New Study Points to Holes In AI Overviews

Why It Matters

Inaccurate or ungrounded AI answers erode trust in search, affecting user experience and advertiser credibility across the digital ecosystem.

Key Takeaways

• Study examined 4,326 Google AI Overviews for accuracy.
• Accuracy rose from 85% with Gemini 2 to 91% with Gemini 3.
• Over half of the 'accurate' answers were ungrounded.
• Ungrounded responses increased under Gemini 3 versus Gemini 2.
• Google acknowledged study’s flaws, citing 'serious holes'.

Pulse Analysis

Google’s AI Overviews have become a cornerstone of the company’s search experience, delivering concise answers generated by large‑language models. A new independent analysis by research firm Oumi, cited by The New York Times, audited 4,326 of these snippets to gauge factual reliability. While the headline figure—91 % accuracy with the latest Gemini 3 model—appears reassuring, the study highlights a deeper problem: more than half of the supposedly correct answers lack proper source grounding. This disconnect raises questions about how users verify information presented in real‑time search results.

The audit revealed that 85 % of Gemini 2‑powered Overviews were accurate, climbing to 91 % after the upgrade to Gemini 3. However, the proportion of ungrounded responses rose alongside the accuracy gain, with Gemini 3 generating more answers that could not be traced to verifiable web sources. Ungrounded content erodes trust because readers cannot cross‑check claims, especially when the answers are embedded directly in the SERP. For enterprises that rely on Google’s AI for brand visibility, even a 9 % error rate can translate into millions of misinformed impressions daily.

Google’s brief response—labeling the study’s findings as having ‘serious holes’—underscores the competitive pressure to showcase AI progress while managing quality control. Industry observers argue that transparent benchmarking and stricter grounding requirements will become essential as AI‑generated snippets dominate information discovery. Companies developing competing LLM‑driven search tools are likely to capitalize on any perceived weakness in Google’s answer engine, positioning their models as more reliable. For users and advertisers alike, the takeaway is clear: accuracy metrics alone are insufficient without robust source attribution.