How does ChatGPT pick the sources it cites?

Direct answer

ChatGPT builds its answer from a search of its own, and OpenAI documents that it returns inline citations along with a wider set of retrieved URLs than it ends up showing. Access to that search depends on a specific agent, OAI-SearchBot, separate from the one feeding model training. In the Observatory, 2,066 of the 2,076 ChatGPT answers in the September 2026 cut arrived with at least one cited domain.

What OpenAI documents

OpenAI documents four agents with separate jobs: OAI-SearchBot to surface sites in ChatGPT search results, GPTBot for model training, OAI-AdsBot to validate the safety of pages submitted as ads, and ChatGPT-User for user-initiated actions. About the last one it makes a point worth reading closely: ChatGPT-User does not determine whether content may appear in search¹. A visit from ChatGPT-User to a page, then, says little about that page's presence on the search surface.

On the answer side, OpenAI documents that its web search tool returns inline citations by default, plus annotations carrying the URL, the title and the location of each source, and a separate field with every URL retrieved, usually wider than the citations displayed². That distance between what is retrieved and what is cited is the measurable part of the process: a domain can enter the answer's context without ending up in a visible citation.

OpenAI also requires that, when web results are shown to a person, inline citations stay visible and clickable². The obligation falls on whoever builds the interface.

Which crawler it uses and how to verify access

The agent behind the search surface is OAI-SearchBot. Verification is a robots.txt matter and takes two steps, with no tooling.

The first is permission per agent. A `Disallow` reaching OAI-SearchBot leaves the site out of ChatGPT search results, whatever the quality of the content. Since the agents are separated by function, a site can allow OAI-SearchBot and block GPTBot: the search decision and the training decision are distinct and get written separately¹.

The second is consistency. A generic `User-agent: *` block reaches OAI-SearchBot unless a block of its own exists for that agent, so a site that believes it has allowed search may be closing it from a general rule written earlier.

What my own data shows for ChatGPT

In the Observatory of brands in AI, which queries the real ChatGPT product, 2,066 of the model's 2,076 answers arrived with at least one cited domain in the September 2026 cut. Source exposure, on that surface and with that collection method, is close to total.

The contrast with my own weekly benchmark measurement is the part worth understanding. The benchmark queries the model API and stores the URLs the answer text writes out. Over the four-week window closed on September 21, 2026, the count of answers carrying a URL in the text stayed at zero across 5,893 ChatGPT answers through the API.

The two figures describe the same model and measure different things. The first measures what the product shows the public. The second measures what the model writes in its prose when queried through the API without search. Any report about ChatGPT sources has to say which of the two surfaces it measured, and that holds for third-party reports too.

According to a study by Profound, a company that sells AI visibility analytics, across 680 million citations between August 2024 and June 2025, Wikipedia held 7.8% of ChatGPT's citations and 47.9% of its top 10 sources³. According to a study by Semrush, a company that sells SEO software, across more than 230,000 prompts between July and October 2025, the rates of Reddit and Wikipedia in ChatGPT moved sharply week to week⁴. Both panels are their own and incomparable between them.

How I measure it

I run a bank of unbranded prompts over the category, repeated several times, and I record two series per answer: which brands get named and which domains get cited. For ChatGPT I measure on the product, which is where the sources appear, and I record the collection method next to every figure.

Before the first run I do the access check from the previous section, once, and I repeat it whenever robots.txt changes.

I always compare inside the same model and the same window. An average across models hides the case of a brand that appears often in one and vanishes in another, and a comparison across different windows mixes site changes with engine changes.

How to improve your chances of ChatGPT citing you

1. Check that OAI-SearchBot can reach your site. It is the search surface's agent, separate from the training one, so you can allow one and block the other according to what you decide.

2. Look at whether a general rule in your robots.txt is catching it without meaning to. A wildcard block written years ago reaches OAI-SearchBot unless a block of its own exists for it.

3. Measure on the real ChatGPT product and leave the API for something else. It suits you because the sources show up in the product, and an API measurement without search describes another surface.

4. Keep the collection method next to every figure you publish. Without it, two measurements of the same site stop comparing with each other and the series loses its point.

5. Work on the pages that answer category questions, with the answer up top. That is what a short intermediate search can retrieve and show without having to read your whole site.

Which AuraMetrics modules address this

Measure the brand with the same category prompts. Prompt Tracking runs prompts on the five engines.

See which sources AI cites when it answers about the category. Prompt Tracking shows, for each prompt and model, the sources the answer cited, and AI Market Leaders groups them by brand.

Review the site's technical signals with the GEO Score and the GEO Framework. The GEO Score scores the site from 0 to 100 on the four pillars of the GEO Framework.

Frequently asked questions

What does a tool need in order to track brand mentions across ChatGPT, Gemini, Claude and Perplexity?

Tracking brand mentions across ChatGPT search takes three pieces: a fixed bank of prompts, repetition per model, and the collection method recorded next to every figure. Without those three, two runs of the same site stay uncomparable.

How should a brand visibility alert for AI answers work?

A measurement that holds compares the same brand against its own earlier runs, inside the same model and the same window. Comparing across engines or across different windows produces readings that describe the method change before the site change.

For how often a brand appears in ChatGPT, with its methods and options, there is the guide on measuring brand frequency in ChatGPT.

How I measured it

Observatory of brands in AI: a fixed bank of unbranded prompts across 35 categories in Uruguay, three repetitions per model, querying the real ChatGPT product. September 2026 cut, read on September 24, 2026: 2,076 ChatGPT answers, 2,066 with at least one cited domain. The Observatory has run three cuts; the figures in this guide come from the latest.

Weekly benchmark: a four-week window closed on September 21, 2026, 5,893 ChatGPT answers through the API. URLs are extracted from the answer text with a regular expression, so the count describes what the model wrote in its prose.