AI visibility is becoming a regular reporting line for brands and agencies, but a single prompt run may be a poor basis for a performance claim. New research argues that rankings of domains cited by AI search surfaces can remain unstable across repeated measurements, particularly where leading domains are closely matched.
Ronald Sielinski's preprint, From Stochastic to Stable: Rank Stability and Structural Sufficiency in AI Visibility Measurement, proposes a framework for deciding when enough AI-output data has been collected to make an inference about which domains are cited most often. Its central message is straightforward: collect repeatedly, test whether the ranking has settled, and show the uncertainty rather than presenting a precise-looking number that may be mostly noise.
For businesses building GEO dashboards, this changes the question from “What was our AI visibility score today?” to “Has the observed result become stable enough to inform a decision?”
Why a single AI visibility reading can mislead
AI-generated answers are probabilistic outputs. Results can vary with repeated runs, prompt wording, the platform, market and language settings, and the underlying state of the AI product. That makes an isolated count of mentions, citations or recommendations a limited observation rather than a conclusive ranking.
Sielinski tested the issue across 30 platform-topic combinations involving Gemini, SearchGPT and Perplexity. The paper evaluates two conditions:
- Rank stability, meaning whether the ordering of domains remains consistent as more samples are added.
- Structural sufficiency, meaning whether the observed pattern has enough separation among leading domains to support an inference.
The study found no universal sampling budget that guarantees stable rankings. Some topics did not converge within the tested budgets. Others became more dependable only when the top domains had a clear enough separation. In practical terms, a dashboard cannot safely assume that, for example, 10, 50 or 100 runs will always produce a settled result.
That conclusion does not mean AI visibility cannot be measured. It means the measurement design must acknowledge variation before it is used to compare brands, publishers or competitors.
The distinction between visibility signals and performance facts
An AI citation or recommendation is useful evidence of how an AI surface responded to a defined query at a particular time. It is not automatically evidence of traffic, conversions, revenue or durable market-wide preference.
This distinction matters because several different ideas are often collapsed into an “AI visibility” metric:
- Whether a brand or domain appeared in an answer.
- How often it appeared across a specified query set.
- Whether it was cited, named or recommended.
- Its position relative to other domains in a sampled ranking.
- The downstream business impact of those appearances.
Each measure can be relevant, but none should be treated as interchangeable. A citation-frequency change may reflect a change in prompts or sampling rather than a meaningful shift in a company's discoverability. Likewise, a stable citation ranking does not by itself establish that users clicked, trusted or acted on the answer.
The paper's recommendation to report distributions or confidence intervals instead of single point estimates is intended to make this uncertainty visible. Rather than saying a domain has exactly a certain share of citations, reporting can show the observed range across repeated runs and whether the relative ordering remained stable.
A measurement-governance model for GEO reporting
The research aligns with the broader caution in AMEC's Practitioner's Guide to GEO Measurement. The guide says there is no single stable metric for AI-led discovery and frames GEO outputs as directional evidence that needs defined governance around sources, prompts, markets and languages.
For smaller businesses, governance need not mean an elaborate research operation. It means documenting enough of the measurement process that a future result can be interpreted and compared responsibly. A useful AI visibility dashboard should retain:
- The AI platform or surface tested.
- The exact prompt set and the business intent each prompt represents.
- Market, language and any other available settings.
- The dates and number of repeated runs.
- The definition of a mention, citation or recommendation.
- Ranges or distributions alongside any headline result.
- A note on whether the observed ranking met a defined stability threshold.
This record helps prevent a common reporting error: treating two results as directly comparable when the prompts, platforms or collection design changed. It also allows teams to distinguish a real movement worth investigating from ordinary output variation.
What an SMB-ready dashboard should show
A practical dashboard does not need to hide uncertainty to be useful. It can separate diagnostic signals from business outcomes and present findings in a decision-friendly form.
For example, a reporting view might show repeated-run citation ranges for a cluster of high-value questions, the domains that recur across those runs, and a stability status for each topic. A topic with a wide range and changing leaders should be marked as inconclusive, not converted into a league-table claim. A topic with a narrow range and consistent leading domains may be a stronger candidate for content or entity analysis.
That approach is also supported by related SparkToro research on inconsistent AI brand and product recommendations. Variation across runs and prompts is not merely a reporting inconvenience. It is a property that measurement programs must design around.
What teams can do now
The immediate response is not to abandon AI visibility monitoring. It is to use it with a clearer standard of evidence.
First, focus on a documented set of prompts connected to real customer questions, rather than a large but loosely defined keyword list. Second, repeat collection before making comparative claims. Third, establish a stopping rule based on stability, not a fixed number of runs alone. Finally, pair AI-output observations with other evidence, such as Search Console trends, referral data where available, conversion performance and qualitative review of the answers.
The preprint remains research that will benefit from replication and peer review. Its 30 platform-topic tests also do not establish a universal rule for every AI system, category, geography or prompt set. But its warning is immediately applicable: precision in an AI visibility dashboard should be earned through repeatable data collection, not assumed from a single snapshot.
If your team is beginning to track how its site appears in AI-led discovery, create a GEO Search Leads account. It provides a starting point for assessing website GEO readiness and identifying optimization opportunities, while giving your team a clearer basis for investigating the content, entity and technical factors behind observed visibility signals.
Frequently Asked Questions
Why is one AI visibility check not enough?
AI-generated answers can vary across repeated runs, prompts and platforms. The cited research finds that domain rankings do not always stabilize within the tested sampling budgets, so one observation may reflect noise rather than a dependable pattern.
What does rank stability mean in AI visibility measurement?
Rank stability refers to whether the ordering of cited domains remains consistent as additional samples are collected. A stable ordering provides stronger support for a comparative inference than a ranking that changes across runs.
What should an AI visibility report include?
It should document the AI platform, prompt set, market and language settings where applicable, collection dates, number of runs and the definition of the measured signal. Reports should show ranges or distributions and identify whether results appear stable.
Does AI visibility prove that a brand will receive traffic or conversions?
No. A mention, citation or recommendation is evidence about an AI response to a defined query. It does not by itself prove clicks, conversions, revenue or sustained visibility across the wider market.
Conclusion
AI visibility measurement can inform GEO work, but it should be treated as a repeated, governed observation process rather than a one-off score. Teams that document their inputs, test for stability and report uncertainty will have a firmer basis for deciding which AI-search signals deserve action.
Top comments (0)