Google’s AI Overview wrong 10% of the times

A major report published yesterday by The New York Times, in collaboration with AI startup Oumi, has confirmed that Google’s AI Overviews currently carry an error rate of approximately 9% to 10%.

The errors matter at scale, as AI responses now appear in 43% of Google searches.

The study used the SimpleQA benchmark—a rigorous test of over 4,000 verifiable factual questions—to measure the accuracy of the summaries that now appear at the top of nearly 50% of all Google searches.

---

1. The Accuracy Gap: Gemini 2.5 vs. Gemini 3

The report highlights that while Google’s accuracy is improving with newer models, the “confidently wrong” nature of AI still poses a massive scale problem.

Model Version	Accuracy Rate	Error Rate	Status
Gemini 2.5	85%	15%	2025 Standard
Gemini 3	91%	9%	Current (April 2026)

While a 91% accuracy rate sounds impressive, analysts warn that with Google handling over 5 trillion searches per year, even a 9% error rate results in tens of millions of incorrect answers every single day.

---

2. High-Profile “Hallucinations”

The NYT and Oumi analysis identified several specific instances where AI Overviews provided false information as absolute fact:

• Bob Marley Museum: When asked when Marley’s home became a museum, Google answered 1987. The correct year is 1986. The AI cited multiple sources, but none supported the 1987 claim.
• Yo-Yo Ma: For a query about his induction into the “Classical Music Hall of Fame,” Google linked to the organization’s site but claimed the hall of fame did not exist.
• Dick Drago: The AI correctly stated the baseball player’s age at death but provided a completely fabricated date of death.

---

3. The “Ungrounded” Citation Problem

Perhaps more troubling than the flat-out lies is the rise of “ungrounded” responses.

• Definition: These are answers that are factually correct but link to sources that do not contain the information or actually contradict the answer.
• The Trend: In October 2025, 37% of correct AI answers were ungrounded. By February 2026, that figure rose to 56%. This makes it nearly impossible for users to verify the information without doing a manual search.

---

4. Google’s Rebuttal: “Flawed Benchmarks”

Google spokesperson Ned Adriance has pushed back against the findings, stating that the study has “serious holes” and does not reflect how people actually use Search.

• Flawed Data: Google claims the SimpleQA benchmark itself contains inaccurate data and that they use a more strictly validated internal tool called “SimpleQA Verified.”
• Dynamic Results: Google notes that AI Overviews use different models (some smaller/faster, some larger/smarter) depending on the query, making a single accuracy score misleading.
• Safety First: The company maintains that its safety systems filter out the most dangerous hallucinations, particularly in “Your Money or Your Life” (YMYL) categories like health and finance.

---

5. Why It Matters: The “Zero-Click” Risk

Critics argue that the 10% error rate is more dangerous now than in previous years because AI Overviews have led to a “Zero-Click” search culture.

• Trust Factor: Unlike the old “ten blue links” where users compared sources, AI Overviews present a single, authoritative summary.
• Organic Drop: Organic click-through rates (CTR) have plummeted by 61% for queries featuring an AI Overview. If the summary is wrong, users are 60% less likely to click a link and discover the truth.

“Google is trading the reliability of the open web for the convenience of a summary,” noted an analyst from Ars Technica. “When 1 in 10 answers is wrong, ‘convenience’ becomes a liability.”