The phrase gets used loosely enough that it is worth pinning down, because most of the confusion around it is definitional rather than technical.
AI visibility is whether, and how, an AI assistant names your business when someone asks it a question your business could answer.
That is the whole definition. Everything below is consequence.
Why it is not a rank
The instinct is to translate it into search terms: what position am I in ChatGPT? The question has no answer, and not because the data is unavailable.
A search engine returns a list. The list is ordered, mostly stable, and the same for two people in similar circumstances. Position two means something.
An assistant generates a sentence. There is no list to hold a position in, and the sentence is constructed fresh each time. Ask the same question twice in two fresh sessions and you can get two different recommendations, in a different order, drawn from different sources — a point covered at length in why AI answers change every time.
So the honest unit is not a position but a frequency. Not where do I rank but out of forty recorded answers, how often was I named, how often first, and what did the assistant cite when it named me.
This is a genuine difference in kind, not a measurement problem waiting for better tooling. Any product that hands you a single "AI visibility score" has compressed a distribution into a number and thrown away the variance — and the variance is frequently the most useful finding.
What it is repeatedly confused with
It is not SEO. Related, overlapping at the edges, different object. SEO is about which links appear on a results page. AI visibility is about whether your name appears inside a generated paragraph. A business can rank first on Google and be entirely absent from assistant answers, because the assistant is leaning on a comparison page that happens not to include it.
It is not your website's quality. Uncomfortable but consistent. When an assistant is asked which one should I use, it needs a source describing several businesses with reasons. Your site describes one, favourably, which is exactly the kind of source a model discounts. Your site confirms facts about you; third-party pages decide whether you are named at all.
It is not brand awareness. A well-known business can be missing from assistant answers in a specific category or city, because the sources covering that specific question do not include it. Visibility is question-shaped, not company-shaped. You do not have one AI visibility — you have a different one for each question a customer might ask.
It is not something you can buy. There is no ad inventory inside an assistant's recommendation. Anyone selling placement is selling either directory work under another name, or nothing.
What it is made of
Three inputs, roughly in order of how much they move the outcome:
Cited sources. When an assistant searches live, it reads pages and tells you which ones. Those pages — directories, "best of" lists, review platforms, forum threads — are the actual battleground. This is where most of the outcome is decided, and it is covered in which sources AI assistants cite.
Training data. What the model absorbed before it was deployed. You cannot edit it and you cannot appeal it. If it is wrong about you, the practical remedy is to make the live-search layer strong enough to override it.
Entity clarity. Whether the model can tell that the four differently-spelled listings all refer to one business. Dispersion here quietly taxes everything else.
Why "how often" beats "whether"
Here is the finding that reframes the whole subject for most people who encounter it.
When we asked four assistants forty questions about dental clinics in Tijuana and recorded 440 answers, they named 61 different clinics. Every one of those 61 was the top recommendation in at least one recorded answer. And when the same question was repeated in a fresh session, the top recommendation changed in 159 of 160 cases.
Read that as a definition rather than a curiosity. In that market, "who does AI recommend" has no single answer, and any clinic asking once and drawing its own name would conclude it was winning. Any clinic asking once and drawing a competitor would conclude it was losing. Both conclusions would be artifacts of sample size.
Visibility in that market is a share of a distribution. That is what there is to measure, and it is why the measurement has to be repeated to mean anything.
The practical version
If you want one sentence to act on: AI visibility is your share of the answers to the questions your customers actually ask, and it is won mostly on pages you do not own.
The method for measuring it is in how to test your AI visibility. What to do once you have measured is in how to actually appear in AI answers.