The order is not the gap

A competitor keeps appearing above you in an AI visibility report. Is that enough to change your priorities? “These domains tend to lead” and “this competitor has a meaningful lead over us” are different claims. A settled leaderboard can support the first while leaving the second unresolved.

Ronald Sielinski’s study, From Stochastic to Stable, explains why. Its proposed measurement framework asks both whether a citation ranking has settled and whether differences in citation share are large enough relative to measurement uncertainty. Those are useful checks, but neither automatically answers the narrower question behind a competitive decision: how confidently can we estimate the gap between these two domains?

A leading group is not a resolved competitor gap

The study measured registered-domain citations across ten topics on Gemini, Perplexity and SearchGPT, using ChatGPT-generated query sets. Its two checks did not always finish together. For Perplexity’s “coffee mugs” topic, the ranking met the stability condition after 34 citation-bearing responses, but the test for separation relative to uncertainty did not pass until response 81.

That example shows why a ranking that has stopped moving may still need more evidence. But the separation test has a boundary of its own: it is a practical heuristic for reporting on the leading domain set as a whole, not proof that every neighboring pair is distinguishable. Even after it passes, your domain’s gap with a named competitor may remain uncertain.

The useful distinction is between evidence for a market overview and evidence for a particular comparison. A report can be ready for one without being ready for the other.

Ask for the difference, not just two positions

If a decision depends on a particular competitor’s lead, ask the report provider for the estimated citation-share difference and uncertainty around that difference. A whole-ranking stability score answers a different question.

First, make sure the measure is legible: which questions, platform, topic and collection period does it cover? What counts toward citation share? Does it measure pages, registered domains or brands? The study measured registered domains, not brands. Its uncertainty intervals also condition on responses containing citations, rather than covering citation outcomes across every submitted query.

Then ask how the uncertainty was estimated. A narrow range is not automatically reliable. The study found that citations clustered within responses could make uncertainty intervals unreliable for sparsely cited domains, even when their widths appeared stable. The provider should explain how its estimate handles citations that arrive together in the same response.

Finally, apply a business judgment the measurement cannot supply: how large a gap would change your priorities? A difference can be convincingly above zero yet too small to matter. A potentially important gap can also remain too uncertain to justify a confident conclusion. If the provider cannot estimate uncertainty for the comparison, treat the ordering as descriptive—not as an established competitive lead.

Precision must describe the right questions

Before paying for a narrower uncertainty range, check whether the sampled questions represent the audience you care about. The study explicitly separates sufficient sampling from representative query design. A citation distribution can settle neatly while describing questions that miss your customers.

A useful comparison therefore connects three things: relevant questions, an estimated gap, and uncertainty small enough for the decision at hand. Sometimes that supports a competitor-level conclusion. Sometimes the defensible finding is simply that both domains belong in the leading group. That may be all you need for a market overview; it is not the same evidence you would need to act on a specific competitive lead.