How ChatGPT Search Works: A Site Owner's Map of Sources, Providers and Crawlers
ChatGPT search works in two modes: it either answers from what its model learned in training, or it searches the web, rewrites your question into targeted queries for search providers, reads the pages it finds and cites them. This guide maps that pipeline from a website owner's point of view, using only what OpenAI documents, and marks clearly where OpenAI says nothing.
Two ways ChatGPT answers: from the model or from the web
Every ChatGPT reply comes from one of two places, and the difference decides whether your site can be a source at all.
- From the model. The model generates text from patterns it learned during training. OpenAI's article on how its foundation models are developed says the training data comes from publicly available internet content, data accessed through third-party partners, and data from users, trainers and researchers. It also says models "do not store or retain copies of the data they are trained on". There is no citation and no link in this mode.
- From the web. ChatGPT runs a search, reads current pages and answers with citations. This is the mode where a small site can earn a visible link.
Getting into the first mode is about training data, a slow process you mostly cannot steer. Getting into the second is about being crawlable and relevant today. The rest of this article is about the second.
When ChatGPT decides to search
OpenAI's help article Searching the web with ChatGPT says ChatGPT "may search the web automatically when your question would benefit from current information", and that users can also start a search manually from the tools menu or by typing / in the composer. The same article says web search is available on Free, Go, Plus, Pro, Business, Enterprise and Edu plans, and to people who are not signed in.
OpenAI does not publish the exact rule that triggers a search. In practice, questions about recent events, prices, local businesses or "best X for Y" comparisons are the natural candidates, because the model alone cannot know what changed after training. Evergreen definitions may be answered without searching. That is why the same topic can bring you a citation for one phrasing and nothing for another.
Query rewriting: your page competes for a different query
This is the part most site owners miss. ChatGPT usually does not send the user's question to a search engine word for word. According to the same help article, when ChatGPT search partners with other search providers it "typically rewrites your query into one or more targeted queries" and, after reviewing the first results, "may send additional, more specific queries".
OpenAI's own example: a long question about drugs that target CCR8 becomes a first query like "CCR8 immunotherapy drug development 2025", followed by a narrower one naming a specific compound and conference. A "restaurants near me" question becomes "top restaurants San Francisco", using a general location estimated from the IP address. The article says ChatGPT does not share the IP address itself or account information with search providers, and that saved memories may shape the rewritten query if memory is on.
What this means for you:
- Your page has to match short, keyword-like search queries, not only the chatty question a person typed. Classic on-page basics (a clear title, a direct answer near the top, specific terms) still do the work.
- Follow-up queries are narrower. Pages that cover specifics (model numbers, versions, local details, dates) have a second chance to be picked.
- Local results depend on the location in the rewritten query, so a local business page should name its city plainly.
Where the results come from: providers, partners and OpenAI's crawler
"Does ChatGPT use Bing or Google?" is the question everyone asks, and OpenAI answers it only partly:
- The 2024 launch post, Introducing ChatGPT search, says ChatGPT search uses "third-party search providers, as well as content provided directly by our partners".
- The current help article says ChatGPT search "sometimes partners with other search providers" and, for how those providers process queries, links to two privacy policies: Microsoft's and Shopify's. It names no other provider and does not say how much each contributes.
- Microsoft announced in 2023 that "Bing is coming to ChatGPT as the default search experience". That is history, not a description of today's system.
OpenAI also runs its own search crawler, OAI-SearchBot, described below. So the honest summary is: Microsoft is one documented provider, Shopify appears in the shopping context, partners supply some content directly, and OpenAI crawls the web itself. Whether ChatGPT uses Google is not stated in any OpenAI document we could find, so we will not guess. Being properly indexed in Bing is cheap insurance; our ChatGPT indexing guide covers the Bing Webmaster Tools setup.
Four OpenAI bots and what each one does
OpenAI lists its crawlers on Overview of OpenAI crawlers. Each has its own user agent and its own published IP list, and robots.txt settings for each are independent.
| Bot | Job | Role in ChatGPT search | robots.txt |
|---|---|---|---|
OAI-SearchBot | Surfaces sites in ChatGPT search results | The one that matters. Opted-out sites are not shown in search answers, though they can still appear as navigational links | Followed |
ChatGPT-User | Visits a page when a user's action in ChatGPT or a custom GPT needs it | On-demand fetch, not automatic crawling; OpenAI says it does not decide whether content appears in search | OpenAI says rules "may not apply" because a user triggered the visit |
GPTBot | Crawls content that may be used to train foundation models | None. Blocking it is a training opt-out only | Followed |
OAI-AdsBot | Checks landing pages submitted as ChatGPT ads for policy compliance | None for organic results | Not relevant unless you advertise |
Two details from the same page are useful in practice: if a site allows both OAI-SearchBot and GPTBot, OpenAI may use one crawl for both purposes, and search changes take about 24 hours to apply after a robots.txt update. The full robots.txt setup, including how to allow search but refuse training, is in our guide to blocking AI crawlers with robots.txt.
