How ChatGPT Chooses Which Sites to Cite

ChatGPT retrieves before it writes. It queries an external index, fetches candidate pages through OAI-SearchBot, and lifts a self-contained passage from one of them. A site absent from that index is never considered, whatever its content quality. This page explains the retrieval path, the two OpenAI crawlers, and what makes a passage quotable.

See pricing

Five questions decide a citation. Each one is answered below, in the order retrieval resolves them.

Retrieval order — the stages ChatGPT runs before writing a sentence
Index coverage — where the candidate pages come from
Crawler identity — which OpenAI agent fetches a page for citation
Render dependency — why browser-assembled content arrives empty
Passage shape — the four properties that make text liftable

How ChatGPT chooses which sites to cite

ChatGPT runs a retrieval step before writing. It queries an external index, fetches candidate pages, and extracts a passage from one. Sources absent from that index are never considered, regardless of content quality.

Retrieval is the lookup step: the assistant turns a prompt into search queries and receives a ranked candidate set. Extraction is the lift step: the model takes from a fetched page the sentences it will quote and link.

Selection resolves in three stages, in this order: retrievability, extractability, corroboration. Retrievability admits a page to the candidate set. Extractability decides whether a passage survives the lift. Corroboration ranks competing passages, because agreement across independent sources outranks a lone assertion. A page failing stage one never reaches stages two and three.

The retrieval path, stage by stage

prompt → search query → index lookup → candidate set → OAI-SearchBot fetch → passage extraction → cited answer

ChatGPT queries the index, fetches candidates with OAI-SearchBot, then extracts one passage. A page joins that chain at the index lookup or nowhere.

Which index ChatGPT actually searches

ChatGPT search retrieves through Microsoft's index and OpenAI's own crawl. A site missing from both is unreachable on this surface. Bing Webmaster Tools registration is free and takes about ten minutes.

Microsoft supplied the original index behind ChatGPT search, announced by Satya Nadella at Build 2023: "We're bringing search grounding and Bing to ChatGPT." Bing still grounds part of that surface, so a Bing entry is cheap insurance.

OpenAI has since built its own retrieval. Nick Turley, ChatGPT's head of product, testified in April 2025, in the remedies phase of the US antitrust case against Google, that OpenAI started its own search index, aiming at 80 percent of ChatGPT searches. OAI-SearchBot fills it. Two retrieval paths feed one answer: Microsoft's index and OpenAI's own crawl.

Bing Webmaster Tools carries no fee. The AI Performance report Microsoft opened there in February 2026 logs citations across three surfaces: Copilot, Bing's AI summaries, and partner integrations.

Index prerequisite

Search site:yourdomain.com in Bing; zero results means no coverage.
Register the domain at Bing Webmaster Tools.
Verify ownership by one of four methods: XML file, meta tag, CNAME record, Domain Connect.
Submit the XML sitemap.

OAI-SearchBot and GPTBot are not the same crawler

OpenAI operates two crawlers with separate purposes. OAI-SearchBot fetches pages for retrieval and citation in live answers. GPTBot collects training data for future models. Blocking GPTBot preserves citation eligibility; blocking OAI-SearchBot removes it entirely.

OpenAI publishes both tokens and both purposes: OAI-SearchBot "to surface websites in search results in ChatGPT's search features", GPTBot "to make our generative AI foundation models more useful and safe".

Each setting is independent of the others, in OpenAI's own words, and OpenAI names the case: a webmaster allows OAI-SearchBot in order to appear in search results while disallowing GPTBot. Sites opted out of OAI-SearchBot are not shown in ChatGPT search answers, though they still appear as navigational links.

The two OpenAI crawlers, side by side

OAI-SearchBot and GPTBot are not the same crawler
Crawler	Purpose, in OpenAI's wording	Effect of disallowing it
`OAI-SearchBot/1.4`	Surface websites in ChatGPT search results	Not shown in ChatGPT search answers
`GPTBot/1.4`	Make foundation models more useful and safe	Content not used to train those models

OAI-SearchBot fetches pages for citation. GPTBot collects training data. Published strings end OAI-SearchBot/1.4; +https://openai.com/searchbot and GPTBot/1.4; +https://openai.com/gptbot, and OpenAI lists matching IP ranges at openai.com/searchbot.json and openai.com/gptbot.json.

