What AI visibility means, and what it does not

Almost every conversation I have about this starts in the same confused place. Someone asks ChatGPT about their category, does not see their brand, and concludes their SEO is broken. Sometimes it is. Usually the thing that failed is not the thing they are looking at.

The three outcomes, separated

Mention

The model names your brand in the body of an answer. This can come from the model's training data, from a retrieved document, or from a live search the assistant ran. A mention carries no link and no traffic — but it is the thing that puts you on a buyer's shortlist.

Citation

The answer attributes a claim to a specific URL on your site, usually as a numbered source or a link chip. This requires that a crawler could fetch the page, that the page contained a passage worth attributing, and that the retrieval layer selected it.

Ranking

A position in a conventional results page. It is decided by a search index you can inspect in Search Console, and it has almost nothing to say about whether an assistant will repeat your name.

Why separating them changes what you build

If the problem is that you are never mentioned, publishing more pages will not help. The model is not reading your site in the moment it answers; it is repeating a description of your category that it learned from how the rest of the web talks about it. The work is to make sure the web describes you, consistently, in the words a buyer would use.

If the problem is that you are mentioned but never cited, the failure is almost always mechanical. The page was blocked, or slow, or rendered entirely in JavaScript, or written in a way that offers nothing a machine can lift as a self-contained passage. That is an engineering problem and it has engineering fixes.

How you would actually measure this

The unit of measurement is a buyer question, not a keyword. A keyword is what someone types into a search box; a buyer question is what they ask an assistant when they are trying to decide something — which supplier, which product, which alternative, which is better for my case. Those are different sentences and they produce different answers.

This is the problem XORV AnswerRank exists to solve, and building it is where most of what I know about this came from. The measurement is harder than the optimisation: getting a defensible number out of four engines that fail in different ways, and refusing to show a number when the data does not support one, is most of the engineering.