Google AI Overviews misinfo problem
Multiple analyses suggest Google's AI‑generated 'Overviews' are producing incorrect answers at scale, with one study estimating the feature is 'spewing out tens of millions of inaccurate answers per hour' and others finding roughly 90% accuracy but persistent harmful errors. The shift toward AI summaries is diverting traffic and ad revenue from publishers while making search reliability uneven — a poor fit for legal, financial or reputational queries that demand near‑perfect accuracy. That tension forces businesses to weigh visibility in AI results against the risk of being misrepresented by intermediaries. ( )
Google turned search results into answer boxes in the United States on May 14, 2024, and by May 2025 it said AI Overviews were reaching 1.5 billion monthly users in 200 countries and territories. That means a feature that started as a shortcut now sits in front of a huge share of the world’s web searches. (blog.google, blog.google) The new problem is not that the system is wrong all the time. The problem is that a tool used at Google’s scale can be wrong one time in ten and still flood the internet with bad answers. (nytimes.com, searchengineland.com) A New York Times analysis with artificial intelligence startup Oumi tested 4,326 searches using the SimpleQA factual benchmark and found Google’s AI Overviews answered correctly 85% of the time in October and 91% in February after a Gemini upgrade. That sounds like progress until you combine it with Google’s more than 5 trillion searches per year. (nytimes.com, searchengineland.com) At that volume, a 9% error rate turns into tens of millions of wrong answers per hour and hundreds of thousands per minute. A passing grade in a classroom becomes a factory for mistakes when the classroom is the biggest search engine on earth. (nytimes.com, computing.co.uk) The stranger finding was about sourcing, not just accuracy. Oumi found that more than half of the correct answers in February were “ungrounded,” meaning the links shown under the summary did not fully support what the summary said. (searchengineland.com, nytimes.com) That is like a student turning in the right answer with the wrong homework attached. The grade looks fine until someone checks the footnotes and realizes the proof is missing. (searchengineland.com, nytimes.com) The reported misses were small on paper and dangerous in practice. The Times said Google gave the wrong year for when Bob Marley’s home became a museum, said there was no record of Yo-Yo Ma entering the Classical Music Hall of Fame despite linking to the group’s site, and misstated the date of baseball pitcher Dick Drago’s death while getting his age right. (nytimes.com, searchengineland.com) Google disputed the analysis and said the benchmark was flawed and did not reflect what people actually search for. The company also said its ranking and safety systems are designed to reduce spam and that it has long warned users that artificial intelligence responses can make mistakes. (nytimes.com, searchengineland.com) While Google argues the product is useful, publishers are dealing with a separate hit: fewer clicks. Semrush said AI Overviews sit above organic links, reduce the need to visit websites, and are turning Google from a search engine into what it called an “answer engine.” (semrush.com) News companies say that shift is already showing up in their numbers. TechCrunch, citing Wall Street Journal reporting and Similarweb data, said organic search made up 36.5% of traffic to The New York Times in April 2025, down from 44% three years earlier, as users increasingly got answers without leaving Google. (techcrunch.com) That leaves businesses and publishers in a bad trade. They need to appear inside Google’s summaries because that is where attention is moving, but they also risk being summarized incorrectly by a machine that can sound certain while citing pages that do not actually back it up. (searchengineland.com, semrush.com)
