Context
A 2026 research preprint examines repeated observations of AI-search visibility and finds that outputs vary across runs, prompts and time, making single observations unreliable for assessing brand or source visibility.
What changed
The study formalizes AI-search visibility as a probabilistic measurement problem and argues for repeated sampling rather than treating a single generated answer as a stable ranking.
Why it matters
AI visibility monitoring and GEO experiments need uncertainty-aware repeated observations before changes are interpreted as meaningful. Snapshot measurements can confuse stochastic variation with real visibility movement.
Visibility OS interpretation
Visibility OS interprets this as evidence for measurement-laboratory design: retain repeated runs and distributions, and distinguish observed variation from intervention effects. Study-specific sampling thresholds should not be universalized.
What remains uncertain
How do required sample sizes change by engine, language, geography, query class and time? How much observed variation comes from retrieval versus generation? Which uncertainty measures best support longitudinal intervention experiments?