How ChatGPT Grounding Works When It Browses the Web

How ChatGPT grounding works

ChatGPT grounding is what happens when the assistant retrieves live content from the web during a session and bases part of its answer on that retrieved material rather than relying only on knowledge learned during training. When browsing or a similar retrieval mode is active, the assistant runs a search step, reads a set of retrieved pages, and can cite specific ones as sources in the text it returns.

Grounding does not happen for every message in every conversation. A large share of ChatGPT's responses are still generated purely from trained knowledge, with no retrieval step involved at all, which is why the same question can sometimes produce a grounded, source-citing answer and other times produce a plain, uncited one depending on how the assistant decides to handle that particular turn.

When ChatGPT retrieves versus answers from memory

Distinguishing a grounded answer from a memory-based one usually comes down to reading the response itself. A grounded answer typically includes visible citations, links or footnote-style references pointing to specific pages the assistant consulted, and sometimes explicit language noting that it searched the web for the question. A memory-based answer has neither, and reads as a direct response drawn from what the model learned during training rather than what it retrieved for that particular prompt. The search step behind a grounded answer works much like a grounding query in any retrieval-augmented system: the assistant reformulates the question into a search-friendly form, runs it, and reads back whatever the search step returns before writing anything.

What ChatGPT's browsing has historically drawn on

OpenAI has not published a full technical account of the search infrastructure behind ChatGPT's browsing feature. What can be said with more confidence is a narrower, historical observation: ChatGPT browsing has historically drawn on Bing-family search infrastructure for at least part of its underlying retrieval, a detail that has been reported and is consistent with how the feature has behaved over time. That does not mean every retrieval ChatGPT performs works identically to a Bing web search, and it does not describe the current state of any specific version with certainty. Treat this as a directional fact about lineage rather than a precise technical specification of how a given ChatGPT response was grounded.

What you can observe versus what is inference

What a publisher can actually observe is limited but real. Visible citations inside a grounded ChatGPT answer are observable, as is language in the response noting that a search happened. Server logs are also observable: GPTBot is OpenAI's documented crawler user agent, and robots.txt governs whether it can access a given page, so checking server logs for GPTBot requests shows whether that crawler has been reaching your content. What is not observable is the internal decision process that determines which retrieved pages get cited in a given answer versus which get read but left out. Treating a citation you can see as confirmed, and everything else as inference, rather than blending the two, keeps conclusions about ChatGPT grounding honest.

Why grounding output can vary across sessions

The same question asked in separate ChatGPT sessions can produce different grounded answers, with different sources cited or no citations at all. This happens because grounding depends on a live search step, and a live search step can return different results depending on when it runs, how the question was rephrased internally, and which retrieval mode the assistant chose to use for that particular turn. None of this is a malfunction. It reflects how a retrieval-based system behaves by design, since it queries a live, changing source rather than a fixed, unchanging one. Why AI answers change between sessions covers this variability in more depth across engines generally, not only ChatGPT.