Google’s AI Overviews, those tidy summaries perched atop search results, sound helpful. Right? They promise quick answers drawn from the web. But a fresh analysis paints a grim picture: hundreds of millions of wrong responses daily, scaling to 225 billion misleading summaries a year. And that’s assuming the feature triggers on just half of Google’s 5 trillion annual searches. TechRepublic broke the story on April 21, 2026, drawing from an Oumi study commissioned by The New York Times.
Oumi tested 4,326 searches using the SimpleQA benchmark, an industry yardstick from OpenAI. Results? AI Overviews hit 85% accuracy with the older Gemini 2 model. Bump to Gemini 3, now powering the feature, and it climbs to 91%. Solid numbers, perhaps. Scale them up, though. A 9% error rate means tens of millions of bad answers hourly. Google processes billions of queries daily; even small slips amplify massively.
Worse still. Even correct answers often lack solid footing. Fully 56% of accurate responses were ‘ungrounded’—citing sources that don’t back the claims. With Gemini 3, that figure drops to 37% for accurate ones, but the overall problem persists. AI Overviews pulled from 5,380 sources total. Facebook ranked second most-cited. Reddit took fourth. Facebook appeared in 5% of accurate summaries, 7% of inaccurate ones. User-generated content steering the ship. Questionable at best.
Google pushes back hard. Spokesperson Ned Adriance called the study flawed. ‘This study has serious holes,’ he said. ‘It doesn’t reflect what people are actually searching on Google.’ Fair point—benchmarks aren’t real-world queries. Yet Google’s internal tests show Gemini 3 alone errs on 28% of queries. Pair it with search, and accuracy rises. Still, the feature sits there, bold and prominent, luring clicks away from deeper dives.
Users buy it hook, line, and sinker. A Pew Research Center survey from July 2025 found AI Overview viewers click traditional results just 8% of the time. No summary? That jumps to 15%. Links in the overview itself? A measly 1%. By mid-2025, the feature reached 2 billion monthly users across 200 jurisdictions and 40 languages. Trust erodes verification.
Health queries expose the stakes. The Guardian reported in January 2026 on AI Overviews dispensing dangerous medical misinformation. Pancreatic cancer patients got advice to shun high-fat foods—the opposite of what’s needed to sustain strength for chemo. Liver test results misled users into thinking severe issues were fine. Vaginal cancer symptoms wrongly tied to Pap tests. Repeat the search. Different wrong answer each time. Mental health tips for psychosis and eating disorders? ‘Very dangerous,’ per experts, potentially deterring real help.
This isn’t new. Back in May 2024, launch glitches went viral: glue on pizza, anyone? Wired noted in May 2025 the feature still hallucinated dates, claiming 2024 lingered into 2025. Mashable tested in August 2025: one in five responses misleading, even on real questions. Anecdotal, sure. But patterns emerge.
Finance fares no better. A study found 37% of AI Overviews on banking, credit, investing, insurance, loans, and aid flat-out wrong. X users echo the frustration. One post from April 21, 2026, flagged the Oumi findings: ‘More confident, less accurate. Classic.’ Another: ‘Users trust AI blindly, that’s why misinformation scales faster than truth.’
Traffic shifts too. An arXiv paper pegged AI Overviews cutting Wikipedia English traffic 15%. SEO Clarity’s October 2025 research showed organic rankings volatile post-rollout. Publishers suffer as summaries siphon clicks.
Google admits models err. They’ve dialed back triggers post-launch debacles. Yet expansion continues—now universal for non-logged-in users since August 2024. Disclaimers linger at the bottom: AI may not be accurate. Too late for many. People scan tops first.
Fixes? Grounding citations tighter. Better benchmarks mirroring daily searches. Human oversight for high-risk topics like health. But scale defies perfection. 91% accuracy thrills in labs. In the wild, with billions querying health scares or investments? Unforgivable.
So next search. Pause. Click through. Question the box. Misinformation at this volume doesn’t just mislead. It reshapes knowledge itself.
WebProNews is an iEntry Publication