ChatGPT Source Links: Documented Access and Observed Answers

Published April 22, 2026 · ClickRadius Institute

OpenAI publishes crawler roles and controls, but it does not publish a complete rulebook for why one source link appears in a ChatGPT answer and another does not. A sound audit therefore separates documented access policy, verified requests, and sampled answer output. It does not turn any one of those observations into a private ranking rule or a promised result window.

OpenAI source-link evidence boundary. A crawler-policy result, a verified edge request, and a sampled answer with a source link are three different observations. None proves the next one. OpenAI's public material reviewed here does not publish source-ranking weights, an internal routing rule, an external-index dependency, or a recrawl-to-link timeline. No eligible closed ClickRadius customer-month cohort establishes a typical ChatGPT source-link outcome.

What OpenAI's crawler documentation establishes

OpenAI's current crawler documentation describes three independently controlled roles:

• OAI-SearchBot: OpenAI documents OAI-SearchBot as the automatic crawler used to surface sites in ChatGPT Search. OpenAI says a site that opts out will not be shown in Search answers, although a navigational link may still appear. Permission removes that documented policy barrier; it does not prove crawling, selection, or a source link.
• GPTBot: OpenAI documents GPTBot for content that may be used to train its generative models. The GPTBot setting is separate from OAI-SearchBot. A policy decision about GPTBot is not a Search setting and does not establish whether previously acquired content was used.
• ChatGPT-User: OpenAI documents ChatGPT-User for some user-initiated actions and says robots.txt may not apply to those requests. It is not an automatic crawler and is not used to determine whether a site appears in ChatGPT Search.

These are access-purpose statements. They do not expose ChatGPT's candidate generation, ranking, passage selection, answer composition, or source-link thresholds.

Three evidence layers that must stay separate

1. Published policy: evaluating the live robots.txt shows which rules apply to a token at a specific path. It does not show whether a provider complied, whether a CDN allowed the request, or whether the URL entered an index.
2. Verified request: a request matched to a provider's current network evidence shows that the identified agent requested a URL at a recorded time and received the recorded response. A copied user-agent string alone is not identity proof. The request does not prove indexing or answer use.
3. Sampled answer: a preserved answer and its source links show what one product returned for one prompt, account state, locale, model or product label, and time. It does not reveal the full candidate set, a stable ranking factor, or what another user saw.

Access policy, transport evidence, and answer output are adjacent measurements, not interchangeable proof.— ClickRadius Institute

What a source link does and does not prove

A source link proves that the recorded answer displayed that link. Preserve the complete answer because a bare URL loses the surrounding claim, qualifications, and ordering. Also preserve whether the response visibly used Search and the exact product label exposed by the interface or API.

The link does not prove that the same page ranked first in an external index, that its schema or wording caused selection, that a crawler visit caused the answer, or that the link will recur. Those are separate hypotheses requiring separate evidence. The absence of a source link is equally limited: it does not identify a training-memory answer, a crawl failure, an indexing failure, or a ranking loss by itself.

Why there is no responsible trigger list

It is reasonable to record whether a sampled answer visibly searched and whether it displayed sources. It is not responsible to convert examples such as current, local, comparison, or recommendation prompts into guaranteed routing triggers. OpenAI's documentation reviewed for this article does not publish a deterministic search-switch rule. Product behavior can also differ by interface, model, account state, locale, and date.

The same limit applies to claims about Google, Bing, licensed publishers, or a fixed number of source slots. Without current primary evidence for a specific dependency or preference, the honest status is unknown. Observing a domain in one answer does not establish how it entered the candidate set.

A ChatGPT source-link measurement sequence

1. Record the policy you actually serve. Evaluate OAI-SearchBot and GPTBot independently at the tested URLs. Do not add a named Allow: / group without copying every private-path restriction you still need, because a matching named group does not inherit the wildcard group's rules.
2. Check every delivery layer. A permissive robots.txt does not establish CDN, WAF, application, TLS, or origin reachability. Record the public response and the layer that produced it.
3. Verify request identity before labeling logs. Match network evidence using the provider's current method, then retain token, URL, time, status, and response layer. Report unverified tokens as unverified.
4. Freeze the answer sample. Store the prompt, returned answer, visible source links, engine, model or product label, locale and account state when available, and timestamp.
5. Compare like with like. Re-run the same saved prompt set on a declared cadence. Describe movement in that sample; do not attribute it to a shipped change without a design that supports attribution.

How ClickRadius bounds the measurement

ClickRadius's current workload policy supports up to 25 saved queries per site in a monthly run across ChatGPT, Gemini, Perplexity, Claude, and Grok. Microsoft Copilot is not currently supported. Each stored response is a dated sample, not continuous observation and not proof that every user received the same answer.

Publisher-controlled work can make facts clearer, keep first-party pages internally consistent, and remove measured access defects. Those are auditable outputs. They are not evidence that OpenAI adopted a fact, selected a passage, displayed a source link, or did so within a fixed number of days. No market-wide answer prevalence or provider-selection formula is established here.

Frequently asked questions

Should I block GPTBot in robots.txt?

That is a content-use policy choice, not a citation setting. OpenAI documents GPTBot for potential training use and OAI-SearchBot for automatic Search crawling as independent controls. Changing GPTBot does not change the OAI-SearchBot setting, retract previously acquired data, or prove how a future answer will describe the business.

Why does ChatGPT describe my business incorrectly?

One sampled answer does not identify its source or cause. Preserve the exact prompt, answer, source links, model or product label, account state, locale, and time. Correct first-party facts you control and separately reconcile external records, but do not label the result a training-memory or live-search defect without evidence of that pathway.

Does ChatGPT use Google or Bing rankings to choose source links?

The OpenAI crawler documentation reviewed here does not publish a Google- or Bing-ranking dependency or a source-selection formula. A returned source link shows only what appeared in that sampled answer. Do not infer an external index dependency from the link without separate, current primary evidence.

The free AI Readiness Score estimates five weighted families of inputs observed at the submitted URL; it does not prove that ChatGPT can retrieve or cite the site. ClickRadius plans add a separate monthly sample across ChatGPT, Gemini, Perplexity, Claude, and Grok.