How to get your documentation cited by ChatGPT

Somebody who used to search for "northwind api rate limit" now asks an assistant, and the answer arrives with two or three links under it. Being one of those links is worth more than a position on a results page, because the reader arrives already told that you are the source.

Two separate things decide whether you get there. An assistant has to be able to fetch your page, and then it has to find one passage worth quoting. The first is a five-minute configuration job. The second is a writing job, and it is where almost all documentation loses.

The part nobody can shortcut

An assistant quoting you is doing something specific: lifting a short passage out of your page and putting it next to your name. If the passage stops making sense once it leaves the page, it does not get used, however good the page is.

Here is the same fact written three ways. Each one is correct, and only one of them survives the trip.

The third version is not better writing in a literary sense. It is worse, if anything, to read top to bottom. It is written for somebody arriving at that heading from a question, which is how most people now arrive.

Five properties of a quotable passage

• The subject is named in the sentence. "We", "it" and "the platform" are fine in a paragraph and useless in a quote. Write your product's name where a reader might arrive cold.
• Numbers carry their unit and their scope. Not 120, but 120 requests per minute per API key.
• One claim per paragraph. Two facts in one paragraph means an assistant lifting either one drags along the other.
• The heading is the question. People ask assistants in whole sentences, and a heading phrased the same way is the strongest match you can offer.
• The consequence follows immediately. A limit without its error, a price without its period, a step without its result: each one invites the model to fill the gap from somewhere else.

The fetching half, in five minutes

None of the writing matters if the page cannot be read. Three OpenAI agents touch your site and they do different jobs, which is why blocking them is not one decision.

Three more things belong on the same checklist. Serve the content in the HTML rather than building it in the browser, because a fetcher that does not run your JavaScript sees an empty page. Keep the answer above any interstitial, since a cookie wall is often all that is read. And give every page a title that names the subject, not just the section.

Where the effort is usually wasted

Keyword density does nothing here: there is no ranking function to feed, only a passage that is either liftable or not. Publishing a file at your root does not get you quoted either, because being listed is not the same as being worth quoting. And schema markup helps machines classify a page, but assistants quote prose, so a perfectly marked-up page of vague sentences still loses to a plain one that answers the question.

The work that pays is unglamorous: take the twenty questions your customers actually ask, give each one a heading, and answer it in the first two sentences underneath.

Questions people ask

Pick the five questions you would hate a competitor to be cited for. Give each one a heading that matches how it gets asked, answer it in two self-contained sentences, and check in a month who gets quoted.