How ChatGPT decides what to cite, and how to get cited
The short answer: ChatGPT makes two decisions, not one. First it turns your question into one or more searches and pulls back a shortlist of documents from the index its search crawler builds. Then, while writing, it attaches a citation to the few passages it actually leaned on. Being crawlable only wins you the first decision. The second is won by the passage, not the page. That is ChatGPT's search behaviour, which is what this post is about. Other engines are described in their own documentation and we do not assume they work the same way.
How ChatGPT cites sources, in one paragraph. It rewrites your question into machine-written searches, pulls a shortlist of documents out of the index its search crawler builds, writes the answer, and attaches a link to the passages the sentences actually came from. The link you see is a claim about one sentence, not a reading list, and a page can be retrieved and then credited to nobody.
How to get cited by ChatGPT, in two steps. First, let OAI-SearchBot in: OpenAI's documentation says a site opted out of that agent will not be shown in ChatGPT search answers, and the rule most sites got wrong was written for GPTBot. Second, give the writing step something it can lift, which means one sentence per fact with the number, the unit and the date inside it. Nobody outside OpenAI can promise you a citation, because nobody outside OpenAI can see the selection step.
Most teams treat citation as one event and optimise the wrong half. The two halves fail for different reasons and take different work to fix.
Two decisions, and only one of them involves your robots file
The retrieval decision happens before any text is generated. Your question becomes a small set of machine-written queries, those queries hit an index, and a handful of documents come back. Nothing about your writing style matters yet. What matters is whether the right crawler was allowed in, whether the page rendered server side, and whether the index holds a reasonably fresh copy.
The citation decision happens during generation. The model has a working set of passages in front of it and is composing a few hundred words. A source gets credited when a sentence in the answer traces back to it. Pages that were retrieved and then ignored are invisible to you and to your analytics, which is exactly why the failure is so hard to diagnose from the outside.
Hold the two apart like this: retrieval is access and index presence, citation is whether a liftable claim existed in the text. The preprint cited further down fitted the same two stage pattern on Gemini as well as on GPT, with different coefficients. We have not measured it on the other assistants, and how each AI engine picks the sources it cites sets out what each one documents.
The crawler people block is not the crawler that cites them
OpenAI's crawler documentation, which we read on 12 October 2026, lists four user agents doing different jobs. GPTBot gathers content that may feed model training. OAI-SearchBot surfaces sites in ChatGPT's search features. ChatGPT-User fetches a page because a person or a tool asked for it in the moment, so it is not a bulk crawler at all. OAI-AdsBot visits only pages submitted as ads, and the documentation says its data is not used to train foundation models.
OpenAI's own documentation is blunt about the consequence: a site opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though it can still appear as a navigational link. The documentation also notes that a robots.txt change can take about a day to take effect on the search side. A site that blocked GPTBot in 2023 for training reasons and later copied the rule down the file to the newer agents has opted itself out of ChatGPT's search answers as well. We find that pattern in audits. We have no sample that would let us say how common it is across the web.
You can verify the allowlist rather than argue about it. OpenAI publishes the IP ranges its crawlers use as machine readable files, one per agent. We fetched https://openai.com/searchbot.json on 8 October 2026 and again on 12 October 2026: both times it carried 39 IPv4 prefixes and a creationTime of 2 January 2026. That is one file read twice, not a monitored series, so refetch it rather than trusting our count. Grep your access logs for those prefixes. If no request from them appears in your logs for a page, rewriting that page is not the fix, and the self-check has the rest of the access tests in the same spirit.
On GPT and Gemini, retrieval looks like most of the battle
A September 2026 preprint on arXiv, *From Prompt to Recommendation: A Fitted Stage Model of Brand Visibility in AI Search*, fitted a model to 34,960 unbranded prompt and engine observations drawn from 75 anonymised monitoring projects, covering 2,854 distinct prompts run repeatedly on GPT and Gemini between June and September 2026. Its central finding is a staircase. It is one panel of monitored prompts on two engines, not a census of either engine.
