Spectral Brand Theory · measurement layer
The Brand Spectrometer reads how different audiences actually perceive a brand — audience by audience, across eight dimensions, from public material — and tells you which of those differences are real and which are just noise.
The problem
The instrument's scale
A brand is one signal made of these eight. Two audiences can hold the same overall opinion while reading those dimensions completely differently — the same score hiding two different brands. Toggle the audiences below to see it.
Each dimension carries two readings. One is how strongly it registers — the score, 0–10 (the line's height). The other is which way it leans — its valence:
Strength is not approval: an audience can read a brand as intensely economic and feel negative about it. On the spectrum strip the dot slides from negative (left) to positive (right) along each line — a line sitting high with its dot on the left is the dangerous one: an audience that cares a lot, and not in your favour.
A real read
A spectrum strip: eight wavelength lanes, one per dimension. Each audience is a line at its score (height = how strongly the dimension registers), the dot on the line is its valence (left −1 negative … right +1 positive), and the band is the spread across audiences. Valence is floored: where the lean sits below the instrument's noise floor the dot turns hollow and grey at centre — the read is sub-resolution, so the tool abstains rather than fake a direction. No radar — a spider chart's silhouette is an artifact of axis order, not the brand. Turn the detail dial up to add sentiment, uncertainty and the noise floor; each step says what it still hides. How to read this →
Who agrees · who doesn't · what's real
Hover any cell to see the pair, the dimensions driving their distance, and whether the gap clears its own noise floor. Flip the switch to grey out every sub-resolution pair — what remains is real.
Distance is 1 − cosine similarity between two cohorts' eight-dimension vectors. The noise floor is the wobble from swapping AI operators alone. Above the floor = real; below = we can't tell.
In this worked example, the owners ↔ online-debaters split is the largest and is resolved; several press-to-press gaps sit below the floor and are honestly reported as sub-resolution, not findings.
The read, in numbers
Hover a row to highlight that cohort in the chart above. These are the raw vectors the whole read is built from.
The managerial reading
It's an instrument, not a slideshow
Bring your own atlas — the whole read re-renders for it. Nothing is uploaded: every atlas is processed entirely in your browser.
Currently loaded —
Bring the result back — pick how:
Verification equipment — the PRISM instrument family, three published measurements about instruments of this kind: what a single blended score destroys (PRISM-M); whether stated readings predict agentic choice (PRISM-C — a companion instrument; choice is not claimed by this meter); whether readings survive model-version changes (PRISM-T — the version floor readings are epoch-stamped against).
Code open-source under MIT; theory published under CC BY 4.0.
Measurement instrument for how a brand is perceived — not a verdict on what it "is." Cohort metameric variance is the measurement; there is no presumed ground truth. No brand assets reproduced (nominative use).