Are Google AI Overviews accurate?
Google AI Overviews answered 91% of SimpleQA questions correctly in the New York Times and Oumi analysis, but most answers were not supported by the sources they cited.
Short answer
Google AI Overviews are usually right on simple factual questions, but their sources often do not back them up. In the New York Times and Oumi analysis reported by Futurism, AI Overviews answered 91% of SimpleQA questions correctly. Yet 56% of answers were ungrounded: the cited sources did not support the claim.
Results
- News audit
- Vendor claim
| Test | What was measured | Result | Date | Source |
|---|---|---|---|---|
| NYT and Oumi analysis of AI Overviews | Accuracy on SimpleQA questions, Gemini 3 era (4,326 searches per round) | 91% | Source 1News audit | |
| NYT and Oumi analysis of AI Overviews | Ungrounded answers: the cited sources did not support the claim | 56% | Source 1News audit | |
| Google internal testingVendor's own figure, not independently checked | Accuracy of AI Overviews against featured snippets | On par | Source 2Vendor claim | |
| Google internal reviewVendor's own figure, not independently checked | Content policy violations in queries where AI Overviews appeared | Less than 1 in every 7 million unique queries | Source 2Vendor claim |
What these results mean
The New York Times worked with Oumi to run SimpleQA questions through Google Search and score the AI Overviews that appeared. Under the Gemini 3 model, 91% of answers were correct, up from 85% under the earlier Gemini model (Futurism, 2026-04-08). Each round used 4,326 searches.
Being right is not the same as being supported. The same analysis found 56% of answers were ungrounded, meaning the sources AI Overviews linked to did not support what it said. Under the earlier Gemini model that figure was 37%. So the newer model got more answers right while showing its working less often. Other outlets reported the ungrounded figure as 51%. The primary New York Times article is paywalled, and the figures here come from Futurism's write-up.
Google disputes the study and said it has serious holes. Google's published figures are a comparison with featured snippets and a count of policy violations, not an error rate for factual answers. Google's own published position, from May 2024, is that AI Overviews are as accurate as featured snippets in its internal tests. It also reported content policy violations on less than 1 in every 7 million unique queries where AI Overviews appeared (Google, 2024-05-30). That figure covers harmful content, not factual errors. Both are the vendor's claims.
A correct AI Overview with a source that says something else is still an answer you cannot verify from the page.
Where AI Overviews go wrong
Sources that do not back the claim
56% of answers were ungrounded in the NYT and Oumi analysis, as reported by Futurism. A link under the summary is not proof.
Viral wrong answers
Google's May 2024 post followed viral answers that suggested glue on pizza and eating rocks. Google said it made more than a dozen technical fixes afterwards (Google, 2024-05-30).
Grounding got worse as accuracy improved
The share of ungrounded answers rose from 37% to 56% when Google moved to the Gemini 3 model.
Errors at search scale
91% accuracy on simple questions sounds high, but AI Overviews sit at the top of a huge number of searches. A small share of wrong answers still reaches many people, and the overview is the first thing they read.
AI Overviews run on Gemini. The Gemini assistant has its own record, covered on is Gemini accurate.
How it compares
The NYT and Oumi analysis tested only AI Overviews, so the fair comparison is within that test, between Google's earlier model and the Gemini 3 model.
NYT and Oumi analysis of AI Overviews (reported 2026-04-08)
| Measure | Earlier Gemini model | Gemini 3 model |
|---|---|---|
| Accuracy on SimpleQA questions | 85% | 91% |
| Ungrounded answers | 37% | 56% |
Source: Futurism, reporting on the New York Times and Oumi analysis. Accuracy went up and grounding went down, in the same test and with the same questions.
For assistants that answer in a chat window rather than a search page, see Perplexity, which had the best sourcing record in two news studies, and Microsoft Copilot. The overall order is on the most accurate AI assistants.
How to check an AI Overview
Open the cited link and find the exact claim
If the page does not say it, do not rely on the overview.
Scroll past the overview to the regular results
Compare the summary with two or three reputable pages.
Watch for humour, forums and satire
If the source is a joke post or a forum thread, the answer may be one too.
Check health, money and legal answers at the source
Go to an official body, a doctor's guidance or the law itself. These are the answers where a wrong summary costs the most.
Check dates
An overview can mix old and new facts. Confirm the date on the source page.
The full routine is on how to fact-check an AI answer. How we weigh news reports against vendor claims is on how we score AI accuracy.
Sources
- News audit
- Vendor claim
Questions
Are Google AI Overviews always right?
No. In the New York Times and Oumi analysis reported by Futurism in April 2026, AI Overviews answered 91% of SimpleQA questions correctly, so some answers were still wrong. More of a problem, 56% of answers were not supported by the sources they cited. Open the source before you trust the summary.
Are AI Overviews reliable for research?
No, not on their own. An AI Overview can be a quick pointer, but the NYT and Oumi analysis found 56% of answers were ungrounded, meaning the linked sources did not support the claim. For research, use the overview only to find terms and leads, then cite pages you have opened and read.
Do AI Overviews make up sources?
AI Overviews link to real pages, but the pages often do not say what the overview says. In the NYT and Oumi analysis, as reported by Futurism, 56% of answers were ungrounded. Other outlets reported 51%. Either way, a large share of overviews could not be checked from their own links, so open each one before relying on it.
Has Google made AI Overviews more accurate?
Partly. Under the Gemini 3 model, accuracy on SimpleQA questions rose from 85% to 91% in the NYT and Oumi analysis. In the same analysis, ungrounded answers rose from 37% to 56%. Google says AI Overviews match featured snippets on accuracy in its own tests, and disputes the study's method.
How do I check an AI Overview?
Open the cited link and find the exact sentence that supports the claim. Compare the overview with two or three reputable pages in the normal results. Be wary of forum and satire sources. Check health, money and legal answers with an official source. Confirm the date on the page, since overviews can mix old and new facts.
