Skip to content

How ChatGPT chooses sources and what you can verify

Understand what OpenAI documents about search access and citations, what a single answer can show, and how to inspect sources without inventing ranking rules.

Shubham Bansal, founder of The Cited Club

· Founder

Published Updated 5 min readAgent Markdown

  • #Core-GEO
  • #AI-search
  • #Strategy
A reader inspects which sources support an answer.

ChatGPT can use web search to find information and link sources in its answer. Those visible links show what was associated with that response; they do not disclose a complete source-ranking formula. An accessible page may still be absent, and a citation does not automatically mean the business was recommended.

Start by recording the product, exact question, date and collection conditions. Follow an important claim to its linked page and check whether the source supports it, including any limits. Distinguish search discovery, citation and recommendation in your report.

Improve inaccurate, missing or inaccessible evidence that you can verify, then observe later answers. You can inspect these records; you cannot infer universal selection rules from one response.

Start with the product you actually observed

Record whether the result came from the ChatGPT app or an API-based tool. Note the date, exact question, market and search behaviour where you can observe it.

OpenAI's API documentation describes web search outputs and citations for that tool. It is a source for understanding the documented API behaviour, not proof that a monitoring tool reproduces every consumer ChatGPT answer. OpenAI web search documentation

If you cannot establish the collection route, keep that uncertainty. Do not choose a familiar product label simply because it makes a report easier to market.

Search discovery is one part of the picture

OpenAI documents separate search and training crawlers. OAI-SearchBot concerns discovery in search; GPTBot concerns potential training use. Those controls do not disclose a complete source-ranking formula. OpenAI crawler documentation

An access check can establish that a rule permits the relevant crawler, or that a request succeeded in the records inspected. It cannot establish that the page will be selected for a buyer's question.

If important information is inaccessible or missing, you have a concrete website issue to fix. Once it is available, the next question is whether it addresses the decision and supports the claims at hand.

What you can check—and what remains a selection outcome
AccessInspect whether the relevant public page can be reached.
Recorded answerSave the product, question, answer and visible source links.
Claim supportRead whether the linked source supports the attached statement.

This is an inspection workflow, not a model of hidden ranking rules. Access does not guarantee selection or citation.

Retrieval and citation answer different questions

A record that a page was found or opened is evidence of retrieval or access under the recorded conditions. A linked citation is evidence that the response associated a source with its answer. Neither automatically establishes a positive business recommendation.

Do not call every linked domain a company the system recommended. A help document can support a technical explanation while the answer recommends a different provider. Equally, a company may be named without a direct link to its website.

This distinction changes what you should improve. A weakly supported factual statement needs evidence review. An irrelevant shortlist needs question and fit review. A wrong company description needs source and accuracy review.

Follow one claim to one source

Take a material statement from the answer and open the associated link. Read the relevant passage in context. Does it say what the response says it says?

Check whether the claim concerns a current service, a historical project, an operating condition or a general category. A company page supports the statement that the company makes a claim. Independent verification requires an appropriate separate record.

For equipment, a source describing a product family may not establish compatibility of a particular variant. For a consultancy, a case describing one implementation may not support a claim of experience across all Salesforce products.

Use supported, partly supported, contradicted or unable to verify as working labels. Record the passage and reasoning, then have specialist claims reviewed by someone qualified to assess them.

Inspect the claim and the source as a pair
Statement in the answerRead the exact assertion attached to a visible source link.
Evidence on the linked pageCheck whether the source supports that assertion, with its conditions.

This inspection tests claim support. It does not reveal hidden source-ranking rules.

A saved report exposes the limits of an absolute explanation

The New York–based Salesforce consultancy report from 5 September 2026 says its category source sets contained none of the consultancy's pages. It also describes platforms being recommended in category answers without their own pages appearing in those particular sets.

The report's displayed pattern therefore does not support the absolute statement that a system cannot recommend a business whose pages it did not read. It also does not establish that changing a source page necessarily changes a future answer.

This is a review of the report artifact, labeled ChatGPT and India with each question asked once. Original answer and tool records were not available. Treat “read” as the report's description until the underlying trace is inspected.

Useful content still needs a defined purpose

A business can improve the accuracy and usefulness of its evidence without claiming to control source selection. Explain the service scope, product facts or decision criteria a buyer actually needs. Keep time-sensitive information current and link related evidence sensibly.

Google's guidance emphasizes distinctive material built around the reader's needs. That is Google guidance, not a disclosed ChatGPT ranking rule. Google AI search guidance

Avoid importing a source preference from another platform as if it applied universally. Surface Diet explains the importance of product, sample and denominator when comparing citation patterns.

What you can check without guessing a hidden algorithm

You can check public-page accuracy, source support, access rules, internal navigation and the answers returned for a documented question set. You can also check whether a service page explains the evidence the buyer needs to assess fit.

You can then change a supported gap, retain the version and observe later responses. Report the conditions and other material changes. A before-and-after movement is an observation, not an isolated causal experiment.

Do not label a question set as real ChatGPT demand unless you actually have appropriate demand records. The prompt-research guide shows how to preserve the origin of questions without inventing usage volume.

Is a citation proof the claim is true?

No. Inspect the source and the statement. The association can be incomplete or incorrect.

Is allowing a crawler enough?

Access supports discovery. The page still needs relevant, accurate information, and its selection remains uncertain.

What should an owner do with this uncertainty?

Use it to keep the diagnosis bounded. Fix what the evidence supports, then measure what actually changes.

The practical citation guide turns these distinctions into a work plan. Request an audit if you want help reviewing your own answers and evidence.

Sources

From insight to execution

Make your company easier for AI and search to choose.

The Growth Retainer connects AI visibility research with GEO, SEO, content, technical fixes, authority, and ongoing measurement. Start with the service model or see your current gaps first.

Shubham Bansal, founder of The Cited Club

Written and reviewed by

Shubham Bansal

Founder of The Cited Club. Shubham leads the research and delivery behind its GEO, SEO, content, technical, and authority work.