How to Know If ChatGPT Cites Your Website

Build a test you can repeat

Start with five to ten questions customers ask before they know your company: how to solve a problem, what to compare, which option fits a constraint, and which alternatives exist. Keep the wording stable. Include a branded query as a separate diagnostic, because it tests recognition rather than discovery.

Run each question in the same product mode, language, region and account state. Record the exact interface and date. ChatGPT Search can return current web sources, but the answer and source set may vary by query and run, so a screenshot without its prompt and context is weak evidence.

• Save the full answer, not only the visible brand list.
• Copy every cited URL and note which claim it supports.
• Mark mention, citation, recommendation and factual accuracy separately.
• Repeat the set later before calling a change a trend.

Diagnose the missing link

If the company is absent, inspect the answer before editing the homepage. A competitor may be present because a comparison page names the use case clearly, because independent sources describe it consistently, or simply because the query exposes a different need. Treat each explanation as a hypothesis to verify against the cited pages.

Check whether the page you expect to be cited is public, returns a successful response, can be crawled by the relevant provider, and states its audience, category, use case and limits in visible text. OpenAI documents OAI-SearchBot access for ChatGPT search; this is an eligibility condition, not a promise that a page will be selected.

• Choose one source gap and one page to improve.
• Change the page only when the answer shows a real ambiguity or omission.
• Keep the original prompt and answer as the before state.

When to call a result inconclusive

Some prompts produce no citations, or different sources on each run. That is not automatically a defect in your page. Check whether the interface searched the web, whether the question needs current sources, and whether your sample contains enough repeated observations to support a pattern. Mark missing or unavailable outputs instead of filling them in from memory.

If the answer names a competitor but provides no source, you can record the recommendation while leaving its explanation unknown. Do not infer which competitor page caused the answer unless the interface exposed that page or you have independent evidence. A careful audit makes uncertainty visible so the next test can reduce it.

• Keep unavailable runs in the denominator policy explicit.
• Do not guess the retrieval path behind an uncited answer.
• Retest before assigning a cause.

Test real buyer prompts

Do not start with your brand name. Start with the question a buyer would ask before knowing you exist: best tool for a use case, how to solve a problem, alternative to a known product, or checklist for a job.

Run the same prompts in ChatGPT, Perplexity, Gemini and Google AI. Save the answer, the brands mentioned, the cited sources, and the language used to describe each option.

• Does the assistant mention your brand without help?
• Does it cite a page from your site?
• Does it describe your product accurately?
• Does it cite competitors instead of you?

Separate mention, citation and understanding

A mention is not a citation. ChatGPT can know your brand name but cite another source. It can also cite your page while misunderstanding your positioning. Track those three signals separately.

The strongest signal is a sourced answer that mentions your site, links to a relevant page, and explains the use case in terms your buyer would recognize.

Fix pages before publishing more

If a repeated test shows that AI answers miss your brand, the first fix is not always a new article. Review the page that should answer the buyer question: clarity of the category, concrete evidence, useful comparison criteria, relevant internal links, and access for the provider in question.

Create a new article only when a real buyer question has no useful existing page. Otherwise, record one change hypothesis and retest the same prompt and provider after a defined observation window.