Measuring AI visibility: which numbers really exist

Why classic metrics don't work here

A position in Google search is unambiguous: third place is third place. A generated answer is not. It turns out a little differently with every query, it names three sources one time and none the next, and it doesn't necessarily link to what it is based on.

On top of that: a large share of these queries ends without a click. The user gets the answer and leaves. None of this shows up in Google Search Console, and none of it in website analytics either. The visibility exists, but it leaves no trace in the tools most companies use.

This means: if you want to know how you stand in AI systems, you have to actively check. Passive tracking doesn't give you the picture.

The question behind the measurement

A position in Google search is unambiguous: third place is third place. A generated answer is not. It turns out a little differently with every query, it names three sources one time and none the next, and it doesn't necessarily link to what it is based on.

On top of that: a large share of these queries ends without a click. The user gets the answer and leaves. None of this shows up in Google Search Console, and none of it in website analytics either. The visibility exists, but it leaves no trace in the tools most companies use.

This means: if you want to know how you stand in AI systems, you have to actively check. Passive tracking doesn't give you the picture.

The three metrics that count

Mention rate. In what percentage of answers to a defined list of questions does the brand name appear? That is the basis. A mention without a link is valuable because it builds trust and often leads to a brand search later.

Citation rate. In what percentage of answers is your own website named or linked as a source? That is the stricter benchmark and the one that brings in visitors.

Share of voice. How often does your own brand appear, measured against all providers named? Two mentions with three providers named is something different from two with twenty.

A fourth factor is tone: is the brand named as an example, as a recommendation or as a side note? This is hard to capture automatically, but you can see it right away when reading the answers.

The manual test in two hours

A first measurement doesn't need a tool. It needs a list of questions and discipline.

Step 1: Formulate twenty questions a customer would actually ask. Not keywords, but sentences. A mix of category questions, comparison questions, price questions and problem questions.

Step 2: Enter every question into at least three systems. Important: without being logged in or in a new window, so personalization doesn't play a role.

Step 3: Note three things for every answer. Does your own brand appear, is your own site named as a source, which competitors show up.

Step 4: Enter everything into a table and calculate the three rates.

The result is a snapshot, not a statistically clean survey. For the question of whether you appear at all, it is completely enough.

What the tools do and don't do

There are now providers that automate exactly this measurement: they regularly send a defined list of prompts to several systems and analyze the answers. That saves time and gives you a trend instead of a snapshot.

There are three limitations you should know about.

The prompt list decides everything. A tool only measures what you have it measure. A poorly chosen list produces a nice curve that has nothing to do with reality.

The answers fluctuate. The same system gives different results to the same question on two different days. Individual values are therefore worthless, only trends over weeks are meaningful.

No tool covers everything. The systems differ greatly in where they get their information from. If you only measure one, you only measure a slice.

From measurement to action

A measurement without consequences is just a report. It gets interesting with the question of why a competitor appears and you don't.

In most cases, it comes down to one of four points: there is no content on the topic. There is content, but it doesn't answer the question directly, only after five paragraphs of introduction. The content is there, but it isn't marked up cleanly in technical terms. Or the brand isn't mentioned anywhere on other sites, so no model has a reason to know it.

The fourth point is underestimated most often. Generative systems rely heavily on what others write about a brand: directories, expert articles, comparison lists, forums. A brand that nobody writes about stays invisible, even with a perfect website of its own.

How often to measure

Monthly is enough in most cases. The underlying models and their sources don't change daily, and measures need time to take effect.

A fixed approach makes sense: the same list of questions, the same systems, the same day of the month. Only then are the values comparable. If you expand the list in between, keep running the old one in parallel, otherwise the series breaks.

One-off additional measurements are worthwhile after major publications, after a mention in a trade publication or after a technical changeover.

The most important date is the first one. If you measure today, you can say in six months whether something has happened. If you don't measure, you can never prove it, neither to yourself nor to others.