How ChatGPT sources the web
ChatGPT is the answer engine publishers ask about most, and the one that documents itself least. OpenAI publishes what its crawlers do, and not what sits in the index behind ChatGPT search, how sources are ranked, or what makes the model search at all. This page keeps two layers apart: what OpenAI documents, and what outside researchers have measured. The second layer is observation of output rather than description of the system, and is labeled that way throughout.
What OpenAI documents
OpenAI documents three fetchers, and they do different jobs.
- GPTBot crawls content that may be used in training OpenAI's generative AI foundation models. Disallowing it indicates that a site's content should not be used for that training.
- OAI-SearchBot surfaces websites in ChatGPT's search features. OpenAI states that sites opted out of OAI-SearchBot "will not be shown in ChatGPT search answers, though can still appear as navigational links."
- ChatGPT-User fetches a page when a user action in ChatGPT or a custom GPT requires it. OpenAI states it is not used to crawl the web automatically, and does not determine whether content appears in Search.
That yields one documented lever. Allowing OAI-SearchBot is the stated precondition for appearing in ChatGPT search answers, and blocking it is the stated way out. Blocking GPTBot is a training decision with no documented effect on answers. The trade-offs are in AI crawlers.
OpenAI also documents that answers built on web search carry citations. Its web search guide states that web search lets models "provide answers with sourced citations", that responses include inline citations by default, and that a separate sources field lists every URL consulted, usually more than it cites. That document covers the API tool rather than the consumer product.
What OpenAI does not document is the part everyone wants. No OpenAI document describes the composition of the index behind ChatGPT search, names an index provider, explains how candidates are ranked within it, or offers publishers a citation report. The per-engine comparison records the same gap. Claims that ChatGPT search runs on a particular third-party index are unconfirmed by OpenAI, and this page does not make one.
Model knowledge and retrieval are not the same thing
Most confusion about ChatGPT visibility comes from treating every answer as a retrieval event. Some answers come entirely from training data, with no fetch and no citation. Others trigger a search and are grounded in pages retrieved in that moment. The same question, asked twice, can be answered both ways.
OpenAI's documentation describes three modes: a non-reasoning search that relays the tool's result, an agentic search where a reasoning model searches inside its chain of thought and decides whether to keep going, and deep research across hundreds of sources. What is documented nowhere is the condition that triggers a search at all. There is no published rule, threshold or query class.
The consequence is practical. An answer that omits your brand may mean your page was not retrieved, or that nothing was retrieved at all. One check cannot tell those apart.
What outside research observes
Everything in this section is third-party observation of ChatGPT's output. None of it is documentation, none of it is confirmed by OpenAI, and all of it can change without notice.
Retrieval concentrates on the first turn
An outside analysis of roughly 730,000 ChatGPT conversations containing at least one web citation, from United States English-language users between October and December 2025, observed that about 18 percent of conversations triggered a web search, and that those searches cluster at the start. In that sample 12.6 percent of first turns carried citations, falling to 4.5 percent by turn 10 and 3.0 percent by turn 20. A measured pattern in one sample, not a documented rule.
Wikipedia sits at the base of the observed citation layer
The same analysis observed Wikipedia accounting for about 5 percent of all citations and appearing in 18 percent of conversations that carried any citation, ahead of Reddit at 3 percent and 13 percent. It observed roughly six unique citations per cited conversation, with the top ten domains taking only 12 percent of all citations. Observed, not documented. General reference material looks like a common baseline, and one citation is not ownership of an answer.
Whether ChatGPT's citations track an existing search index is unsettled
Two published studies measured this and disagree, and the disagreement is the finding. One analysis of about 100 queries and more than 500 citations, published in February 2025, found that 87 percent of ChatGPT search citations matched Bing's top organic results when the user's question was run verbatim as the query, against a 56 percent match in Google at a median rank of 17. A study of 15,000 long-tail prompts found a different shape: 8 percent of ChatGPT's in-text citations and 6.1 percent of its reference citations ranked in Google's top ten for the original prompt.
The two are not comparable. One ran short questions verbatim as queries across a small sample; the other used long-tail prompts, which an engine is likely to rewrite before searching. Neither establishes which index backs ChatGPT search, and neither could, because both measure output. Treat the question as open.
What a publisher can actually do
- Be fetchable by OAI-SearchBot. This is the one documented requirement. Check robots.txt, and check that your edge is not blocking the crawler, per AI crawlers.
- Be readable without JavaScript execution. No AI fetcher documents a rendering guarantee, so content that exists only after client-side rendering is an unverifiable risk. See rendering.
- Write self-contained passages. A retrieved passage is quoted away from its context. See chunking.
- Keep facts current. Updating an existing page beats publishing another one. See freshness.
How to measure your standing honestly
OpenAI publishes no citation report, so there is no first-party number to read. What remains is sampling your own output: a fixed set of prompts, run repeatedly, recorded over time, as described in measuring AI search visibility honestly and prompt sets. Because retrieval is not guaranteed on any answer, a single result carries almost no information. A distribution across many runs carries some.
What stays unanswerable
Four questions cannot be answered from outside: what is in the index behind ChatGPT search, how candidates are ranked within it, what makes the model search on one turn and not the next, and how much of an answer came from training data rather than retrieval. Outside studies measure output. They do not see the system. Anyone selling a ChatGPT ranking factor is describing a correlation in a sample, and should say so.