Why JavaScript-built content is invisible here

AI retrieval crawlers do not reliably execute JavaScript. Content assembled in the browser arrives empty. A page that looks complete in a browser and empty with scripts disabled is invisible to this surface.

Rendering is the step where a browser executes JavaScript and assembles the document. Googlebot operates a rendering service; the AI retrieval crawlers measured to date do not. Vercel and MERJ instrumented server logs on nextjs.org and reported in December 2024 that none of the major AI crawlers render JavaScript, with GPTBot, OAI-SearchBot and ChatGPT-User in the measured set.

The consequence is binary. A crawler reads the first HTML response and stores nothing beyond it. Markup a framework injects after page load is absent from that response, so the page enters the index as a shell around a loading state.

The scripts-disabled test

Disable JavaScript in the browser.
Reload the page and read what remains.
Compare that text against the intended citation passage.
Move missing sentences into the server-rendered HTML.

Text surviving step two is the text available for citation.

What makes a passage quotable

A quotable passage states one fact, stands alone without surrounding context, contains no pronouns referring backward, and carries a number where a number applies. Passages depending on the previous paragraph cannot be lifted.

Extraction lifts a span, not a document. The model selects a contiguous passage that answers the prompt without further reading, attaches the source URL to that span, and discards the remainder of the page. A passage whose meaning arrives from the paragraph above it loses that meaning at the moment of the lift, so the model drops it and takes a competitor's instead.

The four properties of a liftable passage

Singularity
one claim per passage, with any second claim moved into its own sentence.
Self-containment
the subject named inside the passage, never carried in from the heading above it.
Referential closure
no this, that, these or it pointing back at earlier text.
Numeric specificity
the figure written out wherever a figure exists, because a number is the unit a second source corroborates.

Quotability decides nothing while the fetch is refused. OAI-SearchBot and GPTBot are governed separately in robots.txt, and a page blocking the first leaves citation eligibility closed regardless of how extractable its passages are. Sites already failing that test are found by the 40-check examination.

Questions

Common questions

Every answer below ships in the raw HTML, so an assistant reading this page without running a script still receives it.

Does ChatGPT use Google or Bing?

Bing grounds part of ChatGPT search; OpenAI's own crawl carries the rest. Google's index sits outside that set: Turley testified that OpenAI asked Google for search access and was refused. Gemini resolves sources on a separate path, set out in how google gemini grounds its answers.

Does blocking GPTBot stop ChatGPT from citing a site?

No. OpenAI states that each crawler setting is independent. A robots.txt block on GPTBot removes the domain from training data and leaves citation eligibility intact, because citation runs through OAI-SearchBot.

How does a business get mentioned in ChatGPT?

Three prerequisites run in order: an index entry, server-rendered HTML, and a self-contained passage carrying the fact. Each one gates the next, and the three together open the door to structural GEO work.

Do other assistants select sources the same way?

No. Each assistant runs its own index, crawler and ranking. The path described here applies to ChatGPT, and the comparable mechanism on xAI's assistant is set out in how grok selects sources.

Sources

OpenAI, Overview of OpenAI Crawlers — user-agent tokens, strings, stated purposes, and the independence of each robots.txt setting. https://developers.openai.com/api/docs/bots
Satya Nadella, Microsoft Build 2023 keynote — "We're bringing search grounding and Bing to ChatGPT."
Nick Turley, testimony in the remedies phase of United States v. Google, April 2025 — OpenAI's own search index, the 80 percent target, and the refused request for Google search access.
Vercel and MERJ, The rise of the AI crawler, 17 December 2024 — server-log measurement finding that none of the major AI crawlers render JavaScript. https://vercel.com/blog/the-rise-of-the-ai-crawler
Microsoft Bing Webmaster Blog, Introducing AI Performance in Bing Webmaster Tools Public Preview, February 2026 — citation reporting across Microsoft Copilot, Bing AI summaries and partner integrations.

Next step

Find out which assistants name you today.

The audit runs 40 checks across five categories and queries seven assistants for your citation baseline. Findings in ten business days.