Answer volatility measures how much an AI engine's response shifts between observations. Specifically the set of brands it names and the sources it cites. When you ask the same question again, or a slightly reworded one.
It exists because AI answers are probabilistic, not fixed. The model samples a different response each time, live retrieval pulls a different mix of pages, and small wording changes route to different sub-queries. A 2026 study (Zaprev, reported via LinkedIn) found an identical question asked 30 times in a day returned the same brand list only 50-61% of the time, and a synonym swap cut overlap below 30%. An academic preprint measured day-to-day cited-source overlap at roughly 0.34-0.42 (Jaccard). Both are directional, single-method readings.
Volatility is why a single citation-rate snapshot is unreliable. A trustworthy share of voice or citation coverage reading comes from many prompts, sampled repeatedly across every engine, read as a distribution and trend rather than one number.