SAN FRANCISCO, April 14, 2026, 14:04 PDT
Google’s AI Overviews are under the microscope again. On Tuesday, new coverage raised concerns about the accuracy of these machine-generated summaries, which appear above traditional search results. The New York Times, referenced in the latest reports, found that after the Gemini 3 update, the tool delivered the right answer to a standard factual test about 91% of the time. That still leaves a significant number of mistakes, given the scale at which the feature operates.
Here’s why this is in sharp focus: Google keeps accelerating AI’s role in Search. They set Gemini 3 as the new default for AI Overviews back in January, rolled out Search Live to more markets using AI Mode in March, and last July reported AI Overviews had topped 2 billion monthly users across 200+ countries and territories, spanning 40 languages.
Google pushed back on the findings. Company spokesperson Ned Adriance, cited in the report, called out “serious holes” in the study. Google’s support pages themselves caution that AI Overviews “can and will make mistakes,” urging users to double-check key details elsewhere. Computing
AI startup Oumi worked with Google on an analysis that looked at 4,326 searches using SimpleQA — OpenAI’s benchmark for fact-based queries. After switching AI Overviews from Gemini 2 to Gemini 3, accuracy edged up from roughly 85% to 91%. Still, over half of the correct responses were tagged as “ungrounded,” industry jargon for answers where the cited pages don’t really back up the claim. Search Engine Roundtable
Some of the errors weren’t exactly subtle. At one point, AI Overviews provided the incorrect year for when Bob Marley’s old house became a museum—a detail later pointed out in media coverage. Elsewhere, the system claimed there’s no Classical Music Hall of Fame, despite referencing a source that actually listed Yo-Yo Ma as an honoree. “Never trust one source,” said Pratik Verma, CEO of Okahu. “Always compare what you get with another source.” Computing
The numbers escalate quickly. With over 5 trillion searches annually, a 9% error rate would mean AI puts out tens of millions of bad answers every hour—if that benchmark actually holds up against real-world search volumes. Still, Google has expanded the feature, saying in 2025 that AI Overviews is boosting usage for the kinds of queries it powers in major markets like the United States and India.
Google is feeling the heat again, this time in the wider battle over AI search. Reuters notes that both OpenAI and Perplexity are angling to attract users to their AI-centric search and browsing tools, while Google keeps adding generative responses right into its results.
Publishers have a new battleground. In February, the European Publishers Council took its concerns to regulators, lodging an EU antitrust complaint against Google over AI Overviews. Then in March, Google responded, saying it was working on controls that would allow publishers to opt out of its generative AI tools—a move aimed at easing UK competition worries.
Traffic figures give publishers good reason to worry. When an AI summary shows up, users only clicked traditional search results 8% of the time, according to Pew Research Center data—down from 15% of visits when there’s no summary. The takeaway: Google’s AI-generated answers appear to keep more users on its own site rather than sending them off to other publishers.
But it’s not all negative. Google argues the benchmark doesn’t reflect real user queries and points to improved accuracy after switching to Gemini 3. Still, the company continues to advise users to verify key details—evidence that questions around trust in AI search aren’t fading any time soon.