When neither the brand nor its own domain showed up anywhere in the observable retrieval path, GPT mentioned the brand in 2.8% of runs. When the brand's own domain was cited in that path, the rate was 49.0%. When the brand also appeared through a branded follow-up search, it was 91.4%. Comparing runs inside the same project, prompt and engine with the follow-up search absent, own-domain exposure was associated with a mean mention-rate increase of 40.2 percentage points on GPT. The authors call the fitted equation predictive and observational rather than a causal description of engine internals, and so do we: these are associations in their panel, not a mechanism anyone has seen.
Read it as a priority order on the two engines it covers. Getting into the retrieval set is worth more than any on-page tweak you can make to a page that never gets retrieved. Once you are in the set, the second lever in that data is corroboration: something the engine already retrieves saying your name when it runs the follow-up query.
The shortlist is far smaller than the web
Citation volume concentrates hard, at least in the one dated public count we have found. An analytics vendor published a citation analysis of its own monitored prompts in the United States between January and February 2026, covering ChatGPT in web browsing mode and Google AI Mode. In its ChatGPT table, Wikipedia took 13.15% of citations and Reddit 11.97%, with OpenAI's own domain third at 6.21% and nothing else in the top twenty above 3%. Two domains, a quarter of the citations. That is a vendor's panel of monitored prompts in one country over two months, not a measurement of everything ChatGPT cites, and we have no second source to check it against.
That shape has a practical reading, and it is not "go and spam Reddit". It means the long tail of cited domains is thin and contested, so the slot you are competing for is usually the one specific, verifiable claim that the generalist sources cannot supply. Encyclopaedias give definitions. Forums give opinions. Neither can tell the model what your product costs in India, what your throughput was under load, or what your refund window is. Those are the sentences that get lifted.
The same logic drives the retail surfaces, where a product feed and a price are the liftable facts. We covered that separately in Shopping inside ChatGPT, and what it reads.
Four failure modes we keep finding
These are the four we run into most often in our own audit work, which reads a client's pages and runs a prompt set across several engines. They are what we see, not a ranked frequency from a counted sample, and we have not tried to measure how common each one is across the web.
The page answers late. The model is working with passages, not documents. If the answer to the heading sits four hundred words below it, the kept passage may miss it entirely. State the claim in the opening sentence beneath that heading, then explain it.
The claim is unattributable. "Industry leading uptime" cannot be cited because there is nothing to quote. "99.95% uptime measured across the 2026 financial year" can. Specificity is not a style preference here, it is what makes a sentence liftable.
The facts contradict each other across your own site. Pricing on the product page, different pricing in the help centre, a third number in a PDF. In the audits where we have found that, the answers we got tended to hedge on the number or to credit a source that sounded certain. We cannot see why the model did that, so read it as a pattern worth fixing rather than a rule about how engines behave.
The page needs JavaScript to say anything. If the words only exist after a script runs, a crawler that does not execute that script has nothing to index. Rendering the main content server side removes that dependency. It is dull work and it is often the cheapest fix an audit turns up.
A check you can run this week
List ten of the questions your buyers actually ask. Put each through ChatGPT with search on, five times, and record three things: whether you were named, whether the claim attached to your name was true, and which domains were cited instead of you. Repeat the run because the same prompt does not reliably return the same sources: the preprint above found mention rates for one prompt varying across repeated runs on the same engine. Ten prompts five times is a sample you can act on, not a measurement of your visibility.
Then split the misses. If your domain never appeared in the cited set, you have a retrieval problem, so go and read your logs for OAI-SearchBot. If your domain appeared and the sentence came from somewhere else, you have a passage problem, and the fix is in the text. Two different piles, two different owners. Google's equivalent surface behaves differently enough to deserve its own pass, which is the subject of Google AI Mode: what changes for a publisher.
