How to rank in ChatGPT, and why there is no rank to win

The most common thing people want from AI search is a position. They ask how to rank in ChatGPT the way they would ask how to rank in Google, and the question carries an assumption that does not survive contact with the thing itself. There is no ranking. There is no result three that you can climb to from result eight, no list to be on, and no number to check tomorrow to see whether yesterday worked.

What exists instead is a sentence, written fresh each time somebody asks, which may or may not contain your name. That is a different object, and the difference decides almost everything about what you can actually do.

What is happening instead of ranking

When a buyer asks an assistant for a recommendation, the model produces an answer out of two things: what it absorbed during training, and whatever it retrieved at the moment of asking. Neither of those is an index of pages sorted by authority. The output is prose, and your brand is either named in it or it is not.

Three consequences follow, and they are the practical part.

The first is that there is no single answer to be correct about. Ask the same question twice and the wording moves. Ask it in a different phrasing and the shortlist can change. Anyone promising you a fixed position in something that is regenerated on every request is selling you a picture of a thing that does not hold still.

The second is that the assistants disagree with each other, and not by a little. We ran ten identical buying questions about website builders across three models on one day, with no brand names in the questions. Webflow scored 15 on ChatGPT and 60 on Claude. ChatGPT named it in three of ten answers, Claude in nine. That is not a rounding difference. It is two systems holding materially different views of the same market, and a buyer who asked only one of them would leave with a different shortlist depending on which one they happened to open.

The third is that your own website is not where most of the decision is made. This is the part people like least and it is the most useful thing on this page.

Why your homepage is not doing the work

A model writing a recommendation is reproducing a consensus it absorbed from a very large amount of text. Almost none of that text is your marketing site. It is other people writing about your category: roundups, comparison posts, forum threads, review pages, documentation, news.

Your own pages describe you accurately and in your own words, which is exactly why they carry less weight in this particular job. Every competitor's site also says its product is the best one. What distinguishes brands inside an answer is how often and how consistently other people name them when discussing the category.

This is why the work that moves AI visibility looks less like on-page SEO and more like public relations, community presence, and getting included in the lists other people maintain. If the pages that answer "best tools for X" do not mention you, the assistants assembling answers from that same material have very little reason to.

What you can actually influence

Here is what is worth doing, roughly in order of how much it seems to matter, with the honest caveat that this is a young field and nobody has clean causal evidence.

Be present in third party coverage of your category. Roundups, comparison articles, directories with real editorial standards, credible review sites. This is slow and mostly involves asking people, but it is the closest thing to a lever that exists.

Be named consistently. Pick one name and one description of what you do, and use them everywhere. Our own extraction counted "Dashlane" and "Dashlane Business" as two separate brands, and "Sendinblue" and "Brevo" as two, when the latter pair is one company before and after a rename. If a machine reading the web splits your mentions across variants, every one of those variants is weaker than the total would have been.

Make your category obvious in plain language. Answers get assembled around the way buyers phrase things, not the way your positioning document phrases them. If your site describes a category nobody types, you will not appear against the category they do type.

Be readable by the things that retrieve. Clear headings, a direct statement of what the product is near the top, factual specifics rather than adjectives, and an llms.txt if you want to state plainly what you do. This is cheap and sensible. It is not a growth strategy on its own, and treating it as one is the most common way to spend a quarter on nothing.

Publish things worth citing. Original numbers, honest methodology, findings someone else would want to reference. A page that other people quote is a page that enters the material answers are built from.

What nobody can promise you

No one can guarantee that ChatGPT will recommend you, and anyone who does is either misunderstanding the system or counting on you to. The models update, the retrieval changes, and the same question asked twice can come back differently.

You also cannot fix this from inside your own site alone. If the consensus in your category does not include you, no amount of schema markup changes what four assistants say about it. Structured data helps a machine understand a page it has already decided to read. It does not manufacture a reputation.

And there is a limit on measurement itself worth stating plainly. Answers taken through the official APIs approximate what the consumer apps say without being identical to them. One run on one day in one language is a snapshot, not a trend. We say this about our own product because it is true of every tool in this category, including the ones that do not mention it.

Measure it before you try to change it

The useful move is unglamorous. Before doing any of the work above, find out where you actually stand, because most people are surprised and a good number are surprised in the direction of doing less work than they feared.

Write ten questions a buyer in your category would genuinely type, with no brand names in them. Ask each one on more than one assistant. Record which brands get named, in what order, and which sources the answers lean on. Then compare the assistants against each other rather than averaging them, because the disagreement is the most informative thing in the data. A brand sitting at 15 on one model and 60 on another has a very different problem from one sitting at 40 on both.

Do that before you spend anything. Then do it again in a month, and you will know whether the work moved anything, which is the question the ranking framing was trying to ask in the first place.

Spegla runs that loop for you across ChatGPT, Claude, Gemini and Perplexity, in the buyer's own language, and shows you the full answers rather than a score on its own. The free check needs no account. Whether you use it or do it by hand in an afternoon, do the measuring first. Nothing else on this page is worth acting on until you know which problem you have.