how to read Luce
Every term on the board, in plain words.
Luce measures how often AI engines recommend brands. It shows ranges with their uncertainty, publishes only real measurements, and never collapses an engine's answer into a single winner. These are the words it uses.
- Mention rate
- How often a brand is named in answers to Luce’s published question set, with uncertainty estimated across prompt clusters.
- 95% confidence interval
- The range the true mention rate very likely sits in, given the sample. A wide interval means fewer samples, so Luce publishes a range. When two brands' intervals overlap, the data cannot separate them.
- League
- A band of brands the data cannot rank apart. Brands in one league share a shelf. Luce publishes the band and never prints a single position, because an engine's answer varies from one ask to the next.
- Rank range
- The span of positions a brand plausibly holds. Luce never collapses it to one rank. Overlapping ranges read as a tie.
- Measuring
- A category awaiting board publication or an estimate still gathering observations toward its precision gate.
- Provenance tier
- How real a measurement is. L0 is observed in the wild, L1 is a real answer Luce elicited with its own prompt, and the lower tiers are derived or synthetic. Only the real tiers are published, and they are never pooled with synthetic ones.
- Measured surface
- Where a number was observed. The consumer interface (the ChatGPT web app a shopper uses) and the developer API can recommend different brands, so every number states its surface.
- Wave
- One dated round of measurement. A trend needs several waves before Luce shows a significance arrow.
- Divergence
- How much brands’ measured mention patterns or top-league membership differ across surfaces or engines, reported with a confidence interval.
- Effective sources
- How many publishers a set of citations really behaves like, once concentration is taken into account. Ten citations spread evenly over five sites count as five; ten citations where one site takes nine count as close to one. It is the inverse-Herfindahl number of the cited domains, and it counts diversity, so it carries no confidence interval.
- Prompt cluster
- One question from the Luce grammar together with every repetition of it. The n on a cell counts prompt clusters, and the confidence interval comes from a bootstrap over them, because five repetitions of the same question yield one independent fact.
The full method, the estimand, and the provenance ladder are on the methodology page.