Three numbers worth tracking before any others

Most AI visibility dashboards lead with one composite score. It is the least informative number on the page, because three different problems all push it down and each needs different work. If you want the argument for why a share is the right shape of metric, start there; this is how to compute one.

1. Presence: how often are you named at all?

Across a fixed, documented set of phrasings your customers actually use, in what share of answers does your brand appear? That is the base rate, and it only means something with the sample attached: which engines, which phrasings, which languages, how many runs, over what window. Answers vary run to run even at temperature zero, so presence is a proportion with an interval around it, never a yes or no. The sampling design decides whether the number is real.

Two refinements make it more useful. Separate answers where a page of yours was cited from answers where you were named with no citation: those are two different mechanisms with different fixes. And record position within the answer, because being the fourth brand mentioned in passing is not the same as leading the recommendation.

2. Perception: when you are named, what is said?

Presence counts mentions. Perception reads them. The attributes an assistant attaches to your brand are a summary of what its sources say, and they are frequently dated, thin or borrowed from a competitor’s framing. Track the recurring adjectives, the use cases you get recommended for, the caveats that follow your name, and the factual errors. Errors are the highest-value finding here: they are concrete, they are often traceable to a specific weak source, and fixing the source is tractable.

3. Displacement: who is there when you are not?

An answer has room for roughly four sources: one large study found assistants citing a mean of 4.3 URLs and 3.4 domains, against 10.3 and 7.3 for traditional search engines (Zhang et al., arXiv:2512.09483). In a constrained space, presence is mostly a question of whose slot you take.

So record, for every answer that omits you, which brands appeared and which sources were cited. The pattern is usually more actionable than your own score, and the cited sources are often not competitors at all but retailers, reviews and forums.

Why not roll them up

A composite score moves for reasons you cannot read off it. Presence can fall because a competitor gained a slot, because a review site changed its recommendation, or because the engine changed how many sources it cites. Those are three unrelated situations, and only the raw three tell you which one you are in.

The reporting rule that follows: every number ships with its sample, its date and its engine. Anything that arrives as a single figure with no denominator should be treated as marketing, including the citation statistics in wide circulation.