How we measure
Being named is not the same as being recommended. Here is what we ask, what we count, and — the part almost nobody writes down — what the report refuses to tell you.
“Are we mentioned?” is the wrong question
A search engine returned ten links and left the judgment to the reader. An assistant returns one answer, having already done the choosing, and the runners-up are not shown.
That is the change everybody notices. The one that matters more is that an assistant does not merely include you or omit you — it takes a position on you. It can name you flatly, recommend you among others, single you out as the best choice, or name you while steering the reader elsewhere.
So a brand that appears in 60% of answers and is recommended in 4% of them has a completely different problem from a brand with the reverse profile. One has a reputation problem, the other a presence problem, and they need opposite work. A tool that reports only whether you were mentioned reports both as the same number.
Five numbers, and each says what it is a share of
Every metric names its own population inside its own sentence. Two of these are shares of different things, and percentages without denominators look contradictory even when both are correct.
- Unprompted appearance
- Of the answers to questions that never name the brand, how often it comes up on its own.
- Named without endorsement
- Of the answers where the brand appears and the stance could be read, how often it is named but neither recommended nor criticised.
- Actively recommended
- Of the answers where the brand appears and the stance could be read, how often the assistant recommends it.
- Named the best option
- Of the answers where the brand appears and the stance could be read, how often the assistant singles it out as the best choice rather than one of several.
- Share of brand mentions
- The brand's share of all brand mentions — its own and every tracked competitor's — across answers to questions that never name it.
A question that names your brand guarantees the mention it would otherwise be evidence of, so branded questions are excluded from the two presence metrics. They are kept in the three stance metrics, because how an assistant speaks about a brand is a real signal however the brand got into the answer.
A percentage three questions could move by a third is not a percentage
Break a report into slices — by question type, by buying stage, by the rivals a question names — and each slice is a smaller sample than the one before it. At some point one question decides the number.
So a slice of the coverage breakdown is shown as a rate only when at least five questions sit behind it, and no single question is worth more than a fifth of the claim. Below the floor the report shows the counts it honestly has and says how many further questions the slice needs. Not a greyed-out percentage — counts, and a number to act on.
The floor governs the sliced breakdown rather than the five headline metrics, which divide by answers and describe behaviour across the whole bank instead of making a claim about one narrow cut of it.
The floor counts questions, not answers. Three assistants answering one question is one question, not three — counting answers would multiply every slice by the size of the model panel and clear the floor while measuring nothing more.
“Not enough data” is never zero
A metric with nothing behind it is reported as unknown, never as 0%. Printing zero would tell a brand it is invisible when nothing was measured. And because “unknown” alone is barely better, the report says which kind of nothing it is — the distinction that matters most sitting under the three stance metrics.
The brand never came up
There was no appearance to read a stance from. The work is presence: you are not in the conversation at all, and nothing about tone or framing is the problem yet.
It came up, but the stance could not be read
The brand appeared and the answer took no readable position on it. You are in the conversation and being passed over inside it — a different problem, with different work behind it.
The questions are the instrument
Most tools in this space inherited their frame from SEO: crawl the site, score the pages, recommend content changes. For a brand sold in an aisle that produces numbers a marketing team cannot act on. The instrument is the category conversation — what an assistant says when somebody asks about the situation the brand competes in, without the brand being named.
Category entry points
The Ehrenberg-Bass framework for mental availability, from Byron Sharp and Jenni Romaniuk: the seven situations that bring a category to mind — why, when, where, while, with whom, with what, how feeling. It surfaces situations nobody produces by free association.
The means-end ladder
Climbing from an attribute, to what it does for the person, to what that outcome means to them. The last rung names no product at all, and it is where assistants answer with advice instead of a recommendation. That absence is the finding, and a bank written at the first rung cannot see it.
One need at three distances
The same buyer need in category words, in the buyer's own words, and as a full situation. Not paraphrasing, which buys nothing — different retrieval, and the effect is large. A bank written entirely in category vocabulary says almost nothing about buyers who do not speak it.
Where a question would need words you did not supply, none is written: an invented answer would be recorded as though it came from you. Today every generated question is recorded as resting on what you told us, not on measured demand, and the product says so rather than implying otherwise.
A score that moved is only your doing if the same model answered both times
Model names are aliases, and vendors repoint them silently. The string you asked for in March and the string you asked for in May can be identical while the model that actually answered is a different snapshot with different opinions about your category.
So what gets recorded is what actually served each answer, and a comparison between two scans returns three verdicts, never two.
Matched
Both runs recorded what served and the sets are the same. This is the only verdict under which a change may be attributed to the brand.
Drifted
The sets differ. The report separates the two causes you must not confuse: a vendor repointed an alias underneath you, or you changed your own configured panel.
Unknown
At least one run did not record what served, so nothing can be ruled out. This is not a degraded “matched” — rendering it as “no change” would be the same mistake in a place customers read.
The longer argument, including how a bank is sized and reviewed, is in the methodology: read the measurement methodology