Getting cited
ChatGPT cites a page when it searched mid-answer, found it, and could lift one sentence that settled the question. Everything that works follows from that, and most of what is sold in this category does not.
The short answer
Everything else on this page is a consequence of that sentence. Most advice in this category fails because it optimises for being crawled, which is necessary, and stops before being quotable, which is the part that decides it.
Two different things people mean
| From the model’s training | From a live search | |
|---|---|---|
| What you see | Your name in the text, no link. | Your name, and a citation pointing at a page. |
| What decides it | How often and how consistently you appeared across the web before the cutoff. | What the search returned seconds ago, and which of those pages had a liftable answer. |
| How fast you can change it | Not this year. It changes when the model does. | As fast as the record changes and the pages are re-crawled. |
| What goes wrong | Confidently out of date. It will describe the company you were. | It cites whoever answered the question, which is often a third party describing you in their words. |
| Worth working on | Only indirectly, by being consistent everywhere for a long time. | Yes. This is the whole addressable surface. |
The page-level work
An engine is assembling a short answer from several sources under a token budget. It favours text that answers without needing the rest of the page.
| Do this | Why it works |
|---|---|
| Make headings the question people type | Retrieval matches the question to text. A heading phrased as the question is the strongest available match, and it tells the engine which paragraph answers it. |
| Answer in the first two sentences under it | What gets quoted is usually the opening of the matched section. An answer that arrives in the fourth paragraph often does not survive the budget. |
| Write sentences that stand alone | A sentence starting “this is why it matters” is unquotable out of context. One that names its own subject can be lifted intact. |
| Use real tables | A marked-up table is the most structured thing on a normal page, and comparison questions are answered from them disproportionately often. A table rendered as an image is invisible. |
| Be specific, with numbers and dates | A concrete figure is checkable, attributable and worth citing. “Fast and affordable” is in every competitor’s copy and distinguishes nothing. |
| Publish something only you know | Original measurement is the most reliable way to be cited, because nobody else can be the source for it. It is also the slowest, which is why so little of it exists. |
| Keep it out from behind a gate | A crawler does not fill in forms, accept cookies or run your paywall. Anything that needs a click to reveal is, for this purpose, not published. |
The other half
White-on-white paragraphs, zero-height divs, content served to a crawler and not to a reader. It is the oldest trick in search and it is detectable by exactly the same methods. We refuse it as a rule rather than as a judgement call.
Sentences in a page telling an assistant to recommend the company. Engines increasingly treat page content as data rather than instruction, so the best case is that it does nothing and the worst is that it is read as an attempt to manipulate.
Review farms, testimonials from people who do not exist, a dozen thin sites repeating the same claim. Engines weigh agreement between independent sources, which is exactly what this fakes — and what a buyer will check once your name is in front of them.
Forty generated posts on the same topic give an engine forty mediocre candidates for one slot and dilute the page that was actually good. The unit of competition is one paragraph with about three names in it, not a page count.
There is a reason to care beyond scruples: the whole mechanism rewards a public record that agrees with itself. Anything planted is a claim you will have to keep defending, in a record you do not control, in front of a buyer who can check.
Then check it
Asking an engine once and finding yourself named proves very little: these answers vary between runs, between engines and between phrasings of the same question. A number worth reporting needs repeats, several engines, and a set of questions you deliberately left alone so you can tell your own effect from the market’s. The method, in full, is here — and what it costs to run is here, with the numbers we measured.
Questions
Sometimes, and the distinction matters. When it answers from training it is recalling a compressed version of the web as it was, with no link to you. When it searches, it fetches live pages, reads them into context and cites them. Only the second path can be influenced this week, and it is the one that produces a citation.
If you want to be cited, yes. GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and Google-Extended are separate agents, and a named group in robots.txt replaces the wildcard rather than extending it — so a rule written for one crawler can silently block the others. Check what your file actually resolves to rather than what you meant.
There is nothing to submit. There is no index to register with and no console to verify in. Engines find pages by crawling and by searching at the moment the question is asked, which means the only lever is being findable and quotable when that search happens.
On its own, no. It is a convention some clients read, not a registration, and no engine has committed to honouring it. It costs an hour and it does no harm, so we publish one — but anybody selling it as the reason you will get cited is selling you an hour of work.
Related reading
The full definition and the five things that move a mention, in the order we would do them.
What carries over from the search work you already do, and what quietly stops paying.
Measured numbers for what it costs to ask an engine a question properly, and why the search cap matters more than the model.
The same mechanics applied to an industry where one answer can decide a partner conversation.
Start here
The free diagnostic: how public sources describe your company and your main product today, where they disagree, and which disagreement to fix first. Every line links to the page we read it from. Nothing is published anywhere without your written approval.