How ChatGPT Search Works: Retrieval, Reading, Citation

ClickRadius Institute · Published April 5, 2026

ChatGPT Search can return a composed answer with source links, but OpenAI does not publish a complete candidate-ranking and generation trace for each answer. What it does publish is a useful access-control boundary: OAI-SearchBot is the automatic Search crawler, GPTBot is the training crawler, and ChatGPT-User represents certain user-initiated actions. This guide separates those roles from the citations a user can actually observe, without presenting a private retrieval pipeline as established fact.

Published interfaces, not a disclosed ranking pipeline

OpenAI says public sites can appear in ChatGPT search and that Search may display summaries, snippets, citations, and links. Those are observable output surfaces. The provider documentation does not expose every query transformation, candidate page, fetch, weighting, or generation decision behind a particular answer.

That distinction matters. A cited link proves that the captured answer displayed the link. A verified crawler request proves that an identified agent requested a URL. Neither observation alone proves every document considered, the reason a source was selected, or a general rule that will govern another prompt.

Measure the event you actually have: crawler access, a user-triggered request, a displayed link, a mention, a citation, a referral, and a conversion are related possibilities, not interchangeable evidence.

— ClickRadius Institute

Stage one: the access controls OpenAI documents

OpenAI’s crawler documentation assigns a distinct purpose and control boundary to each named agent:

• OAI-SearchBot: the automatic crawler used to surface websites in ChatGPT search results. OpenAI recommends allowing it in robots.txt and permitting its published IP ranges when Search appearance is wanted. An opted-out page is not shown in ChatGPT search answers, although a navigational link may still appear.
• GPTBot: crawls content that may be used to train OpenAI’s generative foundation models. Its setting is independent of OAI-SearchBot, so a publisher can allow Search crawling while signaling a training opt-out.
• ChatGPT-User: may visit a web page for certain actions initiated by a ChatGPT or Custom GPT user. OpenAI says it is not automatic web crawling, is not used to determine Search appearance, and may not be governed by robots.txt.

A server log should therefore retain the exact user agent, requested URL, status, timestamp, and verified source network. Do not collapse OAI-SearchBot and ChatGPT-User into one “retrieval crawler” metric: their triggers, controls, and evidentiary meaning differ.

Stage two: what the returned answer does not disclose

A returned answer can be saved and audited, but it does not expose the complete candidate set or a provider trace showing every page fetched, passage selected, weight applied, or disagreement reconciled. Treat explanations of those private steps as hypotheses unless OpenAI documents the exact mode or a trace accompanies the observation.

Evidence remains useful without inventing a ranking rule

Statistics, attributed quotations, and source citations can make a page easier for any reader to verify when they are accurate and relevant. The Princeton-led GEO experiment reported selected-source visibility changes in its benchmark, including an upper-bound result near 40% for some settings. It did not establish ChatGPT Search’s current ranking or citation rule. The benchmark began with the top five Google sources, randomly selected one source for optimization, and measured that source’s relative visibility/impression in generated responses; it did not measure retrieval or citation-inclusion probability, and reported effects varied by domain.

Keep entity evidence source-scoped

Business, person, and product records can be audited across first-party pages, directories, profiles, and third-party coverage. Preserve source ownership, exact subject, fact scope, public visibility, date, and conflicts. Those records do not reveal how ChatGPT resolved an entity or why it appeared in an answer.

Stage three: citation

When ChatGPT Search displays citations or source links, record the prompt, answer, link target, account and locale context where available, and collection time. That receipt establishes the displayed output for that run. It does not establish how many uncited pages were consulted or why one link appeared instead of another.

Ranked position, mention, quotation, linked citation, referral, and click remain separate observations. State material facts plainly and support them so a reader can verify them, then measure what ChatGPT actually displays. The Princeton study did not establish a current ChatGPT citation formula.

