Why isn't my site showing up in ChatGPT?
Diagnose a missing ChatGPT citation as a set of hypotheses, not a single known cause. Separate OAI-SearchBot policy for ChatGPT search discovery from GPTBot policy for potential model training, then inspect page delivery, answer structure, returned sources, and repeated prompt observations. RankEcho can record what configured prompts returned before and after a shipped change; that narrows the investigation but does not prove why an answer changed.
More: Intro to RankEcho · How it works · Full walkthrough · Written product guide
Hypothesis 1 — Search access is constrained
OpenAI documents OAI-SearchBot for ChatGPT search discovery and GPTBot for potential model training; the two controls are not interchangeable. A robots.txt rule shows declared policy, while CDN or origin logs and provider verification can add request evidence. A request sent with a crawler-looking User-Agent only shows how the site handled that supplied header. Anyone can spoof it, so the result is useful for reproduction but is not authoritative evidence of what an authentic OpenAI crawler received. A block or challenge can matter for a particular route and crawler role, but it does not by itself explain every missing mention or citation.
Hypothesis 2 — The answer is difficult to extract
The initial HTML is one useful compatibility check because not every client executes a site's JavaScript, but provider behavior varies and some systems may render pages or rely on an index. Compare the initial response with the rendered page, keep the substantive answer clear and visible, and validate structured data against the visible text. A missing or buried answer supports an extraction hypothesis; it does not reveal whether ChatGPT fetched, indexed, retrieved, or selected the page.
Hypothesis 3 — Other source coverage differs
Returned sources can show which first-party or third-party pages accompanied a sampled answer. Compare their scope, specificity, freshness, and corroboration with your own coverage, and treat relevant omissions as research leads. The source list does not disclose the provider's weighting or prove why a competitor appeared, so an off-site mention should be tested as one hypothesis rather than treated as the deciding move.
How to diagnose in order
Work from narrow observations to broader hypotheses:
- Record the ChatGPT surface, exact prompt, locale, date, and returned source URLs.
- Inspect OAI-SearchBot policy separately from the GPTBot training choice, then reconcile policy with authentic edge or origin evidence where available.
- Compare the initial HTML, rendered page, status, canonical, and indexing directives for the exact URL.
- Review the returned sources and competing pages for coverage your page may lack.
- Repeat a fixed prompt panel before and after one recorded shipment, reporting later movement as an observation rather than a causal verdict.
Testing off-site source coverage
Third-party coverage can be relevant when sampled answers return reviews, roundups, directories, or community discussions. Start with sources actually observed for the prompt set, assess whether they are relevant and independent, and use accurate, disclosed participation rather than manufacturing mentions.
The aim is not raw mention volume. It is to test whether credible source coverage fills a documented information gap, while recognizing that publication on any one source does not guarantee retrieval, citation, or recommendation.
Technical signals that help extraction
Schema and technical hygiene do not force a citation. Valid structured data that matches visible content, useful initial HTML, accurate canonicals, and a current sitemap are interoperability checks, not evidence of provider selection. Provider rendering and indexing behavior varies, and the dated sources reviewed here do not establish general ChatGPT support for llms.txt.
Use these checks to remove testable delivery ambiguity. Keep indexing, retrieval, answer inclusion, and citation as separate observations rather than inferring them from a successful fetch or valid schema result.
Where RankEcho fits
RankEcho records configured prompt observations across the engines available to the audit, including owned source URLs and competing returned sources. A gap can become a proposed fix package, and the Proof Loop can repeat the same prompt panel after a recorded shipment. The resulting timeline shows what changed after publication; it does not identify which change caused the movement.
Publishing, implementation, and third-party outreach stay with the operator. RankEcho organizes the measurement and shipment record so a team can test the next hypothesis without presenting correlation as proof.
Frequently asked questions
Inspect the robots.txt group selected for GPTBot and the relevant edge rules; that describes the potential-training policy, not ChatGPT search access. Inspect OAI-SearchBot separately for search discovery. A crawler-shaped User-Agent is spoofable and non-authoritative, so reconcile test fetches with provider-verified request evidence plus CDN or origin logs where available.
There is no guaranteed discovery path or timetable. Identify the ChatGPT surface you care about, check the documented access and indexing prerequisites that apply to it, and collect repeated prompt and source observations before choosing a fix.
There is no reliable fixed interval. Timing can vary with the product surface, crawl and index updates, the prompt, and answer variability. Record the shipment date and compare repeated observations on a fixed prompt panel without promising an appearance.
Treat access, page coverage, answer structure, and corroborating sources as separate hypotheses. Establish a repeatable prompt baseline, fix a documented gap, record what shipped, and sample the same prompts afterward. A later citation is an observation, not proof that the change caused it.
The answer itself does not reveal why a competitor was selected. Compare the exact returned pages for relevance, specificity, freshness, technical delivery, and corroboration, then test the smallest supported hypothesis while allowing for provider and run-to-run variability.
There is no universal fastest lever. Correct a verified access or serving fault when one exists; otherwise compare page coverage and the sources returned for the target prompts. Use a fixed baseline and repeated post-ship observations, because one appearance cannot confirm which action produced it.