A mention rate is one number over fifteen questions. The per-prompt breakdown is where it becomes something you can act on — starting with the two readings people find hardest to interpret: none of them, and all of them.
A mention rate is one number stretched over fifteen questions. Northwind Coffee’s is 51% — and 51% doesn’t tell you what to write on Monday.
The per-prompt breakdown splits that number back into the questions it came from, and shows which models named you on each one. Fifteen rows, five dots a row. It is the view where a percentage turns into a sentence you can actually do something with: nobody names us when someone asks for a coffee subscription.
One row per tracked question, sorted worst first — so the questions you lose are the ones you see, and the ones you already own are the scroll. Each row carries four things:
One thing catches people out: the timeframe picker doesn’t apply here. A breakdown is always the latest run — averaging a question across runs would blend answers to a question that may not have existed for all of them.
If every row reads 0%, the card is saying one thing plainly: across fifteen buying questions and five models, not one answer volunteered your name. That is a real reading rather than an error — but it has five different causes, and they don’t have the same fix.
And sometimes 0% is simply true. We keep a brand in our own test set that does not exist — an invented electrolyte company, with an invented name, run through the same pipeline as everything else. In its last run it was named in 0 of 50 answers: every question at zero, on every model. That is the control. It is how we know the scoring isn’t flattering anybody, and it is why a 0% on your card is worth taking seriously instead of filing as a glitch.
It is also, awkwardly, the most useful state this card has. A board of zeroes is unambiguous: there are fifteen questions, they are ranked, and there is no argument about where to start.
Type your company name into ChatGPT and it will very likely tell you about your company. Then look at a card full of zeroes and conclude the tracker is broken.
Both things are true, because they are different questions. Asking “what is Northwind Coffee” tests recall. Asking “what’s the best coffee subscription” tests recommendation. A model can hold a whole paragraph about you and still never put you in front of a buyer who didn’t already know your name — and the buyer who already knows your name was never the one you were trying to reach.
That is why none of the fifteen questions on this card contains your brand. A question that names you can only test whether the model recognises it, which is a much easier test and a much less interesting one.
The pleasant version of the same problem: a card with nothing on it to act on. Three things it might mean, in rough order of likelihood.
Widening the question set is a setup decision rather than a dashboard one: prompts are generated from the brief you write when you create a tracker, and you edit them before the first run. A wider market means a brief that describes one.
Take the top row and treat that one question as the whole job. An average over fifteen questions isn’t something you can act on; a question you lose is.
The dashboard’s How to Improve page runs these same rows through that method for you: it names the question you’re invisible on, the rivals the answer hands it to instead, and the sources they came from. The longer version, including the tactics that don’t work, is in how to show up when someone asks ChatGPT for a recommendation. If you want the headline number explained first, start with AI mention rate, explained.
Usually because nothing outside your own website corroborates you. AI models name brands they find in sources they already trust — directories, roundups, review profiles, forum threads — and rarely name a business on the strength of its own homepage. The other two common causes are questions that are broader than your business, and questions aimed at the wrong market.
No. We run an invented brand through the same pipeline as a control, and in its last run it was named in 0 of 50 answers — so a zero is a real reading, not a failure. Check the questions first: if they are broader than your business or pointed at the wrong country, the zero is correct but useless. If they are the questions your buyers ask, the zero is correct and worth acting on.
No, those are two different tests. Asking a model about your company by name tests recall; asking it “what’s the best X” tests recommendation. A model can hold a detailed paragraph about you and still never volunteer your name to a buyer who has not heard of you. That is why none of the tracked questions contains your brand name.
Usually that the questions are narrower than your market — a question only you could win measures how specific the question was, not how visible you are. Across every real brand in our own test set, none has all of its questions at 100%. Once the rows are full, the measures that still carry information are where you place in the list, who is named alongside you, and how you are described.
A hollow dot means that model answered the question and did not name you — that is a finding. A dash means it returned no answer at all, so there is nothing to read into it. They are drawn differently on purpose: a missing answer and a lost answer are not the same result, and merging them would quietly turn a partial failure into bad news about your brand.
Weeks, not days. A new source has to be published, then crawled, then absorbed into what the models draw on. The whole prompt set re-runs every week, so what you are watching for is three runs moving the same way — these models are not deterministic, and one run’s movement is noise about as often as it is progress.
Fifteen prompts, five models, tracked every week — listed worst first, with the models that named you and the sources the winners cited.