SAN FRANCISCO, April 14, 2026, 14:04 PDT
Google’s AI Overviews, the machine-written summaries that sit above regular search links, faced fresh scrutiny on Tuesday after follow-up reports revived questions over how often the feature gets basic facts wrong. A New York Times analysis cited in the latest coverage found the tool answered a standard factual benchmark correctly about 91% of the time after Google’s Gemini 3 upgrade, but that rate still implies errors at huge scale.
Why it matters now is simple: Google is pushing AI deeper into the core of Search, not easing off. The company made Gemini 3 the default model for AI Overviews in January, expanded Search Live across AI Mode markets in March and said last July that AI Overviews had more than 2 billion monthly users in over 200 countries and territories and 40 languages.
Google disputed the benchmark. Ned Adriance, a company spokesperson quoted in the coverage, said the study had “serious holes,” while Google’s own help pages warn users that AI Overviews “can and will make mistakes” and advise checking important information in more than one place. Computing
The analysis, conducted with AI startup Oumi, examined 4,326 Google searches using SimpleQA, a test of verifiable factual questions developed by OpenAI. Accuracy improved from about 85% when AI Overviews ran on Gemini 2 to 91% after the move to Gemini 3, but more than half of the accurate answers were still classed as “ungrounded” — industry shorthand for responses whose cited pages do not fully support the claim. Search Engine Roundtable
Some of the mistakes were not obscure. In one example highlighted in follow-up coverage, AI Overviews gave the wrong date for when Bob Marley’s former home became a museum. In another, it said there was no Classical Music Hall of Fame even while citing a site that listed cellist Yo-Yo Ma as an inductee. “Never trust one source,” Pratik Verma, chief executive of Okahu, said. “Always compare what you get with another source.” Computing
The math gets ugly fast. At more than 5 trillion searches a year, a 9% miss rate would imply tens of millions of wrong AI answers an hour if the benchmark mapped cleanly to real-world search traffic. Google has nevertheless kept widening the feature, and said in 2025 that AI Overviews was driving more usage for the types of queries that trigger it in big markets including the United States and India.
The renewed pressure comes as Google faces a broader AI search fight. Reuters has reported that OpenAI and Perplexity are among the challengers trying to pull users toward AI-first search and browser products, as Google folds more generative answers directly into its own results pages.
Publishers are fighting on another front. The European Publishers Council filed an EU antitrust complaint against Google in February over AI Overviews, and Google said in March it was developing controls to let sites opt out of generative AI features as it tries to address UK competition concerns.
Traffic data helps explain why publishers care. Pew Research Center found that users clicked a traditional search result in 8% of visits when an AI summary appeared, versus 15% when no summary was shown, suggesting Google’s answers can keep people on the platform instead of sending them to outside sites.
Still, the downside case is not one-way. Google says the benchmark is flawed and not representative of what users actually ask, and the latest test did show better accuracy after the Gemini 3 shift. Even so, the company tells users to double-check important information, a sign the debate over how much confidence AI search deserves is not going away.
