Google explains how it is improving its AI Overviews

Google responded to the bad press surrounding its recently rolled out AI Overviews in a new blog post by its new head of Search, Liz Reid. Google explained how AI Overviews work, where the weird AI Overviews came from, the improvements Google made and will continue to make to its AI Overviews.

However, Google said searchers “have higher satisfaction with their search results, and they’re asking longer, more complex questions that they know Google can now help with,” and basically, these AI Overviews are not going anywhere.

As good as featured snippets. Google said the AI Overviews are “highly effective” and based on its internal testing, the AI Overviews “accuracy rate for AI Overviews is on par with featured snippets.” Featured snippets also use AI, Google said numerous times.

No hallucinations. AI Overviews generally don’t hallucinate, Google’s Liz Reid wrote. The AI Overviews don’t “make things up in the ways that other LLM products might,” she added. AI Overviews typically only go wrong when Google “misinterpreting queries, misinterpreting a nuance of language on the web, or not having a lot of great information available,” she wrote.

Why the “odd results.” Google explained that it tested AI Overviews “extensively” before releasing it and was comfortable releasing it. But Google said that people tried to get the AI Overviews to return odd results. “We’ve also seen nonsensical new searches, seemingly aimed at producing erroneous results,” Google wrote.

Also, Google wrote that people faked a lot of the examples, but manipulating screenshots showing fake AI responses. “Those AI Overviews never appeared,” Google said.

Some odd examples did come up, and Google will make improvements in those types of cases. Google will not manually adjust AI Overviews but rather improve the models so they work across many more queries. “we don’t simply “fix” queries one by one, but we work on updates that can help broad sets of queries, including new ones that we haven’t seen yet,” Google wrote.

Google spoke about the “data voids,” which we covered numerous times here. The example, “How many rocks should I eat?” was a query no one has searched for prior and had no real good content on. Google explained, “However, in this case, there is satirical content on this topic … that also happened to be republished on a geological software provider’s website. So when someone put that question into Search, an AI Overview appeared that faithfully linked to one of the only websites that tackled the question.”

Improvements to AI Overviews. Google shared some of the improvements it has made to AI Overviews, explaining it will continue to make improvements going forward. Here is what Google said it has done so far:

Built better detection mechanisms for nonsensical queries that shouldn’t show an AI Overview, and limited the inclusion of satire and humor content.

Updated its systems to limit the use of user-generated content in responses that could offer misleading advice.

Added triggering restrictions for queries where AI Overviews were not proving to be as helpful.

For topics like news and health, Google said it already have strong guardrails in place. For example, Google said it aims to not show AI Overviews for hard news topics, where freshness and factuality are important. In the case of health, Google said it launched additional triggering refinements to enhance our quality protections.

Finally, “We’ll keep improving when and how we show AI Overviews and strengthening our protections, including for edge cases, and we’re very grateful for the ongoing feedback,” Liz Reid ended with.

Why we care. It sounds like AI Overviews are not going anywhere and Google will continue to show them to searchers and roll them out to more countries and users in the future. You can expect them to get better over time, as Google continues to hear feedback and improve its systems.

Until then, I am sure we will find more examples of inaccurate and sometimes humorous AI Overviews, similar to what we saw when featured snippets initially launched.