One of the many ways that AI chatbots like ChatGPT and Gemini can help you is in searching the web. Indeed, if you load up the Google website these days, AI search is pretty much forced on you—the traditional, ranked list of blue links is some way down the page, if it appears at all.
This new type of information retrieval can feel more natural and convenient—like having a conversation with an expert rather than diving through pages and pages of results—but is it actually any more accurate and trustworthy?
No matter what you’re searching for or how you’re going about it, there are some important details to know when it comes to how AI-powered search works, and how you should weigh up the answers you get back.
How does AI web search work?
Different AIs take different approaches to web search, but broadly speaking, a web search inside an AI chatbot has three stages. First, the query is parsed and interpreted; then, relevant webpages are pulled up from the web (sometimes, but not always, with the help of Google and Bing); and finally, a relevant response is packaged together in that familiar chatbot style.
Your web search may be split into smaller searches, run separately, and then combined again. In some cases, AIs also lean on built-in tracking data rather than checking the web (perhaps for well-established knowledge where the most up-to-date details aren’t required).
Google Gemini, for example, uses a technique called query ‘fan-out’ to run sub-searches and follow-ups. This complexity and comprehensiveness is what enables you to put in much more complex questions, and get much more detailed answers back—as with a query like “the best attractions for elderly tourists in the city of Chicago that cost less than $10 to visit” that will return a more generic list of results in a traditional Google search.
AI search will usually give you a correct result—if it can find the right sources. David Nield
You’ll notice that AI chatbots will often include citations back to web links where it’s found the information for its results, but this doesn’t always happen—and the relevant details may not always be visible in the cited link. It’s a multi-stage process in breaking down the query, retrieving the right information, and presenting it back again, but the lines between those stages aren’t always clear to users.
Part of the work AI has to do is figuring out which online sources match the user’s search, and how reliable and relevant those sources are. This is only partly weighted by a site’s prominence in Google or Bing results, because an AI chatbot will also look at the relevance to the query: A lesser-known site might be given more prominence if it has information more relevant to what you’re asking about.
AI companies don’t fully or publicly explain all these details, but that’s a general overview of what happens when you send a chatbot to get answers from the web. A lot of the time it can feel like a step up from a classic web search, unless you want to take more specific control over the way that the search is run, which sources are chosen, and how the information gets synthesized.
Can you trust AI web search?
There’s an increasing amount of data to help answer this question. A May 2026 study based on tens of thousands of Google AI Overview (AIO) queries found that 11 percent of results (so more than 1 in 10) were not supported by the citations given—as the researchers themselves say, “a failure mode a reader who sees only the AIO summary has no way to detect.”
Another study published in March 2026 looked at source credibility and found that ChatGPT’s GPT-5 model quoted reliable sources 71.4 percent of the time, with a media fact-checking database used to determine the trustworthiness of individual sites. That’s not to say 3 out of 10 answers were wrong, just that they used sources not recognized as authoritative.
A certain amount of skepticism is healthy when assessing answers from AI web searches, then. Of course, standard web searches can turn up wrong information too, but the difference is you’re doing the research work: You’re deciding which websites are reliable, and joining the dots between sources, rather than relying on AI to do it.
Always check the sources AI has used to compile its results. David Nield
It’s also important to bear in mind that if there’s not much information about your query online, or no information at all, that’s instantly visible on a classic Google or Bing search. An AI chatbot is more likely to try and put together some kind of answer for you—perhaps there’s something in its training data that is relevant—and present it with confidence. Again, the convenience and ease of an AI search can be misleading.
Because AI search is a multi-step process, it can fail in several ways: misunderstanding the question, choosing the wrong information online (maybe false, out of date, or just not relevant), or mixing sources incorrectly. These are the checks to run when weighing up whether what you’ve been given is accurate.
OpenAI itself encourages ChatGPT web search users to check citations and sources, and to focus on authoritative sources for important questions (which is presumably all of them). While AI search can usually do an impressive job of bringing together multiple sources of information in a helpful way, check its work—so in our Chicago example, don’t set out for a visit until you’ve verified the opening hours, prices, and accessibility options directly with the location.
Disclosure: Ziff Davis, Popular Science’s parent company, filed a lawsuit against OpenAI in April 2025, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.
The post Why you shouldn’t trust everything you find in an AI web search appeared first on Popular Science.
Why you shouldn’t trust everything you find in an AI web search