All of this is the manual version of what an audit compresses, and the wider framing sits in what is answer engine optimisation.
Honest limits
This post does not establish how ChatGPT ranks or selects sources. We cannot see the ranking function, and everything above is inferred from OpenAI's published documentation, from two dated third-party datasets, and from runs we made ourselves, so it describes observed behaviour rather than code. The 2.8 / 49.0 / 91.4 staircase is an association measured on one panel of monitored prompts on GPT and Gemini between June and September 2026, and the authors say so themselves: it is not a causal claim and we do not know whether it holds on your prompts or on the engines nobody measured. The citation shares are one vendor's monitored prompts in the United States over two months, so treat them as a dated snapshot and re-measure before acting; we found no independent source to check them against. The four failure modes and the hedging we describe come from our own audit work rather than a counted sample, and we deliberately report no figure for how often each occurs because we have not measured it.
FAQ
Does blocking GPTBot stop ChatGPT citing me? Not by itself. GPTBot relates to training. The agent behind ChatGPT's search features is OAI-SearchBot, and OpenAI's documentation states that a site opted out of it will not be shown in ChatGPT search answers. That is OpenAI's rule for ChatGPT; other engines document their own.
Does structured data make ChatGPT cite me? It helps a machine read your page correctly and it resolves who you are. Treat it as removing ambiguity rather than as a ranking lever. A clean Article or FAQPage node with accurate content is worth more than a large graph full of optimistic claims.
Why do two identical prompts return different sources? Search behaviour and retrieval both vary between runs, and the index itself refreshes. The September 2026 preprint shows mention rates for the same prompt moving across repeated runs. One run is an anecdote. Several runs across a few days are a sample, which is why audit prompt sets are run repeatedly rather than once.
How does ChatGPT cite sources? In two steps. It turns your question into one or more machine-written searches and pulls a shortlist of documents from the index OAI-SearchBot builds, then, while writing, attaches a link to the passages a sentence traces back to. A page can be retrieved and never credited, and from outside the two look identical, which is why access and wording are separate problems with separate fixes.
How do I get cited by ChatGPT? Start with access. OpenAI's documentation says a site opted out of OAI-SearchBot will not be shown in ChatGPT search answers, so read robots.txt for that agent by name rather than for GPTBot, and check your logs for requests from OpenAI's published search crawler IP ranges. Then make the claim liftable: one sentence carrying the number, the unit and the date, stated directly under the heading that asks the question. Nobody outside OpenAI can promise the citation itself, because the selection step is not published.
Is ChatGPT traffic worth chasing if nobody clicks? We have no referral figure we can source, and assistant referrers are often missing or collapsed in analytics, so clicks are a poor instrument here either way. The value we can observe is the mention itself, arriving before the buyer has opened a browser tab. Measure named appearances and claim accuracy, and treat clicks as a bonus.
Sources
- OpenAI, overview of OpenAI crawlers and user agents, read 12 October 2026: developers.openai.com/api/docs/bots
- OpenAI, OAI-SearchBot published IP ranges; fetched by us 8 and 12 October 2026, 39 IPv4 prefixes both times, creationTime 2 January 2026: openai.com/searchbot.json
- From Prompt to Recommendation: A Fitted Stage Model of Brand Visibility in AI Search, arXiv preprint 2609.23162, submitted 19 September 2026; 34,960 unbranded prompt and engine observations, GPT and Gemini, June to September 2026: arxiv.org/abs/2609.23162
- Vendor citation analysis of monitored prompts in the United States, ChatGPT web browsing mode and Google AI Mode, January to February 2026, published April 2026; ChatGPT domain shares quoted above. We describe it generically and do not link it, because it is an AI visibility vendor's own blog rather than a primary source, and we have found no independent dataset that reports the same shares.
- Google Search Central, AI features and your website, read 12 October 2026: developers.google.com/search/docs/appearance/ai-features