Millions of Mistakes an Hour: Why You Can’t Take Google’s AI Overviews Seriously

Google rolled out its AI Overviews in search results back in 2024, but the problem of answer reliability still hasn’t been fixed. The investigation found that even after model updates, generation accuracy tops out at 91%. That might sound decent, but given the scale of search traffic, it translates into tens of millions of incorrect answers every single day.

The researchers used the SimpleQA benchmark, developed by OpenAI in 2024, which includes over 4,000 questions with verifiable facts. The previous version of Gemini 2.5 achieved 85% accuracy, but after switching to Gemini 3.1, that number rose to 91%. Even when the AI gets it right, more than half of its links don’t actually back up what it said. Oumi analysts looked at 5,380 links cited by AI Overviews. Facebook and Reddit came in second and fourth place for citation frequency — and when the answer was wrong, social media was mentioned even more often.

A Google spokesperson, Ned Adriance, criticized the study’s methodology, saying the SimpleQA test contains flawed data and doesn’t reflect real search queries — and that Google prefers to use its own “verified” version of the benchmark. Another technical complication is that AI models are non‑deterministic: ask the exact same question a few seconds apart, and you might get a correct answer the first time and a wrong one the next.

The system handles over five trillion queries a year, so even with a 9% error rate, it spits out hundreds of thousands of incorrect answers every minute. Google is gradually improving accuracy, but at the same time, the share of answers that can’t be fact‑checked is growing.

The real problem isn’t the error rate itself — it’s how the AI “packages” those mistakes. The answer sounds authoritative, with no hint of doubt. You won’t see words like “maybe,” “according to some sources,” or “I’m not sure.” The AI never hesitates, so even a 9% error rate turns into a massive flood of false information. Sure, Google warns you to double‑check the answers — but they do it in tiny fine print, after you’ve already been served the “truth.”

Have you ever run into Google giving blatantly wrong information in its AI Overviews? Or do you always fact‑check those answers using other sources?