Web search: the query it ran was not the one you typed

Search turns an answer from something the model recalls into something it looked up. That is a real improvement and it introduces two steps where a claim can come loose: the query it actually ran, and the gap between what it read and what it cited.

Your question gets rewritten

OpenAI states that ChatGPT search sometimes partners with other search providers, and that when it does, it "typically rewrites your query into one or more targeted queries that it sends those providers".

Two consequences follow, and neither is obvious from the interface.

The search that ran was not your sentence. If an answer misses the point, the rewrite is a likely culprit, and rephrasing your question is a more direct fix than arguing with the answer.

Your question also leaves OpenAI. Whatever you type in a searching conversation may reach a third-party search provider in some rewritten form. That is worth knowing before pasting a client name or an unreleased product into a question you want searched.

Citations are a subset

Responses that use web search may carry citations, and you can select one to open its source or hover to preview it on desktop.

Selecting Sources, where available, gives you cited sources and other relevant links. That second list is the broader one. The inline markers show what the answer leaned on most visibly; they are not a complete account of what was consulted.

The ten-second version of that check: open the cited page and search it for the number, name or phrase the answer attributed to it. Either the string is there or it is absent. That catches the common failure — a real source, a real page, a claim it does not make — without reading the page end to end.

For the shapes of claim most likely to drift, and how to check something outside your own expertise, see verifying output you are not qualified to check.

Making it search, and making it search better

ChatGPT may search on its own when a question would benefit from current information. To force it: View all tools → Search, or type / in the composer and pick Search. An existing answer can be regenerated with web results through the refresh control.

When a result looks wrong, OpenAI's suggested move is to ask ChatGPT to search again naming a specific source, date or location. That is more effective than repeating the question, because it constrains the rewrite step described above.

Web search is available on every plan, Free included, and to people who are not signed in at all. Searches count against your plan's usage limits.

Try this

Ask something current with a specific number in the answer. Open the citation and search that page for the figure. Then open Sources and count how many links are there that were not inline citations. The difference is the part of the research you had not seen.

What goes wrong

Treating a citation as verification. It records that a page was consulted.

Assuming the inline citations are everything. Sources is the longer list.

Rephrasing the same question when the answer misses. Naming a source, date or location constrains the rewrite; rewording rarely does.

Pasting confidential context into a searching conversation. It may be rewritten and sent to a search partner.

Reading a published date as a verified date. OpenAI specifically asks you to check when a source was published or updated.

How to check it worked

Pick the single load-bearing claim in the answer, open its cited page, and find the exact string. If the claim is aggregated across several citations, check each one, because a sentence citing three links is three chances for one of them not to support it. If a source turns out to be older than the answer implies, ask for the search again with a date constraint.