What Are ChatGPT Citations and When Do They Appear?

Three things get mixed up in most conversations about ChatGPT visibility, and they behave differently. A mention is your brand name appearing in the answer text. A citation is a link to a specific page, attached to a sentence or passage of the answer. A source is any page ChatGPT read while searching, whether or not it ended up linked.

You can be mentioned without being cited (ChatGPT names you but links a directory listing about you), cited without being mentioned (your blog post supports a general statement and your brand never appears), or consulted without either. A useful measurement keeps the three apart, which is why Leadee's tracker reports mention rate and citation rate as separate numbers.

Citations appear when an answer was built with web search. An answer written only from what the model learned in training has nothing to link to, so it comes without citations even when it names companies. That is the single biggest reason two people can ask the same question and see one answer with sources and one without.

In OpenAI's API the difference is explicit. With the Responses API and the web_search tool, the model decides for itself whether a question needs a search. With the search model in the Chat Completions API, OpenAI says it always retrieves from the web before answering. OpenAI also says that Chat Completions search model gives developers access to the fine-tuned model and tool used by Search in ChatGPT.

Inside the consumer ChatGPT app, whether a given question triggers search depends on the question, the product mode and settings we cannot see. We do not test the consumer app — Leadee measures through OpenAI's official API only — so we cannot give you a rule for when the app searches.

OpenAI's documentation is direct about this: the sources field returns every URL the model consulted, inline citations show only the most relevant ones, and the first list is often longer. A page can be fetched, read and still lose the citation to another page that states the same fact more plainly or more specifically.

This is why crawl access is necessary but not enough. If OAI-SearchBot cannot read your page, it cannot be a source at all — see what each OpenAI bot does. Once it can, the passage itself decides: a direct one-sentence answer, a specific number and a named source give a model something it can attach a citation to. Our guide to writing pages ChatGPT can quote shows that on real pages, and how AI chooses suppliers covers the retrieve-then-write process behind it.

On 2026-10-25 Leadee measured the 7 websites of Netbase JSC, the company behind Leadee, across three engines (the full write-up is the Netbase portfolio case study). The OpenAI part used gpt-5 with the web_search tool through the official API. It stopped after 25 runs because the account ran out of credit, so the OpenAI sample is small and every rate below is an indicator, not a measurement you should plan around.

In those 25 runs, a Netbase brand was named in 3 and its own domain was cited in 3 — 12%, with a 95% Wilson interval of 4% to 30%. All three were CMSmart, on two US buyer questions about WooCommerce product designer plugins. The other six brands were neither named nor cited in any of their OpenAI runs, including Netbase JSC's own site (netbasejsc.com, Netbase-owned) in 0 of 7. Several of the same brands were named by Perplexity and Google AI Mode in that round — Netbase JSC's own site, for example, in 11 of 48 US runs on those two engines — which is one more reason to measure each engine separately.

What the data does not show: how many answers carried citations at all, and which other domains OpenAI cited. Leadee's tracker records the url_citation links of each answer to decide whether your domain was cited; the per-domain source lists in our analyses are combined across engines, so we do not publish them as "ChatGPT's sources".

Each tracker run sends one buyer question to OpenAI's Responses API with the web_search tool and reads the url_citation annotations in the answer. A run counts as "cited" when at least one of those links points to your domain, and as "mentioned" when your brand name or domain appears in the text in an unambiguous form. Rates come with a Wilson confidence interval, because with a few dozen runs the uncertainty is larger than most dashboards admit. The details are on our methodology page.

Two honest limits. First, the API is not the ChatGPT app: OpenAI ties the API's search model to Search in ChatGPT, but the app can add memory, personalisation and product settings that an API call does not have, so your own ChatGPT session can differ. Second, one answer is one sample. The same question asked again can cite different pages, which is why a single screenshot of ChatGPT citing (or ignoring) you proves very little.

The research base is thin. The GEO paper (Aggarwal et al., KDD 2024), which built a benchmark of generative-engine queries, reported that content changes could raise a source's visibility in generative answers by up to 40%, and that what works varies by domain. That is a benchmark result, not a promise for any one page, but it supports working on the passage itself rather than only on technical settings.

In practice we work in this order: make sure OAI-SearchBot can read the page, rewrite the passages that answer buyer questions so each one can stand alone, and get listed on the third-party pages ChatGPT already cites for your topic. The last step is often the biggest, and it has its own guide: how to get cited when ChatGPT prefers other websites. The full set of ChatGPT guides and tools is on the ChatGPT SEO overview.