- This piece asks a narrower question than "is the score accurate": does a score measured today predict anything real later, more citations, more leads, more revenue? No study meeting basic disclosure standards (sample, methodology, a defined prediction window) was found that tests this directly.
- That's a distinct gap from measurement volatility, already covered in AI visibility score. A properly-averaged, well-sampled score could still turn out to have no predictive relationship to future outcomes, and volatility alone doesn't answer that question either way.
- SparkToro/Gumshoe.ai's finding that identical prompts run simultaneously return the identical brand list under 1-in-100 times makes the predictive-validity question harder to even test cleanly. A score has to be internally reproducible before "does it predict the future" is a well-posed question at all.
- This piece doesn't claim scores are proven worthless. It claims the specific study that would establish predictive validity, one way or the other, doesn't appear to exist yet, and treats that absence honestly rather than filling the gap with a confident claim in either direction.
- The practical implication: treat an audit score as a directional signal worth tracking as a trend, not as a validated predictor of a specific future business outcome, until a study actually tests that relationship.
The five stages for winning in the AI search era already calls a one-time AI visibility audit "a repackaged audit with a scarier name," sold as a diagnostic without much shown for what the diagnosis actually predicts. This piece takes that framing and asks the specific, testable version of the question directly: does an AI visibility score, measured at one point in time, predict anything real about what happens afterward?
What question is this piece actually asking?
Not: is the underlying citation count accurate. Not: is the score volatile day to day. Both of those are separate, already-documented issues; see AI visibility score for the University of St. Gallen's volatility findings and the vendor-methodology differences between Semrush and Ahrefs. This piece asks the question one step further downstream: granting that a score is measured reasonably well, does a high score at time T actually predict more citations, more AI-referred traffic, or more revenue at a later time T+1, the way, say, a credit score is validated against actual default rates, or a lead-scoring model is validated against actual conversion?
That's a different and considerably higher bar than "does the number reflect something real about the present." A thermometer can accurately measure today's temperature without predicting tomorrow's weather. The question here is whether an AI visibility score functions more like the thermometer (an accurate present-state reading) or the forecast (a validated predictor of what comes next), and whether anyone has actually tested which one it is.
What does a search for that specific study turn up?
Searching directly for research establishing that an AI visibility score, audit result, or similar composite metric predicts future citation volume, referral traffic, or revenue outcomes surfaces marketing claims, not studies. Vendor case studies describing a single client's before-and-after numbers are common; a study with a disclosed sample size, a defined scoring methodology, a measured baseline, a later outcome measurement, and a demonstrated statistical relationship between the two is not. This piece did not find one meeting that bar for this specific category, across the sources available. That's a meaningfully different finding from "a study proved scores don't predict anything": it's an absence, not a refutation, and the honest thing to do with an absence is report it as one rather than dressing it up as a confirmed negative.
Why doesn't volatility alone answer this question?
It's tempting to treat the St. Gallen volatility data (65% day-to-day source turnover, 32-43% overlap on identical simultaneous prompts) as itself proof that scores can't predict anything. That's an overreach in the other direction. A noisy measurement can still be predictive on average, over a large enough sample and a long enough window; plenty of genuinely predictive metrics in other fields (stock analyst ratings, weather forecasting itself) are individually noisy but still carry real predictive value in aggregate. Volatility makes predictive validity harder to establish cleanly. You need a properly averaged, multi-run, multi-week score before "does this predict the future" is even a fair test, but it doesn't logically prove the answer is no. It just means nobody has done the harder, better-controlled study that would actually answer it either way.
Why is this genuinely hard to test cleanly?
Beyond volatility, SparkToro and Gumshoe.ai's consistency finding (under a 1-in-100 chance of two runs of the identical prompt returning the identical brand list, per 2,961 prompt runs across ChatGPT, Claude, and Google's AI surfaces) points at a structural problem for anyone trying to run this study properly. A predictive-validity test requires a stable, well-defined measurement at time T to correlate against an outcome at time T+1. If the measurement itself is this unstable run to run, establishing a clean baseline score to test predictions against is already a nontrivial methodological problem, before the harder question of finding and measuring the right downstream outcome even starts.
What does this mean practically, without overclaiming in either direction?
The responsible position, given the absence rather than a confirmed answer: treat an AI visibility score as a directional trend worth tracking over weeks (consistent with the St. Gallen paper's own 2-4 week rolling-window recommendation), not as a number validated to predict a specific future business outcome. That's a real, legitimate use for the measurement. Noticing a category-wide upward or downward trend is meaningfully different from claiming the score itself has been shown to forecast revenue, leads, or citation volume with any measured accuracy. Vendors selling a score as a predictive tool, specifically, are making a claim this piece could not find support for. Vendors offering a score as a trend-tracking input are making a more modest, better-supported claim.
Frequently asked questions
Does a high AI visibility score predict more future citations?
No study meeting basic disclosure standards (sample size, methodology, a defined prediction window) was found that tests this directly: that a score measured today at time T reliably predicts citation volume at a later time T+1. This is a distinct, narrower, and largely unaddressed question from whether the underlying measurement (citation counting) is accurate at all.
Is this the same question as whether AI visibility scores are volatile?
Related but different. Volatility (documented in the University of St. Gallen's 45-day study, 65% day-to-day source turnover) is about whether a single measurement is a reliable snapshot. Predictive validity is a further question: even granting a well-sampled, properly-averaged score, does that score predict anything about the future, the way a credit score predicts default risk or a lead score predicts conversion? Volatility makes predictive validity harder to establish, but the two are conceptually separate gaps.
Does this mean AI visibility scores are worthless?
Not necessarily worthless. Unvalidated. A score can be a reasonable directional indicator without anyone having actually tested whether it predicts a specific future outcome with any measured accuracy. The honest position is that predictive validity for this category is an open question, not a settled no, because the studies that would settle it in either direction don't appear to exist yet.
What would it actually take to prove an AI visibility score predicts something?
A study with a disclosed sample, a score measured at time T, a defined outcome (citation volume, referral traffic, revenue) measured at a later time T+1, and a demonstrated statistical relationship between the two, ideally against a comparison baseline. That's a meaningfully higher bar than a single case study or a vendor's own before/after client anecdote, and it's the specific kind of study this piece was unable to locate for this category.
Sources referenced
- Schulte, Bleeker & Kaufmann, "Don't Measure Once: Measuring Visibility in AI Search (GEO)" (University of St. Gallen)
- SparkToro / Gumshoe.ai AI recommendation consistency study (Jan 2026, 2,961 prompt runs)
This piece is a direct callback to the "repackaged audit with a scarier name" framing in the five stages for winning in the AI search era. For the measurement-volatility research this builds on, see AI visibility score; for the funding and incentive layer around this industry, see the AI visibility score industry: who's really being served.
