A mention is not a citation

AI visibility asks a set of language models the questions your buyers ask, and reads each answer for two different things: your brand named in the prose, and your domain linked to. They are counted separately, and the separation is the whole design rather than a detail of it.

A name is a guess and a domain is a fact

A brand whose name is an ordinary word — Gap, Apple, Next, Boots — matches prose that is not about it. Word boundaries do not fix that, and neither does capitalisation, because the first word of a sentence is capitalised too. Rankgap does not claim to solve it. What keeps it contained:

• The sentence around every mention is stored, so you can read it and decide for yourself in one glance.
• A domain is matched by a hostname rule of its own rather than by the rule used for words, because what counts as the edge of a word is not what counts as the edge of a host.
• The two are counted separately, in separate columns. Averaging a fuzzy signal into a precise one gives you a number that is neither.

An answer that never came back is not an answer saying no

The prompt grid tells a call that failed and a call still running apart from an answer that named nobody, rather than folding the three into one cell. A failed call reads as no answer and a running one reads as waiting; neither becomes “not named”, because a sweep where every call failed would otherwise render exactly like a sweep that measured your market and found you absent — and those are an outage and a finding, which a reader has to be able to tell apart.

The rates follow the same rule. The denominator is the answers that actually came back rather than the number of calls attempted, and the mention rate and the citation rate count theirs separately because they read different records. A sweep with nothing to divide by reads as a dash, never as a rate of nothing.

What the order among names is, and is not

When an answer names several brands, Rankgap records where yours came among the ones you asked about — your own domain and the rivals you track. It is not a place among every brand the model named. Track one more rival and your own position in that list can move without the model having said anything different, which is why the column says it is a place among the brands you track.

And what is being measured

The thing answering is a model reached through an API, not a consumer assistant. That call carries none of a chat product’s own instructions, memory or retrieval, so a figure here is about the model named beside it and about nothing else. A sweep is also a sample of a system that does not answer the same way twice, which is why the screen shows a rate over several recent sweeps rather than a verdict from the latest one.