Three tools, three different numbers

If you are evaluating AI visibility platforms, you have probably already hit this. Two trials running against the same brand, in the same week, reporting numbers that do not reconcile. The instinct is to assume one of them is broken.

Usually neither is. They are answering different questions and giving the answers the same name.

We went through the public documentation of several tools in this category and lined the definitions up. What follows is quoted from their own docs, with links, so you can check us.

Share of voice

Two definitions, both reasonable, that produce different numbers from identical data.

Why that matters: imagine two brands across ten tracked answers. Brand A is named five times inside two answers. Brand B is named once each in six answers.

Counting mentions, A leads. Counting the answers you appeared in, B leads by a distance — and B is almost certainly in the healthier position, because it is reaching more separate conversations.

Neither definition is dishonest. But “our share of voice is 34%” is a meaningless sentence until you know which denominator produced it.

Citation

This is the one that causes the most confusion, because three things happen in sequence and the category uses one word for all of them.

1. Retrieved The model pulled your page while assembling the answer. The reader never sees this.
2. Cited Your page appears as a visible source attached to the answer.
3. Mentioned Your brand name appears in the text of the answer — which can happen with no link to you at all, and can happen in an answer whose citations all point elsewhere.

Peec AI’s documentation is unusual in separating the first two, and in being explicit that they are averages rather than percentages: retrieval rate is “average number of times a URL from this domain appeared as a source per chat,” citation rate is “average number of times the domain was explicitly cited when used”.¹ Both are averages per chat, not percentages.

Profound’s glossary defines citation share differently again: “how often a specific page or domain is cited as a proportion of all citations.”² That is a percentage of a total, not a per-chat average. Ansvisor’s glossary lists citation frequency as “how often a source appears in AI-generated responses” without separating retrieval from citation at all.³

So “citation rate: 0.8” and “citation share: 0.8%” and “citation frequency: 80” can all describe the same domain in the same week.

Position

Both Peec and Profound define position as a rank within an answer — where you sit relative to the other brands named in the same response.¹²

That definition has a property worth noticing: it only exists when more than one brand is named. An answer that recommends you alone has no position to report. Depending on how a tool handles that case — excluded, counted as 1, ignored — your average position can move in opposite directions for the same underlying performance.

Ansvisor lists rankings as a product feature but its public glossary does not define the unit, so we could not determine whether it means intra-answer rank or a leaderboard position across tracked competitors.³

Sentiment

Every tool in the category measures it. None of the public documentation we could find explains how.

Peec states a 0–100 scale.¹ Profound describes sentiment as “whether an answer portrays your brand positively, negatively or neutrally, usually reported as a percentage for each.”² Neither publishes the classification method — whether a model judges it, whether humans validate a sample, how mixed or conditional sentiment is handled (“great for small teams, underpowered for enterprise” is both), or what agreement rate the method achieves.

That is not an accusation; sentiment classification is genuinely hard and nobody in adjacent categories publishes this either. But it does mean sentiment is the metric you should hold most loosely, and the one where a five-point move is least likely to mean anything.

What this means when you are buying

1. Compare formulas, not dashboards Two trials will not reconcile. Ask each vendor to write down the numerator and denominator for every headline metric, then compare those.
2. Insist on the underlying answers Any tool worth paying for lets you click a number and read the actual responses behind it. If you cannot get to the text, you are buying a score you cannot audit.
3. Check what happens to single-brand answers Ask specifically how position handles an answer that names only you. The handling is rarely documented and it moves the average.
4. Treat cross-tool comparison as invalid Until the category agrees on definitions, benchmarking one tool's number against another's tells you about the tools, not about your brand. Pick one and track the trend inside it.

Questions we get asked

The short of it

The category is about two years old and is still using borrowed vocabulary. That is normal for a new measurement discipline, and it will settle.

Until it does, the number on the dashboard is less interesting than the sentence that defines it. Ask for the sentence.