LEGAL
AI Measurement Methodology
How a visibility number is produced, and why two runs can legitimately differ.
Provider, surface and model
A provider is the company. A surface is the interface measured — an official API, a grounded API or an approved consumer surface. A model cohort is reported only when the provider exposes one.
Measurement origin
Each run records the origin of its evidence, so an API measurement is never presented as a consumer-surface measurement.
Auto-routed and unknown versions
When a provider routes a request internally or does not disclose a version, the run is labelled auto-routed or model version unknown. VENRAC does not guess a version.
Sampling
A metric is computed over a sample of prompts and runs. Coverage is reported alongside the number.
Prompt versioning
Prompts are versioned. A metric is comparable only across runs that used the same prompt version and locale.
Locale and country
Measurement locale is a product dimension and is separate from the language of the interface.
Confidence and volatility
Generative systems vary between runs. Where variance is material, confidence is reported rather than hidden.
Why results differ
Different users, sessions, locales, routing decisions and provider updates can all change an answer. A difference is evidence of variation, not proof of an error.
No ranking guarantee
There is no universal ranking to be optimised, and VENRAC does not promise placement in any answer.