Can you trust the results: How accurate are Google’s AI overviews in search results?

Artificial intelligence is rapidly changing how people search for information online. One of the biggest shifts came when Google introduced AI Overviews, a feature that places AI-generated answers at the top of search results.

The system is designed to provide quick summaries instead of simply showing links. But as AI answers become more visible, questions about accuracy, transparency and trust are growing. According to a recent analysis, AI Overviews are correct most of the time but errors still occur at scale.

What are Google’s AI Overviews?

AI Overviews were introduced in 2024 as part of Google’s broader push to integrate artificial intelligence into search. The feature uses AI models from the Google Gemini system to generate direct answers to user queries. Instead of scrolling through multiple websites, users see a short AI summary at the top of the results page. The summary usually includes links to websites that support the information.

This change has effectively turned Google from a search engine that directs users to websites into a platform that sometimes publishes its own AI-generated answers.

What the accuracy study found

An analysis by AI startup Oumi evaluated how accurate these AI-generated answers are.

At the request of The New York Times, the company tested 4,326 Google search queries using a benchmark known as SimpleQA, which is commonly used in the AI industry to measure system accuracy.

The results showed:

AI Overviews were 85 per cent accurate when powered by Gemini 2

After an upgrade to Gemini 3, accuracy increased to 91 per cent

More than 5,000 sources were cited across the analysed responses

These findings suggest the technology is improving, but it is still not perfect.

Given that Google processes more than 5 trillion searches per year, even a small error rate could lead to millions of incorrect responses being generated regularly.

The problem of “ungrounded” answers

Another issue raised in the analysis concerns how AI Overviews cite sources.

Researchers found that more than half of the accurate responses were “ungrounded.” This means the links cited by Google did not fully support the information presented in the AI summary.

This creates challenges for users trying to verify the accuracy of the answer.

The study also observed that social platforms such as Facebook and Reddit appeared frequently among cited sources.

For example:

Facebook was cited 5 per cent of the time in accurate answers

It appeared 7 per cent of the time in inaccurate answers

Google disputes parts of the analysis

Google has acknowledged that AI Overviews can sometimes contain mistakes. However, the company also criticised the methodology used in the analysis. A Google spokesperson said the research relied on a benchmark created by OpenAI, which the company argues contains incorrect information.

According to Google, relying on such benchmarks could affect the reliability of the findings.

Another challenge with AI evaluation is that researchers often use AI systems themselves to verify answers. This means the checking process may also introduce errors.

Why the debate over AI search matters

AI-powered search tools promise faster answers and a simpler user experience. But experts warn that users may trust AI summaries without verifying them. Some technologists say the technology is already useful and improving quickly. Others argue that the average user may not realise that AI-generated results should still be checked.

As AI systems become central to how people access information online, ensuring accuracy and transparency will become increasingly important.

For now, the research suggests that while Google’s AI Overviews are often correct, users should still treat them as a starting point not the final word.