Reading a score
A score is a ranking from 0 to 1 on one model, made on one date by one model version. This page covers how to read one, and where its limits are.
What a score claims
Section titled “What a score claims”- A rank, not a probability. A higher score means the person ranks higher on the model than people with lower scores. It does not say a person has a stated chance of acting. Ranking well and giving well-calibrated probabilities are different properties, measured differently (Niculescu-Mizil and Caruana, 2005).
- A band, when used. Models are used as picks: the band of people at or above a measured cut.
- Relative or absolute, by kind. Habit scores are relative to the population. Buying windows are absolute and expire.
- No evidence, no score. A person the model cannot see is left unscored. The product shows them as not covered, never as low.
Audience level and person level
Section titled “Audience level and person level”Browsing is tied to a person, but it is most reliable across many people. Moonbrush reads behavior at two levels and words each one accordingly.
| Level | How to read it |
|---|---|
| Audience group | The reliable unit. A group’s departures from its audience are measured across all of its members. |
| One person | What the browsing attached to that record leans toward. Record facts such as age band or job title are person-level facts. |
Explanations of why a group was formed are group-level, never claims about a single person. The research on intention and behavior finds the link strong on average and loose for any one person (Webb and Sheeran, 2006); see research note 1.
Reading one person
Section titled “Reading one person”The person view shows:
- the scores where this record departs most from the audience around it;
- a why panel with the contributing evidence for each;
- national percentiles, so a score can be read against the whole population.
An example of scored columns
Section titled “An example of scored columns”The shape of score columns in an export. Values are invented, and column names show the shape only.
row_key,scored_on,catalog_version,waits_for_the_markdown_window,answers_direct_response,whole_home_system_windowr_5c1e9a,2026-10-14,2026.10,0.82,0.64,0.21r_7aa310,2026-10-14,2026.10,,0.12,0.77The empty cell in the second row is a model that cannot see that person. It is left empty, not set to zero.