How ChatGPT picks which sources to cite
OpenAI's public statement on ranking is one sentence: ChatGPT "ranks search results using multiple factors intended to help users find relevant, reliable information", and "placement is not guaranteed". The factors are not listed. Anyone who sells you the "ChatGPT ranking algorithm" is describing their own guesses.
What follows from the documented parts is enough to act on:
- Eligibility first. The help article gives two conditions: allow OAI-SearchBot, and make sure your host or CDN allows traffic from OpenAI's published search bot IP addresses.
- Retrievability second. Your page must come back for the rewritten, keyword-like queries, from OpenAI's own crawl or from a provider.
- Selection third. Among retrieved pages, ChatGPT chooses what to cite. "Relevant, reliable" is the only hint, which points to clear answers on the page and a reputation built outside your own site. Our guide on how to get cited by ChatGPT covers that part.
What you can and cannot control
| Part of the pipeline | Can you influence it? | How |
|---|---|---|
| Whether ChatGPT searches at all | No | Decided per question; write for queries that need fresh or specific information |
| Rewritten query wording | Indirectly | Use the plain terms people search for; name places, versions and dates |
| OAI-SearchBot access | Yes, fully | robots.txt plus firewall and CDN rules that allow OpenAI's search IPs |
| Presence in provider indexes | Partly | Bing Webmaster Tools, sitemaps, IndexNow |
| Readable content | Yes | Main text in the HTML, fast responses, no login walls on public pages |
| Being chosen as a citation | Partly | Direct answers, specific facts, independent mentions of your site |
| Training data | Only opt-out | Disallow GPTBot; this does not affect search |
A worked example (illustrative)
Imagine a small bakery-supply blog with a page called "How to store sourdough starter in the fridge". A user asks ChatGPT: "I'm going on vacation for two weeks, will my starter survive?"
- ChatGPT decides fresh sources would help and searches.
- It rewrites the question into something like "sourdough starter fridge two weeks" and sends it to a provider.
- Results come back; if the blog is in OpenAI's own index too, it can be considered from there.
- ChatGPT reads the candidate pages and cites the ones that answer clearly. If the blog's answer is buried under a long story and the firewall returns 403 to OpenAI's IPs, the page loses at steps 3 and 4 regardless of quality.
The example is invented to show the flow; it is not a measured case.
What to do next
Run the three checks that cover the documented eligibility rules: read your live robots.txt for OAI-SearchBot, search your server logs for that user agent and its status codes, and confirm your CDN does not challenge OpenAI's IP ranges. If ChatGPT already sends visitors, our post on tracking AI traffic in GA4 shows how to measure it.
The fastest start is Satelo's free AI crawler checker: enter a URL and see, bot by bot, whether your robots.txt and server responses let OAI-SearchBot, GPTBot and other AI crawlers in. It shows the actual response each bot would get, which robots.txt alone cannot tell you.
FAQ
How does ChatGPT search the web?
When a question would benefit from current information, ChatGPT rewrites it into one or more targeted search queries, sends them to search providers it partners with, reviews the results, may run narrower follow-up queries, and answers with citations. OpenAI also runs its own search crawler, OAI-SearchBot.
Does ChatGPT use Bing or Google?
OpenAI says ChatGPT search uses third-party search providers and content from partners, and its help article links to Microsoft's and Shopify's privacy policies for how providers process queries. OpenAI does not name Google, and it does not say how much each provider contributes.
Does ChatGPT actually search the web or just use training data?
Both. Without search, answers come from the trained model and have no citations. With search, ChatGPT reads current pages and links to them. It may search automatically, and users can also start a search manually.
Where does ChatGPT get its information?
Training data comes from publicly available internet content, third-party partner data, and data from users, trainers and researchers. Search answers come from search providers, partner content and pages crawled by OAI-SearchBot.
Can I make ChatGPT cite my website?
Not directly. OpenAI says placement is not guaranteed. You can make the site eligible by allowing OAI-SearchBot and OpenAI's search IP addresses, and improve your chances with clear answers and independent mentions.
Does blocking GPTBot remove my site from ChatGPT search?
No. GPTBot is the training crawler. Search visibility depends on OAI-SearchBot, and the two are controlled separately in robots.txt.
Read next
Check your site's Quality Score
Free, no sign-up: a 0–100 score from mostly open data. Then add your site to Satelo and get partner articles from sites of your level.
Enter a domain like example.com
Add my site