Methodology

What we ask

We ask the questions your customers actually ask — hundreds of them, across ChatGPT, Gemini and Perplexity, repeated weekly. One-off checks lie; calibrated series don't. Every answer is stored as text, timestamped, and every published number links back to the answers behind it.

How we experiment

We change your catalog the way a lab would: matched product pairs, one variable at a time, success metric locked before anything ships. You get the difference between treated and control — with confidence intervals, not adjectives. If nothing moved, the report says so.

Our parser's measured error rates

Sensitivity 96.1% — of the brands actually mentioned in an answer, this is the share our parser finds. Specificity 100.0% — of the brands not mentioned, this is the share it correctly leaves out. Measured on 800 labelled cells (parser v1).

2 of these labels are human; a human label always overrides a machine one.

Measurement surface

Today we measure through the engines' APIs, with the same questions and the same protocol every week. An API answer and the answer a shopper sees in the consumer app are not always the same: the consumer surface may apply different retrieval, personalization, or a different model version. We say so rather than implying our numbers are the shopper's exact experience.

On the roadmap: a consumer-surface protocol across three surfaces, with model version, build and geography recorded for every observation, so the gap between the two can itself be measured instead of assumed.

Independence

Today we measure with standardized, repeated benchmarks across ChatGPT, Gemini and Perplexity — same questions, same protocol, every week. Our architecture is engineered toward engine-independent measurement: first-party crawler telemetry, consented shopper panels and checkout attribution, so our customers' history never depends on anyone's API terms.

No customer can buy their way into our index — how the index is governed.