In a benchmark beginning with the top five Google sources, researchers randomly selected one source for optimization. Quotations, statistics, and source citations produced reported relative visibility/impression gains of up to roughly 40% in some settings. The experiment did not measure retrieval or citation-inclusion probability, and effects varied by domain.— Aggarwal et al., "GEO: Generative Engine Optimization," Princeton (KDD 2024)

How this differs from a classic search result

It helps to see the two models side by side.

1. The output. A conventional search result and a ChatGPT search answer expose different layouts. Capture the actual links, citations, and wording rather than converting either layout into a promised click or citation rate.
2. The controls. Search indexing, model-training consent, and user-triggered page access have separate OpenAI agents. A single allow or disallow statement cannot describe all three.
3. The evidence. A search-crawler request, user-triggered request, displayed citation, and referral answer different questions. Preserve them as separate event types.
4. The limits. None of those events reveals the full ranking and generation process, and none guarantees the same result on a later prompt.

What this means for your site, in practice

The pipeline turns into a short, unglamorous checklist. None of it is a trick; all of it is making yourself legible to a reader.

• Set each access policy deliberately. Allow OAI-SearchBot if you want automatic ChatGPT Search crawling, make a separate GPTBot training choice, and do not present robots.txt as a reliable ChatGPT-User control.
• State material facts in accessible text. Record who the facts describe, what they establish, their scope, source, and effective date. This makes the public record inspectable; it does not establish whether ChatGPT will fetch, select, or cite it.
• Publish attributable evidence. Operational statistics still need a method, sample, period, owner, and limitations; quotations need accountable attribution; citations must support the exact claim.
• Answer clearly. Use descriptive headings and complete answers with necessary context. No required sentence count or desired quotation proves extraction.
• Reconcile entity records. Compare the site, profiles, directories, and third-party mentions by controller, entity, fact scope, and date. Agreement need not be character-identical and does not prove ChatGPT use.
• Verify per supported engine. Use the monthly AI Presence protocol across ChatGPT, Gemini, Perplexity, Claude, and Grok, preserving run-to-run variation and avoiding causal conclusions from competitor citations.

The honest limits of what anyone can promise

The internal selection process is not fully observable, and a provider can change it. Nobody can guarantee a ChatGPT citation. The defensible posture is to follow the current published access controls, maintain accurate and inspectable content, and preserve repeated answer observations under a fixed protocol. A change after an edit is temporal evidence, not causal proof that the edit produced it.

No market-wide AI-answer prevalence is established here. Measure the business on a fixed, dated prompt set and record the engines, locale, device or account state where relevant, and whether each result is a mention, linked citation, or recommendation. That sample can identify an observed gap; it does not establish that a category is empty, a citation position is unclaimed, or an early entrant will become a default answer.

Frequently asked questions

How does ChatGPT Search find pages to answer a question?

OpenAI documents OAI-SearchBot as its automatic search crawler: allowing it helps content appear in ChatGPT search results, while opting out prevents the content from appearing in search answers, apart from possible navigational links. ChatGPT-User is different. It may visit a page for certain user actions, is not an automatic crawler, does not determine Search appearance, and robots.txt rules may not apply to those user-initiated requests.

Why does ChatGPT cite some sources and not others?

A returned ChatGPT answer can establish which links or citations were displayed for that prompt, account state, locale, and time. It does not expose every candidate, page fetched, or ranking and generation decision. Princeton’s GEO study began with the top five Google sources, randomly selected one source for optimization, and measured that selected source’s relative visibility or impression in generated responses. It did not measure ChatGPT Search retrieval or citation-inclusion probability, and reported effects varied by domain.

Can I control whether ChatGPT uses my site?

Partly, and each agent has a separate purpose. GPTBot is the training control. OAI-SearchBot is the robots.txt control for automatic Search crawling and appearance in ChatGPT search answers. ChatGPT-User represents certain user-initiated actions; it is not automatic crawling, is not used to decide Search appearance, and robots.txt rules may not apply. None of these controls guarantees a citation.

Inspect five families of readiness evidence at the submitted URL with a free AI Readiness assessment, or see ClickRadius plans for separate monthly AI Presence samples from five supported engines. The score does not simulate ChatGPT or establish how it reads the site.