Read the Intent Map — What Buyers Asked, What the AI Did, Where You Were

The Intent Map is the screen where the three intent levels meet. Every answer collected in the window is placed on four dimensions — question intent, answer treatment, your presence, and engine — and the screen reads them top-down: a headline, the flow, four findings, then the per-engine matrix. The point of the map is a delta nobody sees in a blended mention rate: the AI can take a commercial question and answer it with purchase-steering pages, and your brand can be present in that flow or absent from it. That is where the buying decisions happen, and the map shows whether you are in the room when they do.

​

How to find it

Navigate to app.trybluemoon.com and select Intent Map from the left sidebar. Pick a domain; the map renders for that domain over the selected window — 7, 30 or 90 days — and states the number of answers it is built on in its header (N answers · last 30 days). The same reading appears in two other places, narrowed:

• Prompts → a question — How the engines treat this question: the answers to that one question over the last 30 days, per engine, classified by treatment, with cells that read named/total.
• Competitors — a Recommended column: for you and each tracked rival, recommended in X of N answers that named it over the last 30 days.

​

The headline

Two figures sit at the top. Of the answers that steer buyers toward a point of sale, the share that name your brand. These are the answers whose treatment is transactional — at least two retailer sources cited, or prices in the text with at least two brands named. The number of such answers in the window is printed underneath. When no answer in the window steered to a point of sale, the tile says so instead of printing a percentage over nothing. Of the answers that name you, the share that recommend you. This is the brand-role level: Recommended in X of N answers that named you, where N counts only the answers naming you whose roles have already been read. Until at least one has, the tile reads Not measured yet — brand roles are read after each scan. It never reads “0 of 0”.

​

The map

A flow diagram from What buyers asked (left) to How the AI answered (right). Both axes carry the same three labels — Informational, Commercial, Transactional — so the diagonal is the “no drift” line and every off-diagonal ribbon is the AI moving a question up or down the funnel. Each ribbon is split by your presence: blue where your brand is named, grey where it is not. Pointing at a ribbon shows its count. Engines are aggregated away here; the matrix below keeps them. Under the diagram, the screen states what the right-hand axis is made of, because it is the part that is easy to mistake for a judgement: Answer treatment is computed from evidence, not guessed: cited retailer sources, prices in the answer, brands named, and comparison pages cited. Three further notes appear only when they apply:

• N ambiguous answers were refined by an AI read of their text — the evidence verdict is kept alongside, never overwritten. This is the size of the AI-refined band in this window.
• N answers excluded: their question has no intent score yet. Unscored questions are excluded from the map and counted, not silently folded into a bucket.
• Only N retail sources identified so far — the transactional signal leans on knowing which cited sites are shops, and grows as more sources are classified. A quiet transactional column early in an account’s life is a thin table, not a finding.

​

The four tiles

Each tile is one flow of the map, picked by a fixed rule so it does not flicker between visits. The three flow tiles appear together only when every one of their thresholds is met; the fourth, Named, not recommended, is independent of them and appears whenever at least one role-read answer leaves a gap. The fourth tile says “named, not recommended” rather than “listed” or “mentioned”: the map’s cells carry only recommended-versus-measured counts, so it can prove the gap but not which of the other three roles fills it. The answer viewer can.

​

The matrix — by engine

One row per engine, one column per answer treatment. Each cell reads your presence in that engine’s answers of that treatment — the count of answers naming you over the total — and the tint follows your presence; the numbers are the measure. Under each cell where you are named, a smaller figure carries the role level: a filled dot — the recommended mark — followed by recommended / role-read, for example 4/35, with the full sentence (Recommended in 4 of 35 answers that named you here) as its label. When none of the answers behind the cell has a role row yet, the figure is a dash labelled Brand roles not read yet for these answers. The three views — headline, matrix cells, Competitors column — sum the same rows, so if one ever disagreed with another it would be a visible bug rather than a rounding difference.

​

The roles legend

Wherever a brand role is shown — the answer viewer, the Competitors page — the four roles are told apart by weight and shape, never by hue. Colour is reserved for whose brand it is: blue for yours, the neutral foreground tone for a rival. A text label always sits beside the mark. Each role also states its source — ruled from stored evidence or read by AI from the answer text — and, for the evidence case, which cue decided it: a recommendation phrase near the brand, first in a ranked list, a caveat phrase, or no cue at all.

​

The funnel stage

In the answer viewer, every answer that has been fully classified shows a Stage: Purchase, Decision, Consideration or Awareness. It is derived by rule from treatment × roles — transactional answers are purchase; an answer that recommends any tracked brand is decision; a commercial answer, or one that lists a brand, is consideration; the rest is awareness. An answer whose roles have not been read yet shows no stage rather than one computed without them.

​

Windows and cadence

Two things about timing are easy to misread, so they are stated plainly. The map window is yours to pick; the AI refinement and the roles cover the last 30 days. The evidence treatment is computed at read time for every answer in the window you select. The AI refinement of ambiguous answers and the brand roles are written by jobs that look back 30 days. Those jobs only look back 30 days: an answer classified while it was inside that window keeps its verdict and roles afterwards; one that never was is not measured. On the 90-day view, older answers therefore carry their evidence treatment and, unless they were classified in time, no roles — a dash on every role surface, never a zero. The screen fills in as the answers land, not all at once. After a scan, the flow and the matrix are readable immediately from evidence; the AI-refined verdicts and the brand roles follow within about two minutes of each answer landing. The map itself is served from a cache that refreshes within about ten minutes, and the role tiles say not measured yet until rows exist.

​

What it does not claim