Quick Answer: ChatGPT gets its information from three sources: its training data, a live web search run through third-party search providers, and pages fetched on request. Training data is public web content, licensed partner data and human-generated data, stored as patterns rather than copies. Only the search path produces citations, and OpenAI says placement in those results is not guaranteed.
Most confusion about ChatGPT's sources comes from treating it as one system. It is two: a model that answers from what it learned, and a search layer that fetches current pages and cites them. OpenAI documents both, and the documents are more specific than most summaries of them.
This page sticks to the mechanism. What each source is, how OpenAI says it is collected, and how a source ends up shown to the user. If you want the playbook for getting cited by ChatGPT, that is a separate page.
What are the three sources ChatGPT draws on
ChatGPT draws on training data, a live web search, and pages it fetches when a user asks about them. OpenAI's help centre says its foundation models are developed from "three primary sources of information": public internet content, information accessed through third-party partnerships, and information provided by users, human trainers and researchers, per OpenAI's model development doc.
That description covers the model. The product adds two more paths that operate at answer time, and they behave very differently from the training set.
At a glance
| Source | When it is used | What it produces | The OpenAI bot involved |
|---|---|---|---|
| Training data | Every answer, as the model's baseline knowledge | An answer from memory, no citation | GPTBot (training crawl) |
| Live web search | When ChatGPT decides the question needs current information, or the user turns Search on | An answer with citations and a Sources sidebar | OAI-SearchBot (search index) |
| On-request fetch | When a user asks about a specific page or a Custom GPT calls a site | The page's content in that one conversation | ChatGPT-User (user-triggered fetch) |
Why the split matters
A page can be in one path and absent from the other two. A site that blocks GPTBot but allows OAI-SearchBot can still be cited in search answers, because OpenAI says each robots.txt setting is independent of the others, per its crawler overview. The reverse is also true: a page the model "knows" from training may never be fetched or cited live.
How ChatGPT's training data is collected
ChatGPT's training data is collected from freely and openly accessible web content, third-party datasets accessed under agreements, and data generated by users, trainers and researchers. OpenAI states it does not intentionally gather content from behind paywalls or from the dark web, and it filters out hate speech, adult content, personal-information aggregators and spam before training.
What the model keeps
The model keeps patterns, not pages. OpenAI's wording is that its models "do not store or retain copies of the data they are trained on" and that ChatGPT does not "copy and paste" from its training data. Parameters are adjusted as the model learns, which is why a memorised answer about your brand is a probability, not a lookup.
Where the public web content comes from
GPTBot is the crawler that collects public content that may be used in training. Website owners control it with a robots.txt Disallow, and OpenAI's doc says disallowing GPTBot signals that a site's content should not be used in training its foundation models. That switch has no effect on search or on-request fetches.
Licensed and synthetic data
OpenAI also uses datasets accessed through third-party agreements, and content from publisher partners. At the launch of ChatGPT search it named Associated Press, Axel Springer, Condé Nast, Financial Times, Le Monde, News Corp, Reuters, Time and Vox Media among the publishers it worked with. It also says it "increasingly" uses synthetic data, such as model-generated prompts and multilingual examples, to fill sparse areas.
How ChatGPT search finds and ranks pages
ChatGPT search rewrites your question into targeted queries, sends them to search providers, reads the results, and may send further queries before it answers. OpenAI's announcement says the feature "leverages third-party search providers, as well as content provided directly by our partners", per Introducing ChatGPT search.
Which search providers are involved
OpenAI does not publish a provider list or say which provider serves which query. Its help centre says ChatGPT search "sometimes partners with other search providers" and points readers to Microsoft's privacy statement and Shopify's privacy policy for how those providers handle queries, per Searching the web with ChatGPT. That is the full extent of what OpenAI states, so anything more specific is inference.
How a question becomes a search
The help centre gives a worked example. A researcher asking about drugs that target CCR8 for cancer might trigger an initial provider query of "CCR8 immunotherapy drug development 2025", and after reviewing the results ChatGPT may send a more specific follow-up such as "CHS-114 conference 2025".
Two other inputs shape the query. ChatGPT shares a general location derived from your IP address with providers (not the IP itself), and if memory is on it may fold saved details into the search, turning "restaurants near me" into "good vegan restaurants San Francisco".
What decides which results are used
OpenAI says ChatGPT "ranks search results using multiple factors intended to help users find relevant, reliable information" and that "placement is not guaranteed". The one documented eligibility rule is technical: allow OAI-SearchBot in robots.txt and let traffic from OpenAI's published searchbot IP ranges through your host or CDN. Sites opted out of OAI-SearchBot are not shown in search answers, though they can still appear as navigational links.
How ChatGPT chooses and shows its sources
ChatGPT shows sources as inline citations in the answer and as a Sources button that opens a sidebar of references. Selecting a citation opens the page; on desktop, pointing at one previews it. If the response includes images, each image links to its origin.
What a citation is and is not
A citation means the page was retrieved and used in that answer, in that session. It is not a permanent listing, and OpenAI's own caution is that "search results and citations can be incomplete, outdated, or incorrect". The same question asked twice can cite different pages, because the query rewrite and the provider results both vary.
The on-request path
The third source is a page ChatGPT fetches because a user pointed at it or a Custom GPT called it. That fetch arrives as ChatGPT-User, is not an automatic crawl, and OpenAI says robots.txt rules "may not apply" to it because a person initiated the request. It is also not used to decide what appears in Search, so a page that is fetched this way has not been indexed.
Does ChatGPT get its information from Google
No, not according to anything OpenAI has published. The only search providers OpenAI's documentation refers to are the ones whose privacy policies it links, Microsoft and Shopify, and its 2024 announcement describes "third-party search providers" without naming Google.
That does not mean Google rankings are irrelevant. Pages that rank well are usually the pages that are crawlable, fast and clearly structured, and those are the same properties OAI-SearchBot and any provider index reward. The overlap is in the page, not in a pipe between the two companies.
What OpenAI's search model is
At launch, OpenAI described the search model as a fine-tuned version of GPT-4o, post-trained with synthetic data including distilled outputs from o1-preview. OpenAI's help centre now references GPT-5.6 models, but the shape described then is the one its current docs still describe: a model that decides when to search, what to search for, and how to compose the result.
What the three sources mean for a page that wants to be seen
A page needs to be reachable by OAI-SearchBot to enter the search path, and it needs to be extractable to be cited once fetched. Those are two different jobs, and most sites fail one of them without noticing.
Reachability
Check that robots.txt does not block OAI-SearchBot and that your CDN or firewall is not challenging OpenAI's published IP ranges. The robots.txt rules for each AI bot differ by vendor, and OpenAI notes that a robots.txt change takes about 24 hours to reach its search systems.
Extractability
Once fetched, the page competes with every other result for the sentence the model will quote. Pages that lead with the answer in a structure the model can lift get quoted; pages that bury it under context do not. That is the whole of answer engine optimization in one line, and the ChatGPT search playbook works through it page type by page type.
Memory versus search
The training path is slower and less controllable. It rewards a consistent, well-described entity across the public web, which is what entity signals for AI models are about. If you want to know which path is naming you today, tracking your brand's mentions in ChatGPT with and without search turned on tells you in an afternoon.