Good question, and you're right that identical runs move a lot. In our 14-day benchmark, 36% of ChatGPT answers and 68.8% of Perplexity answers changed day to day. What we do: track 20-50 prompts per topic and run each one daily, then judge share of voice on a 7- or 14-day rolling window rather than a single run. Any one prompt is noisy, but the aggregate across many prompts and days settles down fast. If a topic only has a handful of prompts, we'd want at least two weeks of data before trusting a shift.

Those Perplexity numbers line up with what I saw, it barely gave the same answer twice. A 14 day window with 20+ prompts sounds like the minimum before anyone should report a change to a client.