How to get cited by ChatGPT

The practical consequence: citation is the half you can win quickly and deliberately, one page at a time, and it is the half a small company can win against a large one. A well-built page on a narrow question can be the cited source for that question within weeks, regardless of how well known the brand behind it is.

Four page-level failures that a site-level check will not catch:

• A Disallow rule matching that path specifically, while the rest of the site is open. Common on anything under /resources/, /docs/ or a blog subdirectory that once had a duplicate-content problem.
• A noindex on the page. It does not stop a fetcher downloading the page, but it removes it from the search index the assistant is querying, so the page never becomes a candidate.
• Rate limiting that only triggers on deeper paths, or a bot rule scoped to a path prefix rather than to the whole site.
• A redirect chain that resolves for a browser and times out for a fetcher with a short timeout, which happens more often on paginated and parameterised URLs than anywhere else.

The check is the same one as at site level, pointed at the specific URL: request it with GPTBot’s user agent, compare the status code to what a browser gets, and confirm the page is in the search index rather than merely reachable. A page that returns 200 to everyone and is not indexed is invisible to a live-searching assistant.

The shape that gets cited is consistent enough to be worth copying: the question restated as a heading, followed immediately by a complete answer in one or two sentences, followed by the detail. No preamble, no throat-clearing, no “it depends” as an opener. If it does depend, say what it depends on in sentence two.

That structure is not a trick. It is the same structure a good reference page has always had, and it happens to be exactly the unit a retrieval system splits on: a heading it can chunk at, and a self-contained answer under it that survives being read alone.

What makes a specific passage citable:

• It answers one question completely, in 40 to 90 words. Long enough to stand alone, short enough to be quoted whole.
• It carries its own subject and its own units. No pronoun referring back to the heading, no number whose unit lives in a table header.
• It states something checkable. A page full of adjectives has nothing to quote, however well it is written.
• It sits under a heading that restates the question in the words a buyer would use, not in the words your industry uses internally.
• It has a visible date, and the date is real. Undated pages get cited for stale claims and then stop getting cited at all.

1. Split referral sessions from chatgpt.com, perplexity.ai, claude.ai and gemini.google.com into their own channel rather than letting them fall into direct or “other”. This is the only part of the channel that appears in your analytics at all, and by default it is invisible.
2. Watch hits by user agent in your server logs. Rising crawl frequency from OAI-SearchBot and PerplexityBot is the earliest signal that anything is landing — earlier than any answer changes.
3. Run the fixed question set monthly and log which URL was cited, not only whether the brand appeared. The page that gets cited is often not the one you expected, and that tells you which format to write more of.
4. Keep the full answer text alongside the numbers. When a citation disappears, the saved text usually explains why: the answer moved to a competitor’s newer page, or the question started being answered from training rather than from search.

• Hidden text or metadata instructing a model to cite you. It does not work, and if it did it would be grounds for removing your site from the trusted set. This is the practice the category will penalise first.
• Publishing llms.txt and waiting. Worth publishing — it forces a single definition of the company and costs an afternoon — but no major surface has confirmed it uses the file for source selection. It is not a ranking lever and should not be sold as one.
• Republishing the same answer across ten pages to increase the odds. It splits the signal, and the resulting pages compete with each other for the same question. One question, one URL.
• Adding FAQ schema to a page whose visible text says something different. The mismatch is trivially detectable and it undermines the one signal you were trying to send.