Skip to main content
Narrative measures your brand in two places. On Google, the unit is a keyword, and keywords form topics. In AI assistants, the unit is a prompt: a question Narrative asks ChatGPT, Perplexity or Gemini on your behalf and reads the answer to. Prompt mapping is the bridge between the two. It says which of your topics each prompt is about, so a single topic can carry a search reading and an assistant reading side by side, and so a campaign for a topic knows which prompts it should move.
A prompt fans out into three sub-queries; one matches a keyword exactly and maps to its topic, the others are judged against the topic list with a probability bar each.A prompt fans out into three sub-queries; one matches a keyword exactly and maps to its topic, the others are judged against the topic list with a probability bar each.

A sub-query that is one of your keywords maps the prompt outright. The rest are judged against every topic.

​
Two methods, in order

Every active prompt is mapped to the topics of its own language. A prompt is about as many topics as the evidence supports: on a typical brand about half of the prompts map to one topic, most of the rest to two, and a few to three or four. There is no fixed number.
1

Exact match through fan-out

When an AI assistant answers a prompt it usually searches the web first, and to do that it breaks the prompt into a few shorter searches. Narrative records those sub-queries. When one of them is, word for word, one of your clustered keywords, that is not an opinion: the assistant searched your keyword to answer the prompt, so the prompt is mapped to that keyword’s topic at full confidence. Topics that several sub-queries hit rank first.
2

Judged coverage for the rest

For the topics fan-out did not settle, the prompt is judged against each one, reading the topic’s name and the ten most-searched keywords in it. This happens in two stages. First, every topic is asked on its own whether the prompt is about it, and only topics that clear a probability of one half stay on the shortlist. A prompt that is about none of your topics, a question about corporate tax rates to a payments brand, leaves the shortlist empty rather than being forced onto the nearest topic. Second, the shortlist is judged together, with one probability budget shared across it, which is what stops a broad topic from claiming every prompt in sight. The leader is kept, together with any topic that took at least 40% as much of the budget as the leader.
The second stage matters. Judged one at a time, a vague topic with two near-identical keywords can look like a fit for almost any prompt, and a specific topic with fourteen keywords and real volume can be passed over. Judged together, breadth and precision compete. Measured against a person reading the same prompts, the links a person would reject went from 17 to 0. A prompt in a language none of your topics are in, or a mixed-language prompt, is judged against your primary search language. A prompt nothing covers stays unmapped and is listed as such; Narrative does not invent a topic to hold it.

​
Measured and content-only topics

Mapping decides which topics carry an assistant reading.
  • A topic with at least one mapped prompt is measured: its mention rate and citation rate come from the answers to those prompts, averaged over every mapped prompt the latest scan asked, not only the ones that happened to name you. A prompt nobody has asked yet leaves the denominator rather than counting as a loss.
  • A topic that was judged and that nothing mapped to is content-only. That is a measurement: people search it, no prompt Narrative asks is about it, so its only reading is from Google. It is a good candidate for a prompt you add yourself.
  • A topic in a language nothing judged, because you have no prompts in that language yet, is not measured, which is different from content-only and shown differently.
A rebuild that finds no evidence for a topic does not quietly switch it off, and one that finds some does not quietly switch it on; the mapping writes the evidence and leaves your decisions alone.

​
What you can change

  • Add prompts to a topic from the topic’s record, or add a topic to a prompt from the prompt’s own page. A link you make is yours: no mapping run reads it, rewrites it or removes it, and adding is additive, so a prompt keeps the topics it already had.
  • Deactivate a prompt and it leaves every topic it was mapped to.
  • Add a prompt and it is mapped on the next run.

​
Where it shows

On the Prompts page: the lead topic each prompt is mapped to, the one the mapping was most confident about, with a count of any others, and “unmapped” where nothing covers it. On a topic’s record: how many prompts are mapped to it, and the mention and citation rates those prompts produce. On the Competitors tab: for every rival, the prompts and keywords it appears in, which is the same mapping read from the other side.