How we measure.

Sampling

Fresh sessions, random delay.

Each tracked prompt is run against each of nine engines on the cadence and sample size of the customer's tier (Free n=5/30d → Enterprise n=30/1d). Samples are spread across the cadence window with random delay to avoid temporal clumping. Each sample is a fresh session — no memory carry-over.

Classification

Frozen, versioned classifier.

Every response is classified by Claude Haiku with a frozen, versioned system prompt. Classification returns strict JSON: {level, sentiment, position_index, context, competitors_named}. Holdout agreement: ≥95% on level, ≥90% on sentiment. Classifier version bumps are opt-in re-classification of the prior 30 days.

Confidence intervals

Wilson 95% CI on every rate.

Every citation rate is rendered as a Wilson 95% CI. When n < 10, the bar renders in grey with the label "directional only — sample too small." Two-proportion Z-tests gate every "X is higher than Y" claim at p < 0.05.

Evidence base

Every claim cites a primary source.

Every recommendation in the product cites a primary source from the evidence base (Princeton/Georgia Tech KDD 2024, Ahrefs 75K-brand correlation, Profound 680M citations, Pew, Authoritas, SoCi, Wix 1M citations, Shopify Q4 2025, BrightLocal, SE Ranking, and the rest). The full base is reviewed at every minor release; failed-replication studies are downgraded or moved to the SKIP list.

What we will not do

Hard limits, non-negotiable.

We will not publish a leaderboard with n < 100. We will not score any tactic on the SKIP list. We will not extrapolate the Princeton +41% statistics-tactic figure to production engines. We will not surface a percentage without its CI.

Engine versions tracked (April 2026)

Nine engines. One measurement.