TL;DR
ChatGPT chooses its sources in two stages: it decides whether a prompt needs a web search, and if it does, it rewrites the prompt into search queries, pulls results from third-party search providers and pages reached by OpenAI's own OAI-SearchBot crawler, then cites the pages that supported the answer it wrote. OpenAI documents that pipeline but does not publish a ranking formula for which pages get cited, so anything beyond it is inference. What is documented clearly: sites that block OAI-SearchBot are not shown as sources in ChatGPT search answers, and GPTBot, the training crawler, is a separate switch. In practice, cited pages tend to be ones that already rank in the underlying search results, answer a sub-question plainly, and agree with other sources.
Two ways ChatGPT answers, and only one of them cites
Before asking how ChatGPT chooses its sources, it helps to know when it uses sources at all. ChatGPT can answer a prompt in two different ways, and they behave very differently for your brand.
- From what the model learned in training. No search happens, so there are no source links. If your brand is mentioned, it is because the model absorbed enough about you from its training data to recall you.
- With a live web search. ChatGPT looks things up, writes the answer from what it found, and shows citations: inline source links next to specific claims, plus a fuller list of sources the answer drew on.
OpenAI says ChatGPT decides on its own whether a question would benefit from information on the web, and users can also trigger a search directly. Questions about recent events, prices, product comparisons and anything time-sensitive usually trigger a search. Stable general knowledge often doesn't. Every ChatGPT citation you see comes from the second path, which is why the rest of this post is about search.
What OpenAI publicly documents about ChatGPT search
OpenAI's help and publisher documentation is short, but it settles more than most GEO advice admits. Here is what it covers.
- ChatGPT search uses third-party search providers, along with content provided directly by OpenAI's partners. OpenAI has content partnerships with a number of news publishers.
- ChatGPT can rewrite a prompt into one or more targeted search queries before sending them to those providers, and it may use context such as your general location to make results more relevant.
- OAI-SearchBot is the crawler OpenAI uses to surface websites in ChatGPT search. Sites that opt out of it in robots.txt are not shown as sources in ChatGPT search answers, though OpenAI notes they can still appear as plain navigational links.
- GPTBot is a separate crawler that collects content which may be used to train OpenAI's models. Blocking GPTBot and allowing OAI-SearchBot is a supported combination; each setting is independent.
- ChatGPT-User is the agent used when ChatGPT visits a page because of a user's request, for example when someone pastes a URL and asks about it.
- Links from ChatGPT search carry a utm_source=chatgpt.com parameter, so publishers can see that traffic in their analytics.
That is close to the full public picture. There is no published list of ranking factors, no citation score, and no submission form. Anyone who claims to know OpenAI's exact selection formula is describing their own observations, not OpenAI's documentation.
The ChatGPT citation pipeline, step by step
Putting the documented parts together with what you can observe in the interface gives a reasonable working model. The first three steps are documented; the last three are observable behavior and inference.
- A prompt arrives, and ChatGPT decides whether it needs to search. (Documented.)
- If it does, the prompt is rewritten into one or more search queries, often narrower than the original question. (Documented.)
- Those queries go to search providers, which return ranked results using their own systems, and OpenAI's crawler supplies access to page content. (Documented in general terms; the providers' ranking is their own.)
- ChatGPT reads what it can access from the top results: snippets, and page content for pages it is allowed to fetch. (Inference from how retrieval systems work.)
- It writes the answer and attaches citations to the claims each page supported. (Observable in every search answer.)
- The sources list shows more pages than the inline citations, which suggests more pages are read than end up credited in the text. (Observable.)
The important consequence sits in step 3. ChatGPT is not running its own ranking of the whole web for each prompt. It largely works from what search providers return for its rewritten queries. A page that doesn't appear in those results is unlikely to be cited, however good it is.
What likely decides which page gets cited
Everything in this section is our inference, drawn from watching answers across many prompts and from how retrieval systems generally work. OpenAI has not confirmed any of it, so treat it as a set of working hypotheses, not rules.
The page has to be in the candidate set
Retrieval comes first. If your page doesn't rank in the underlying search results for the rewritten query, it never gets read. OpenAI doesn't publish a full list of its search providers. Microsoft Bing is widely understood to be one of them, so Bing indexing and rankings are a sensible proxy to watch, but we wouldn't treat Bing as the whole story.
Relevance to the sub-query, not the whole prompt
Because prompts get rewritten into narrower searches, a page can be cited for one part of an answer. A long prompt about choosing a payments API might cite one page for fees, another for supported countries, and a third for developer experience. Pages that cover one sub-question thoroughly often get cited where broad overview pages don't.
A clear statement the model can attach a citation to
A citation is pinned to a claim. Pages that state the claim plainly, in text, near the top of a relevant section, are easier to pin to than pages where the fact is implied, spread across paragraphs, or locked in an image or a gated PDF.
Agreement with other sources
When several retrieved pages say the same thing, the answer tends to state it confidently and cite more than one of them. A page making a claim no other source supports is a riskier thing to repeat. This is one reason consistent descriptions of your product across review sites, comparisons and your own pages matter so much.
