ChatGPT search is the feature that lets ChatGPT look things up on the web before it answers. When a question needs current or specific information, the model writes search queries, retrieves pages from a search index, reads the most relevant passages and writes a reply with links to the pages it used. Without search, ChatGPT answers only from what it learned during training.
The process is a form of retrieval-augmented generation: the system retrieves text first and generates the answer from it second. The quality and the sources of the answer depend on what the retrieval step returns.
When does ChatGPT decide to search the web?
A user can turn search on explicitly, but in most conversations the model decides for itself. It searches when the question is likely to need information that is newer than its training data or too specific to trust from memory.
Typical triggers include:
- Questions about recent events, prices, schedules or availability.
- Requests for product recommendations or comparisons.
- Questions about a specific local business, person or company.
- Requests that mention a URL or ask the model to check a source.
For general explanations, writing tasks or coding help, ChatGPT often answers without searching. The same brand can therefore appear in one answer because a page was retrieved and in another because the model remembered it from training, and the two answers can say different things.
How does ChatGPT turn a question into searches?
The model writes its own search queries from the prompt and the earlier conversation, so what reaches the search engine is usually different from what the user typed. A follow-up such as "which one is cheaper?" gets rewritten into a full query that names the products discussed before.
Broad questions are often split into several narrower queries that run separately. A question about the best accounting software for a small agency might produce queries about pricing, about features for agencies and about recent reviews. Each query returns its own set of results, so a page can enter the answer through a query the user never typed.
If the first results leave a gap, the model can search again after reading them. As a result, the set of retrieved pages depends on how the conversation is worded as well as on its topic.
Where do the search results come from?
ChatGPT search draws on third-party search providers, with Microsoft's Bing the best known, together with pages that OpenAI fetches with its own crawlers. A page that Bing has not indexed, or that OpenAI's search crawler cannot reach, has fewer ways into an answer. Bing Webmaster Tools shows whether Bing has indexed a site's pages.
OpenAI runs separate bots for separate jobs, and each has its own user agent (the name a bot sends when it requests a page):
| Bot | Purpose |
|---|---|
| OAI-SearchBot | Crawls pages so they can appear in ChatGPT search results |
| ChatGPT-User | Fetches a page live when a user's request requires it |
| GPTBot | Collects content for training future models |
Because the bots are separate, a site can block training with GPTBot while still allowing OAI-SearchBot, or the reverse. Rules for each bot are set in robots.txt, the file at a site's root that tells crawlers which URLs they may fetch. What are AI crawlers? covers how these bots behave across different AI companies.
How does ChatGPT pick passages and cite them?
Once pages are retrieved, the system strips menus, footers and other page furniture, then splits the remaining text into passages. Not every passage fits into the model's context window, the limit on how much text it can read at once, so a ranking step keeps the passages that best match the question.
The model then writes the answer from those passages. Sentences that rely on a specific source get a citation, shown as an inline link or a source chip, and the full list of sources appears beside the answer. The general process is described in how AI search chooses which sources to cite.
Several things make the output vary between runs:
- The model samples its words, so two answers to the same prompt are rarely identical.
- The queries it writes can differ slightly, which changes the retrieved pages.
- The search index changes as pages are added, updated or removed.
A single check shows one possible answer. To learn how often a brand or page appears, you have to run the same prompts many times.
What does ChatGPT search mean for a website?
For a page to influence a ChatGPT answer, it has to be reachable, indexed and readable. In practice, OAI-SearchBot and ChatGPT-User should not be blocked in robots.txt, the page should be in Bing's index, and the main text should be present in the HTML instead of loaded only by scripts that crawlers may not run.
The passage-ranking step favors text that answers a question on its own. A clear opening sentence, a definition, a list of options or a table with real values is easier to select than a paragraph that depends on the three before it. Product pages also benefit when the product is mentioned on review sites, in comparison articles and on forums, because ChatGPT often builds recommendation answers from those third-party pages.
When a user clicks a citation, the visit arrives as AI referral traffic, usually with chatgpt.com as the referrer and often with a utm_source=chatgpt.com parameter added to the link. Many users read the answer without visiting any source, so click counts miss much of a brand's exposure. Tracking how often a brand is mentioned or cited across repeated prompts gives a more complete measure.
