Your AI Visibility Score Isn’t Wrong. Your Ruler Is.

The same brand, five different numbers

Run the same brand through three AI visibility tools in the same week and you will get three different scores. One says your brand appears in 12% of relevant AI answers, another says 31%, a third reports a “visibility index” of 48 with no unit at all. The instinct is to ask which tool is broken. Most of the time none of them is. They are measuring different things and printing the same word — visibility — on the result.

So how do you measure brand visibility in AI answers when every tool disagrees? You stop measuring the score and start measuring the ruler: the prompt set, the engines, the sample count and the time window. A visibility score is a downstream number. Change any one of those four and the score moves without your brand doing anything at all.

This article is about making that ruler explicit, so that two numbers — from two tools, or from two months — can finally be compared. Nothing here needs a new tool. It needs a one-page measurement card and an afternoon.

How to measure brand visibility in AI answers: the five variables

Every AI visibility number is produced by five choices, and most of them are invisible in the dashboard.

Variable	What it changes	How to align it
Prompt set	The denominator. A score is “appearance rate across these questions”; swap ten prompts and the score moves.	Freeze one panel of eight to ten questions and use it across tools and months.
Engine	ChatGPT, Perplexity, Gemini and AI Overviews retrieve differently and answer differently.	Report one score per engine. Never publish an all-engine average as “your visibility”.
Region and language	The same question retrieves different sources in German than in English.	One panel and one score per market. No cross-market averages.
Sample count	Generative answers vary run to run. One sample is an anecdote; twenty is a rate.	At least five samples per question per engine, reported as a range, not a single point.
Time window	Answers drift as new pages get indexed.	Fix a seven-day window, and compare months with windows of the same length.

Two honest tools can disagree by a factor of three when they sit on different rows of this table. That is not a data problem. It is a specification problem, and it is fixable without changing anything about your content.

Three things every AI visibility number must declare

Before you compare two numbers, ask each one to declare three things. A number that cannot declare them is a mood, not a measurement.

Declaration	Why it matters	What to do when you get it
The panel — which questions were asked	It is the denominator of the whole score.	Ask for the full prompt list. If a tool will not export it, do not compare its absolute numbers with anyone else’s.
Sampling and window — how many runs, over how long	One run and twenty runs are two different instruments.	Accept only results labelled with sample count and date window. Treat a single snapshot as a trend hint, never as a baseline.
The base — which brands count in the denominator	Share is meaningless if the competitor set silently changes.	Fix three to five competitors and keep the set for the whole quarter.

A number that declares its panel, its sampling and its base is comparable. A number that declares none of them is a screenshot with a percentage sign.

Share of AI voice: comparing your visibility against competitors

Position is a poor unit in a generated answer. There are no fixed slots, the answer is prose, and the same brand can appear first in one reply and seventh in another without any ranking rule you can audit.

Share of voice replaces position with a ratio. Define it once, then freeze the denominator:

AI share of voice = your mentions ÷ total mentions of every brand in your fixed competitor set, on the same prompt set and the same window.

That ratio survives model upgrades and content changes, which makes it the one number that can be compared month over month. It also answers the question buyers actually ask — where does my brand stand in AI search next to the four brands I lose deals to? — with a figure instead of an anecdote.

Keep the competitor set at three to five names and do not change it mid-quarter. A share of voice that quietly redefines its base is worse than no share at all. Our weekly competitor benchmarking method is built on exactly this frozen base.

A 5-layer prompt panel that standardizes your ruler

This is the copyable part. Run the five layers in order, once a month, on your own domain plus the competitor set. Each layer produces one artifact you keep.

1. Definition — “list the eight questions buyers in my category actually ask about this product, in their own words, not marketing language.” Artifact: the eight-question panel, frozen for the quarter.
2. Coverage — “for each of these eight questions, tell me whether my brand is mentioned, and which of these four competitors are mentioned.” Artifact: a mention matrix, brand against question.
3. Calibration — “for every result above, report the prompt used, the engine, the region, the number of samples and the date window.” Artifact: the measurement card, one page.
4. Share — “calculate my share of mentions against these four competitors on the same eight questions, and show me the denominator.” Artifact: one share-of-voice number per engine.
5. Drift — “compare this month’s answers with last month’s, question by question, and list what changed and which page was cited.” Artifact: a change log, not a score.

Notice what the panel deliberately does not do: it does not average across engines, and it does not compress your brand into a single number. Both operations destroy exactly the information you need in order to act.

Fix the ruler in one afternoon

1. Export whatever history you have and label every row with panel version, engine, region, sample count and window. Rows without those labels get archived, not compared.
2. Freeze the eight-question panel and the competitor set for the next quarter.
3. Re-run the panel once, on every engine you care about, with at least five samples per question.
4. Write the measurement card and keep it beside the dashboard, so anyone reading a score can read its conditions in ten seconds.

After that afternoon, a drop from 20% to 9% means something, because nothing else moved.

When the difference is real

Not every gap is cosmetic. Once the panel and the window are aligned, a remaining gap usually has one of four causes, and each one needs a different response.

What you see	Most likely cause	Correct move
Two tools differ by a factor of two	Different panels — one uses brand names, the other uses category questions	Align the panel and re-run. Until then, never compare absolute values.
Your score halves month over month	Model version change or a shift in retrieved sources	Check the engine version and the window first, then look at content.
The score rises but pipeline does not	Growth concentrated in low-intent questions	Split the panel by intent and score each layer separately.
A competitor jumps sharply	They shipped a page that now gets cited on several questions	Trace the cited page, then decide whether to compete on content or on the question.

The rule behind the table: cosmetic gaps are fixed by specification, real gaps are fixed by content. Rewriting pages to fix a panel mismatch is the most expensive way to change nothing.

Two scores, one decision

You will still end up with more than one number — one per engine, one share, one mention rate. That is correct, and it is not a problem to be averaged away.

The decision rule is short: use the number whose failure would hurt you most. If revenue comes from buyers who ask one specific comparing question, judge yourself on the share of voice for that question set, not on the brand-wide average. A brand-wide 40% concentrated in questions nobody asks before a purchase is a vanity number wearing a dashboard.

What to fix first

In order: panel, engine separation, sample count, window, share base. Fix them out of sequence and you end up re-running the panel every time, which is the real reason most teams give up and buy another tool instead.

If you want the measurement to survive a change of tool, keep the panel and the card in your own repository rather than inside the vendor’s product. Then a new tool is simply a new ruler applied to the same panel, comparable from day one. If you would rather have the panel, the card and the share-of-voice reporting built and maintained for you, that is what our AI answer monitoring service already does.

A visibility score is not a measurement of your brand. It is a measurement of the questions someone decided to ask. Fix the questions first, and the score becomes an instrument instead of an opinion.

Related reading

• How to Build a Prompt Panel That Actually Tracks Your Category
• Competitor Benchmarking in AI Answers: A Repeatable Weekly Method