What is AI visibility, and how do you measure it?

AI visibility

How often AI assistants name your business in answers to the questions your customers actually ask.

Tracked question

A question you have decided to be measured on, phrased the way a customer would phrase it, and then held constant so that a change in the number is a change in the world.

Check

One tracked question, asked of one engine, at one scheduled time. The check is the denominator of everything downstream.

Appearance rate

Checks that named you, divided by checks run, published with both numbers and never as a bare percentage.

AI Visibility Score

The name Apex publishes appearance rate under. The same counted ratio, always with its counts beside it. Not a composite index, and never a score out of a hundred.

Why your analytics cannot tell you this

An AI answer is composed for one person and then discarded. If the assistant names you and that person calls, your analytics records a direct visit or a phone call with no referrer. If the assistant names a competitor instead, your analytics records nothing whatsoever, and the second case is the one you most need to know about.

There is no report you already own that contains this. It has to be elicited on purpose.

What honest measurement requires

• A question set that stays fixed. If the questions drift, the trend is a fact about the questions rather than a fact about the business.
• Every engine that matters, not the flattering one. Apex asks ChatGPT, Gemini, Perplexity and Claude. They disagree often enough that a single-engine number will mislead you in one direction or the other.
• A schedule rather than a spot check. Assistants are non-deterministic; one answer is an anecdote, and the same answer asked repeatedly is evidence.
• Counts, not scores. Record whether the answer named you, then publish the numerator and the denominator together.
• A record of who else was named. The names that come back instead of yours are the most actionable output of the whole exercise.

What a trustworthy number looks like

It carries its denominator rather than standing alone as a percentage. It is broken out per engine, because an average across engines can hide being completely invisible on one of them. It has a date on it. And it is not a composite score out of a hundred. The moment several things are blended into one index, you can no longer tell which thing moved, or whether anything did.

The free scan is deliberately smaller than what a tracked client runs. It is enough to establish whether you appear at all, and not enough to call a trend. We say so on the report rather than dressing it up.

How long before the number means anything

A single morning is a snapshot. Assistants are non-deterministic, and one morning can look unusually kind or unusually brutal for no reason connected to your business. A run of checks is the first point at which a change is worth acting on. It is also why Apex will not show a share-of-voice figure until enough evidence sits behind it for the number to hold still.

What to do with the number once you have it

• Read the per-engine split before the headline. A zero on one engine is a bigger finding than a good average.
• Read the competitor names before the trend. They tell you which sources the engines are leaning on for your area.
• Change one thing at a time. Checks that repeat often enough let you attribute a movement to something; occasional ones leave you guessing.
• Read the trend, not the morning, and give it longer than feels necessary before you conclude anything confidently.