A repeatable AI visibility claim needs a frozen question set, provenance and a comparable follow-up.

Define the company or local business, market, category, buyer persona when relevant, geographic context and the buyer-intent question portfolio before the measurement begins. A follow-up cannot silently substitute easier questions and still be called the same benchmark.

Every observation remains attached to the provider, model or measured product surface, prompt version, timestamp and run identity that produced it. API measurements are not relabelled as consumer-interface captures.

AI outputs are probabilistic. Promptence uses the repeat policy frozen for the measurement contract and stores individual observations rather than reporting one lucky answer as a stable market fact. The configured repeat count is part of the evidence.

A timeout, provider failure or unusable observation is not converted into an absence, zero score or competitor win. Coverage and failure state remain visible so the reader can judge how much of the requested benchmark actually completed.

A mention, citation, competitor appearance or source pattern is observed evidence. An explanation for why it happened is a diagnosis. Promptence stores and communicates those as different layers.

An approved recommendation is not treated as an implemented change. Before a before/after claim can be made, the relevant implementation must be verified on the intended public surface when verification is technically possible.

Follow-up measurement uses the same governed demand portfolio and preserves relevant provider policy. If the provider, model, geography or another material part of the measurement context changed, that difference must be disclosed rather than hidden.

A positive visibility delta after a change is a measured correlation. Search traffic, analytics events and verified payments are additional evidence layers. They are not automatically attributed to the intervention as causal revenue.

A public case should expose the measurement contract, evidence window, intervention class, verification state, follow-up basis, material limitations and whether the customer identity is disclosed or anonymized with permission. If the proof chain is not complete, Promptence does not call it a customer outcome case.