How do you separate real AI-visibility gains from model randomness?

Been in a good thread here about tracking brand mentions across ChatGPT/Perplexity for buyer-intent prompts. One thing that keeps coming up: model responses vary run to run, so a single "yes, we're mentioned" or "no" doesn't tell you much.

Curious how people are handling this, running prompts multiple times and tracking frequency? Logging exact prompt wording to control for drift? Something else entirely?

Trying to build a more reliable method before I trust any of this data enough to act on it.