Mentioned Is Not Recommended

The first metric every AI visibility programme adopts is mention rate: of the prompts we track, in what share did an engine say our name?

It is the right place to start. It is binary, unambiguous, and it answers the question a nervous executive actually asks. It is also, past the first month, the metric most likely to make a bad position look like a good one — because it awards identical credit to two outcomes that are not remotely equivalent.

Consider two answers to “best project management tool for small teams.”

A. For small teams, Acme is the strongest option — it’s built specifically for teams under 20 and handles onboarding without an admin.

B. There are many options depending on your needs, including Asana, Monday, ClickUp, Notion, Trello, Basecamp and Acme.

Mention rate scores these the same. One is a recommendation. The other is a disclaimer with your name in it.

Position is not a vanity distinction

The instinct is to treat this as splitting hairs — a mention is a mention, and surely the user reads the whole answer.

Mostly they do not, and there is reason to think the model does not weight the whole answer evenly either. The best-known result here is Lost in the Middle, which found that language models use information placed at the beginning and end of a long context far more effectively than material buried in the middle — performance degrades measurably for content in the middle of a long input. That study is about how models read context, not about how users read answers, and it would be overreaching to present it as direct evidence about answer position. But it does establish that position within a block of text is not neutral to these systems, which is the weaker claim this section actually needs.

The stronger argument is behavioural and hardly needs a citation: a seven-item list is a non-answer, and users treat it as one. They either take the first name, ask a follow-up, or leave. Appearing seventh in a list buys you a mention and roughly nothing else.

A taxonomy that survives contact with real answers

Counting stops being useful the moment you accept the above. What replaces it is classification — sorting each mention into a position that carries a different meaning.

First mention. Your brand is the first named in the answer. This is the closest thing to the AI-era featured snippet, and in short answers it is frequently the only one a user acts on.

Prominent mention. You are named with substantive framing — a reason, a fit, a differentiator — regardless of ordering. “Acme is the strongest option for teams under 20” is prominent even in third place, because it carries an argument. This is often more valuable than a first mention with no reasoning attached.

Listed mention. You appear in an enumeration with no distinguishing detail. Present, undifferentiated, easily substituted. This is the largest bucket for most brands and the one most likely to be mistaken for progress.

Absent. The engine answered and did not name you. Worth separating from the fifth case below.

No usable answer. The engine failed, refused, or returned nothing relevant. This is not an absence — it is a missing observation, and folding it into “absent” quietly inflates your denominator and depresses your rate for reasons that have nothing to do with your brand.

That last distinction matters more than it looks. Errored answers should be excluded from the calculation entirely, not counted as failures. A scan where two engines timed out is a scan with less evidence, not a scan with bad news.

What the distribution tells you that the rate cannot

Once mentions are classified, the shape of the distribution becomes the diagnosis.

Heavy listed, light prominent. The engines know you exist and cannot articulate why you would be chosen. This is a positioning problem expressed as a visibility problem. The web contains your name and not your argument — usually because your own material describes what you do rather than who you are the best choice for, and because third-party sources have nothing sharper to draw on. The remedy is not more content; it is a clearer, more consistently repeated claim about fit.

Prominent on narrow prompts, absent on broad ones. Common and healthy for a specialist. “Best CRM for solo real-estate agents” names you; “best CRM” does not. That is category-leadership economics, not a defect, and burning budget attacking the head term is usually the wrong call. Branded vs unbranded AI queries covers reading the split.

First mention on some engines, absent on others. Almost always a sourcing difference rather than a quality difference — one engine is grounding in live retrieval where you are well covered, another is answering from weights that predate your presence. Multi-engine monitoring exists to make that split visible instead of averaging it away.

High rate, all listed, no prominence anywhere. The uncomfortable one: you have maximised the metric you were measuring and gained little. This is the specific failure mode of optimising for mention rate.

Drift is the signal, not the level

A single measurement of position is a snapshot of a non-deterministic system. The same prompt run twice can return differently ordered names with nothing having changed — which means one reading of “we moved from second to fourth” is noise.

What is not noise is a sustained change in the distribution across many prompts and repeated scans. That is position drift, and it is the earliest available warning that something has shifted — a competitor’s campaign landing, a model update rewriting an entire category’s defaults, a piece of your content ageing out of relevance. It usually shows up in the position mix well before it shows up in mention rate, because you get dropped from prominence long before you get dropped from the list.

Reading drift requires enough samples to distinguish it from resampling, which is a real constraint and covered in how many runs before you trust an AI visibility number.

The counter-argument

The honest objection: this is more complexity than most teams need, and mention rate has the enormous advantage of being explicable in one sentence to someone who does not care about any of this.

That is fair, and it is right about reporting. Mention rate should stay on the executive summary — it is the correct altitude for that audience, and replacing it with a four-bucket distribution in a board deck is how a measurement programme loses its sponsor.

The argument here is about the working metric, not the reported one. The team doing the work needs the classification, because mention rate cannot tell you whether last quarter’s effort produced twelve new listed mentions or three new prominent ones — and those are wildly different outcomes for identical movement in the headline number. Report the rate; work the distribution.

There is a second objection worth conceding: classification requires interpreting prose, and interpretation can be wrong. An answer can be ambiguous about whether framing is genuinely favourable. That is real, and it is why the underlying answers stay readable rather than being discarded once scored — a classification you cannot audit against the source text is a number you should not defend.

Where to start

Take your last scan. Ignore the score. Open the twenty prompts that matter most commercially and sort the answers into the five buckets by hand — it takes under an hour and it is the fastest way to find out whether your visibility is real.

Most teams doing this for the first time find the same thing: the rate is healthier than expected and the prominence is much thinner. That is not bad news. It is a far more tractable problem than absence, because being in the consideration set and being the recommendation are separated by a much shorter distance than being unknown and being known.

Answer engine ranking is where this classification lives in the product, and what is answer engine ranking is the reference definition if you want the mechanics first.