If you track how your brand shows up in AI answers—ChatGPT, Perplexity, Google AI Overviews, Gemini—you have probably noticed the number moves. Up one week, down the next, with nothing obvious behind it. New research from our co-founder Ron Sielinski, From Stochastic to Stable, explains why that happens and, more usefully, how to tell when the number is solid enough to act on.

Here is the plain version.

The engines don’t answer the same way twice

Ask an AI engine the same question twice—even just moments apart—and you will often get different answers citing different sources. In our testing, even the steadiest engine repeated only about half of its cited sources when we ran the same query again. So, any AI visibility “score” is a snapshot, not a fixed fact, and the ranking built on it wobbles right along with it. The brand at #1 and the brand at #2 can, in the actual data, be a tie.

There is no magic number of prompts

A lot of tools run a set number of queries and call it done. The research tested that idea across 30 platform-and-topic combinations and found the point where a ranking stabilized ranged from about 40 queries to never. And piling on more data helps slowly: every time you double it, the margin of error shrinks by only about a third. So “we ran 50 prompts” tells you nothing, on its own, about whether the answer can be trusted.

The two questions that actually matter

Instead of chasing a query count, the paper says a measurement has to pass two simple checks before you bet a budget on it:

1. Has the ranking stopped moving as more data comes in? Is it stable?

2. Are the gaps between brands bigger than the measurement’s own wobble? Is the signal bigger than the noise?

Only when both are true is the number what we call decision-grade—solid enough to act on.

How to use this as a marketer

• Before you set a goal like “move from #5 to #2,” check that the gap is real. You might already be tied with #2, or further back than the leaderboard shows.

• Don’t celebrate or panic over weekly swings until you know they clear the noise. A 6-point “gain” can just be the engine being an engine.

• Put four questions to any vendor, dashboard, or case study: How many times was each query run? What is the margin of error on each score? Would the ranking hold if you ran it again next week? Which positions are separable, and which are tied?

• Spend content and media budget against numbers that pass those checks, not against a single reading.
  
  

The short version

Make sure whoever hands you an AI visibility number can show their math. If they cannot tell you whether the ranking has settled and whether the gaps are real, the number is not ready, however clean the dashboard looks.

Want to see how your own brand’s numbers hold up under these checks?