Freshness and source type that fit the question
For time-sensitive questions, newer and recently updated pages tend to win. The type of source also tends to match the question: review and comparison sites for "best tool" prompts, official documentation for technical how-to questions, community threads for questions about real-world experience, and encyclopedic references for definitions.
| Claim you may hear | Status | What we can actually say |
|---|---|---|
| Blocking OAI-SearchBot removes you as a source in ChatGPT search | Documented by OpenAI | True. Allow it if you want to be cited. |
| Blocking GPTBot removes you from ChatGPT search | Contradicted by OpenAI | GPTBot is the training crawler; it is a separate setting. |
| ChatGPT search uses third-party search providers | Documented by OpenAI | Yes, though the full list of providers isn't published. |
| Pages that rank well in search get cited more | Reasonable inference | Consistent with how the pipeline is described, not confirmed. |
| Schema markup makes ChatGPT cite you | No public support | Nothing OpenAI publishes says this. It may help parsing; it isn't a known factor. |
| Domain authority scores decide citations | No public support | DR and DA are vendor estimates. Google doesn't use them, and nothing suggests ChatGPT does. |
Why ChatGPT cites different sources each time
Run the same prompt twice and you will often get a different set of citations. That isn't a bug, and it changes how you should measure AI citations.
- The model generates text with some randomness, so wording and structure vary between runs, and different claims need different sources.
- The rewritten search queries can differ from run to run, which changes the results that come back.
- Context differs: location, conversation history and, for logged-in users, memory and personalization can all shape the search.
- Search indexes change constantly, as pages are crawled, updated and re-ranked.
- OpenAI updates its models and search features, sometimes changing citation behavior noticeably overnight.
The takeaway is to judge citation visibility across many runs of a fixed prompt set, not from a single screenshot. One answer that leaves you out proves very little; a pattern across a month of runs proves a lot.
Why your brand gets mentioned but not cited, or the reverse
A mention and a citation come from different places. ChatGPT can name your brand from training data with no search and no link, which is common for established brands. It can also cite your docs for a technical detail while recommending a competitor in the same answer, because your page supported a fact but other sources shaped the recommendation.
That is why we track mentions, citations and recommendations separately. Getting cited often means fixing your own pages and their search visibility. Getting recommended usually means changing what independent sources say about you, which is the work we describe in what AI brand mentions SEO is. Both sit inside the broader practice of generative engine optimization.
How to check whether ChatGPT can cite your site
Before any content or outreach work, confirm the plumbing. This takes an afternoon and catches the problems that make everything else pointless.
- Robots.txt: confirm OAI-SearchBot is allowed on the pages you want cited. Our robots.txt tester checks each AI crawler's rules for a given URL.
- CDN and firewall: some security settings block AI crawlers by default or with a single toggle, regardless of robots.txt. Check yours.
- Server logs: look for OAI-SearchBot requests. If it never visits, something upstream is stopping it.
- Rendering: make sure the content you want quoted is in the HTML the server returns, not only injected by JavaScript after load.
- Search visibility: confirm your key pages are indexed in Bing as well as Google, since Bing is widely understood to be one of the providers ChatGPT draws on.
- Analytics: filter for utm_source=chatgpt.com and the chatgpt.com referrer to see whether citations are already sending visits.
Once access is confirmed, the work shifts to content and third-party coverage. We lay that out as an ordered playbook in how to get cited by ChatGPT.
How Chung Digital uses this in practice
We build our Mentions work around what is documented and test everything else. As a generative engine optimization agency, we start by checking crawler access and search visibility, then track a fixed set of buyer prompts across ChatGPT, Gemini, Perplexity and AI Overviews every month, recording which sources get cited and whether your brand is mentioned, cited or recommended.
That citation log tells us which third-party pages shape answers in your category, and those become the targets for brand mention placements. We don't claim to know OpenAI's selection formula, because nobody outside OpenAI does, and we don't guarantee citations. What we can do is make your brand easy to retrieve, easy to quote, and consistently described by the sources ChatGPT already reads.
Frequently asked questions
Does ChatGPT use Google search results?
OpenAI says ChatGPT search uses third-party search providers but does not publish a complete list of them. Microsoft Bing is widely understood to be one. Since the full picture isn't public, the safest approach is to be well indexed and visible in both Google and Bing.
If I block GPTBot, will ChatGPT stop citing my site?
No. GPTBot is OpenAI's training crawler, and OpenAI treats it as a separate setting from OAI-SearchBot, the crawler behind ChatGPT search. You can block GPTBot to opt out of training and still allow OAI-SearchBot so your pages can be cited.
Can you pay to be cited in ChatGPT answers?
There is no documented way to buy a citation inside ChatGPT's answers. OpenAI does have content partnerships with some publishers, but those are licensing arrangements, not a paid placement product for brands. Getting cited depends on being retrievable and relevant.
Does ChatGPT read schema markup when choosing sources?
OpenAI hasn't said. Structured data may help any system parse a page, and it helps search engines that feed retrieval, but there is no public evidence that schema is a direct factor in which page ChatGPT cites. Treat it as useful hygiene, not a lever.
How often does ChatGPT update what it knows about my brand?
It depends on the path. Search-grounded answers can reflect a page as soon as it is crawled and indexed by the search providers. What the model knows without searching only changes when OpenAI trains and releases a new model, which happens on OpenAI's schedule.
Do ChatGPT citations send real traffic?
They can. Citation links carry utm_source=chatgpt.com, so the visits are visible in analytics. Many users read the answer without clicking, though, so a citation is often worth as much for visibility and trust as for direct clicks.
Why does ChatGPT cite a competitor's blog instead of our product page?
Usually because the competitor's page answered the specific sub-question more directly, ranked better for the rewritten search query, or was corroborated by other sources. Product pages written as marketing copy often lack the plain, specific statements an answer needs to cite.
Are AI citations the same across ChatGPT, Perplexity and Gemini?
No. Each platform uses different search infrastructure and different models, so the same prompt often produces different sources. Perplexity shows sources for nearly every answer, Gemini and Google AI features draw on Google's index, and ChatGPT relies on its own crawler plus third-party providers.