How Do You Rank in ChatGPT?

The short answer: there is no ranking to win

ChatGPT does not rank websites. It composes an answer, and sometimes names a handful of sources. So the goal is not position, it is inclusion: being retrieved, being quoted, and being named.

That distinction matters because it changes the work. Optimising for a position means competing on relevance and authority against everyone else on a results page. Optimising for inclusion means making yourself easy to find, easy to quote and easy to trust, which is a different and in some ways more tractable problem.

It also means the usual proof is unreliable. A screenshot of ChatGPT naming your brand proves very little, because the same question asked twice can produce different answers. Measurement has to be a repeated prompt set, reported as a trend.

How does ChatGPT actually decide what to say?

Two routes, and they behave differently.

• Training: what the model absorbed when it was built. You cannot influence a finished training run, and anything recent or specific is unlikely to be in it accurately.
• Retrieval: the model searches, fetches a few pages, and writes from passages inside them. This happens live, on every relevant question, and it is where all the available leverage sits.

For anything commercial, current or specific, assume retrieval. That is good news, because retrieval is influenced by exactly the things you control: whether your page can be fetched, whether it contains a clean answer, and whether other sources corroborate it.

What actually changes whether you get cited?

In our experience, three causes account for most of the difference, and they are rarely the ones people expect.

The last row of the left column is worth dwelling on. A site rebuilt in a modern JavaScript framework can be perfectly indexed by Google, because Google renders pages, and serve GPTBot an almost empty document, because most AI crawlers do not run JavaScript. Nothing warns you.

Step one: make your entity unambiguous

Before a model can recommend you, it has to work out who you are. That sounds trivial and it is where most businesses lose.

Check whether your company name, address, service names and descriptions are the same on your website, your schema, your Companies House record, your LinkedIn page, your Google Business Profile and the directories that list you. Where they disagree, the model resolves the conflict using whichever source it trusts more, which is frequently not you.

1. List every place your business is described online.
2. Write down the version of the facts you want to be true.
3. Correct the ones you control, starting with your own site and schema.
4. Request corrections on the ones you do not.
5. Keep them aligned when something changes, which is the part that gets dropped.

This is unglamorous and it is the single highest-yield change for most sites. It is covered in more depth on our AI SEO page.

Step two: write so you can be quoted

Retrieval pulls a chunk of text. If that chunk contains a complete answer, it gets used close to verbatim and your framing survives. If it does not, the model paraphrases from several sources at once and your contribution dissolves into someone else's wording.

The practical rule: under every heading, answer in the first two or three sentences, then explain. No warm-up, no restating the question, no promise that you will cover it below.

This is the same discipline as answer engine optimisation, which is worth reading alongside: AEO targets the direct answer slot, and the writing style that wins it is the style assistants can quote.

Step three: let the crawlers in

Check your robots.txt for GPTBot, OAI-SearchBot and ChatGPT-User. Blocking them is a legitimate choice if you would rather your content was not used, but it should be a decision, and a surprising number of sites block them by accident or by inherited default.

Then check what those crawlers actually receive. Fetch a key page without executing JavaScript and read what comes back. If the text is not there, no amount of content or entity work will help, because the page the model sees is blank.

One more trap: a named user-agent group in robots.txt replaces the wildcard group rather than adding to it. If you add a GPTBot section, it inherits none of your existing rules unless you repeat them.

What about llms.txt?

It is a proposed convention, not an adopted standard: a plain-text file at your site root describing the organisation and linking to key pages. Support is partial and unverified.

Add it, because it costs almost nothing and does no harm. Do not treat it as a ranking factor, and treat anyone who describes it as one with caution.

Which AI crawlers should you actually allow?

There is no single right answer, but there is a wrong one: not knowing. Here is what each crawler does, so the decision is deliberate.

A useful split: if your concern is being quoted, the search and user-facing crawlers are the ones that matter. If your concern is training, GPTBot and Google-Extended are the relevant controls. Blocking the first group while allowing the second is the combination that makes least sense, and we see it fairly often.

Google-Extended is the one most often misunderstood. Disallowing it does not remove you from Google Search or from AI Overviews built on Search; it governs Gemini and model training. If someone tells you blocking it will hide you from AI Overviews, they are mistaken.

A worked example of rewriting for retrieval

The difference between a page that gets quoted and one that does not is usually a paragraph of throat-clearing. Consider a page about invoice automation that opens like this:

In today's fast-moving business environment, finance teams are under more pressure than ever. Many organisations are looking at automation as a way to reduce costs and improve accuracy. In this article we will explore what invoice automation is, how it works, and whether it is right for your business.

There is nothing in there to lift. A model asked what invoice automation costs, or how it works, finds no answer in the opening and moves to a source that leads with one. The rewrite:

Invoice automation reads supplier invoices as they arrive, extracts the line items, matches them against your purchase orders, and sends only the mismatches to a person. For most mid-sized finance teams it removes two to four hours of daily re-keying. It works best where invoice volume is high and the exceptions are a minority.

Same information, same page, entirely different retrievability. Nothing was added; the order was changed.

Common mistakes we see

Six things that waste money in this area, in rough order of how often we encounter them.

1. Publishing more posts. Volume makes the entity picture noisier. Correcting what exists usually beats adding to it.
2. Treating one screenshot as proof. In either direction. Outputs vary between runs.
3. Assuming Google indexing means AI visibility. They are different crawlers with different capabilities, and one does not imply the other.
4. Adding schema for its own sake. Markup that validates but states nothing useful about the business does not help a model resolve who you are.
5. Ignoring third-party sources. What directories, trade press and review sites say about you is weighed more heavily than what you say about yourself.
6. Buying guaranteed placement. Nobody controls what a model says. Treat the offer as a signal about the seller.

How do you measure any of this?

Pick the questions your buyers actually ask before choosing a supplier. Run them across ChatGPT, Google AI Overviews, Perplexity and Gemini on a schedule. Record three things each time: whether you are named, what is said about you, and which URL is cited.

Because model answers vary between runs, a single result tells you almost nothing. Repeat the same prompts and report the trend. If your agency shows you one screenshot as evidence, ask what the other nine runs said.

If you would rather not set this up yourself, our free AI visibility audit does exactly this and sends you the findings in writing, with no obligation.