Measuring AI Visibility: Repetitions, Waves, Noise Band

One measurement is no measurement. Research from the University of St. Gallen (2026) shows: repeat the same query on the same day and the cited sources overlap only 32 to 43 percent. AI answers are probabilistic. Serious measurement therefore needs a frozen question set, several repetitions per question and engine, measurement waves across multiple days, and a declared noise band: small shifts of a few points are noise, not change.

The causality ladder sets the language of the report. Strongest rung: the engine demonstrably cites the page we published as a source. Middle rung: publication, indexing and the rise in mentions coincide in time, and server logs show the AI crawlers' visits. Weakest rung: mere correlation. An honest report names the rung and keeps a confounder log: model updates, algorithm updates, seasonality, parallel marketing activities.

Realistic timelines instead of promises: content changes typically surface in AI answers after days to weeks; first citations usually take several weeks, robust visibility takes months. Anyone guaranteeing results within 30 days is promising something that can neither be controlled nor measured seriously. What was measured gets reported; nothing gets guaranteed.