What is the most useful AI visibility metric?

Recommendation rate: the share of relevant answers that positively recommend your brand. It counts a clear event and has a fixed denominator, so it can be compared over time.

Is a single AI visibility score reliable?

Only if you can see what it is made of. A composite whose weights and inputs are hidden can shift in meaning between brands and assistants, so it is safer to track the underlying measures.

Why is a mention not the same as a recommendation?

A mention can be neutral or even negative, such as naming a brand as one to avoid. A recommendation presents the brand as a choice to make.

Do I need confidence intervals?

If you are comparing small samples, yes. Without an indication of uncertainty, a move from 10 percent to 12 percent can be nothing more than variation.

AI visibility metrics that actually mean something: Citewake

The AI visibility metrics worth tracking each count one clear event over a stated denominator: how often answers recommend you, how often they cite your site, how often pages they cite list you, and your share of all recommendations. A single unexplained score is the metric to distrust.

A good metric tells you what to do next. A bad one gives you a number to report and nothing to act on.

What makes an AI visibility metric trustworthy?

Four tests separate a useful measure from a decorative one.

It counts a stated event. "Answers that recommend the brand" is an event. "Visibility" is not.
It has a fixed denominator. You must know what it is a share of: valid answers, questions asked, pages cited. If the denominator shifts, the number cannot be compared month to month.
It is traceable. You can click from the number to the actual answers that produced it.
It admits uncertainty. Answers vary between runs, so a number from a small sample should come with a margin.

If a tool cannot tell you these four things about its headline score, treat the score with caution.

What does recommendation rate measure?

Recommendation rate is the share of valid answers that positively recommend your brand.

It is the closest thing to the outcome you care about: an assistant telling a buyer to choose you. It should be computed over questions where a good answer would name products, not over informational questions where naming none is correct.

Report it per assistant and overall. A rate of 40 percent in one assistant and zero in another is a different story from 20 percent in both, though the overall averages match.

What does citation rate measure?

Own-site citation rate is the share of valid answers that cite a page on your own domain.

It tells you whether assistants are using you as a source. It is not the same as being recommended. You can be cited for documentation while a competitor is recommended, and the reverse is also possible. Keep the two apart.

Pair this with the kinds of pages being cited. The kinds of pages AI assistants cite most helps interpret what a citation pattern says about your content.

What does cited-page presence measure?

Many recommendations are shaped by third-party pages: roundups, reviews, comparisons. Cited-page presence asks: of the third-party pages that answers cite, how many list your brand positively?

This is often where the gap is. If the pages assistants rely on never mention you, improving your own site only goes so far. It is a leading indicator in practice, though whether it reliably comes before recommendations is a hypothesis that has to be tested, not assumed.

To act on it, see how to get onto the roundup and review pages AI relies on.

Competitive share is your positive recommendations divided by all brands' positive recommendations in the same set of answers.

It is useful because it shows where you stand against rivals. It is also relative. If a competitor gains, your share can fall while your own recommendation rate holds steady. Always read it alongside the absolute rate so you can tell which of the two moved. More on this in share of voice in AI answers.

What about position and "rank"?

Being listed first is not evidence of preference by itself. Lists can be alphabetical, by date or arbitrary. Count position only when the answer expresses preference, such as "my top pick" or a ranking presented as a ranking. A metric built on that is a top-pick rate.

Anything called your "AI rank" without that qualification is measuring something looser. There is no real rank #1 in AI answers explains why.

What should you be wary of?

A few common metrics deserve scepticism.

A composite score with hidden weights. If you cannot see the inputs, you cannot tell what moved it. A composite is only worth having once it has been shown to add information beyond the separate measures.
Counts of mentions. They mix recommendations, neutral naming and warnings.
Averages over unequal samples. A score based on three answers is not comparable with one based on three hundred.
Metrics that quietly drop missing data. Failed or refused answers should be counted and reported, not silently filled in or ignored.
Daily numbers. Answers vary so much from run to run that daily movements are mostly noise. Monthly measurement over a fixed set of questions is steadier.

How do you track these over time?

Hold the question set fixed, collect on a regular schedule, and compare each question's answer with the same question's previous answer from the same assistant. Call something a move only when the change is larger than normal variation. Otherwise, say "no clear change". How to track AI visibility over time gives a routine.

Which metric should you report first?

If you can only report one, report recommendation rate per assistant over a fixed question set, with its margin. Add own-site citation rate and presence on cited pages when you want to explain why. Keep the sample size visible. A small, clearly described set of numbers that a colleague can recompute is worth more than a long dashboard nobody can explain.

The method behind these measures, including the denominators and the uncertainty intervals, is set out in the methodology. If you want to see where you stand before building a tracking routine, you can try the Citewake checker.