-generated summaries that now appear in search results reaching more than 2 billion users, many of whom never realize the text is AI-generated rather than human-written. Over 40 days in March and April 2026, the team ran 55,393 trending queries across 19 topic categories to measure how often the feature shows up and how accurate it is. AI Overviews appeared for 13.7% of queries overall, jumping to 64.7% for question-style searches, while politically sensitive topics triggered it far less often. Breaking the answers into 98,020 individual factual claims, the researchers found cited sources were generally more credible than the links on the regular results page - but nearly 30% of those sources didn't appear in first-page results at all, and 11% of claims weren't actually backed up by what they cited.
That's a real editorial shift. Google isn't just ranking pages and letting readers pick anymore - it's picking the answer for them, using a source-selection method separate from its own search ranking. And since more than half of the cited pages carry display ads, publishers can lose referral traffic and ad revenue even as Google's own sponsored ads keep running right next to the summary.
Search used to be a librarian handing you a stack of books. Now, for over 2 billion users, it's one confidently-worded answer from Google - and roughly one in nine of its claims doesn't check out.