Log inGet Started

Tracked on Growth and Scale

Vendor documentation checked: 09.08.2026

How ChatGPT chooses and cites sources

ChatGPT answers in prose and sometimes links what it used. This page separates what OpenAI documents about that from what nobody outside OpenAI knows — and states what we measure on the surface.

What this surface is

ChatGPT is a conversational assistant. For questions that depend on current information it can run a search step first and then write an answer in its own words, attaching links to some of what it used. For questions it can answer from training data, it may not search at all.

From a brand's point of view that is two systems in one: a model that has read a version of the web, and a retrieval step that reads the live web at the moment of the question. A brand can be named by the first and cited by neither — or cited without ever being recommended.

How it differs from a results page

A results page returns a ranked list and lets the reader choose. ChatGPT returns one composed answer, and the sources are supporting material rather than the product. There is no position one to win.

The output is also not stable. The model decides per turn whether to search and what to search for, so the same question asked twice can produce a different answer with different links. Reading a single conversation as a ranking is the most common mistake on this surface.

What the vendor documents

Each line below is read from the linked document. Nothing in this section comes from our own testing or from a third-party estimate.

What is not documented

These are open questions, not omissions. Anyone answering them with a precise number is guessing.

  • OpenAI does not publish how retrieved pages are ranked, or which of them end up cited. We do not know the weighting, and we will not invent one.
  • Whether a search happens at all is a model decision made per turn. That makes any single answer a sample of behaviour rather than a measurement of standing.
  • We cannot see which ChatGPT surface a given person used — a logged-in chat, a custom GPT, an app integration — and those need not behave identically. What we measure is one consistent path, described below.

What GetCited.me measures here

We ask ChatGPT real buyer questions on a schedule and read the answer: whether your brand is named, which source URLs the answer exposed, which competitors appeared next to you, and how you were characterised. The questions are the ones your customers would type, run with the country and locale settings of your market.

The headline figure is the organic mention rate — questions that never name your brand. Questions that already contain your name are reported separately as recognition and as head-to-head comparisons, and are never blended into the headline, because a brand a model knows perfectly can still lose every comparison.

ChatGPT tracking is part of the Growth and Scale plans; Launch (Free) tracks Gemini only. Card payments are temporarily disabled while payment-provider onboarding completes, so paid plans cannot be purchased right now and every account runs on Launch.

What you can actually do

Concrete steps, each one checkable with a free tool or explained in the glossary.

Decide what OAI-SearchBot may readSearch access and training access are separate rules in the same file. Read what your robots.txt currently says to each OpenAI agent before assuming either one is allowed.Answer the question on the pageA retrieval step can only use what is on the page. Check whether yours states the answer plainly enough to be lifted out, or circles it for three paragraphs first.Make your name unambiguousIf an assistant confuses you with a similarly named product, more content will not fix it. Entity disambiguation is the discipline that does.Be present where the answer looksAnswers often cite third-party pages rather than a vendor's own site. Community threads and reference pages you do not own can carry your brand into an answer.

Free tools for this surface

No account, no card — results in the browser.

Related terms

Answer engineAn answer engine is a system or product surface that generates a synthesized response to a query instead of returning only a list of links.AI citationAn AI citation is a source reference or link exposed with an AI-generated answer; it is not the same as an unlinked brand mention.robots.txt (for AI crawlers)robots.txt controls which compliant crawlers may fetch specified URL paths; different AI and search uses may rely on different crawler tokens.Citation-ready contentCitation-ready content is accurate, accessible content whose claims, structure, and sources are easy to inspect and reuse.

Frequently Asked Questions

No one can promise that, and you should be careful with anyone who does. You can make the site eligible and easy to use — allow the search crawler, answer the question on the page, keep your identity unambiguous — and then measure whether it happens. OpenAI does not publish the selection rules.
No. OpenAI documents GPTBot as the training crawler and OAI-SearchBot as the agent that surfaces sites in ChatGPT's search features, and states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers. They are separate rules and can be set separately.
Because the model decides per turn whether to search and what to search for, and the live web changes underneath it. That is why we track the same questions repeatedly over time instead of reading one conversation.
No. Launch (Free) tracks Gemini only; ChatGPT, Claude, Perplexity and Google AI Overviews are on Growth and Scale. Card payments are temporarily disabled during our beta, so paid plans cannot be purchased at the moment.

See how AI answers about your market

Free scan: market map, AI visibility snapshot, and a GEO score in about a minute. No card, no call.

Scan your site free