Correlation is not causation: how to honestly measure AI visibility work
The claim every vendor in this market wants to make is “our work got you mentioned.” The claim almost nobody can prove is the same sentence. AI answers change weekly for reasons that have nothing to do with anyone’s work: model updates, index refreshes, phrasing sensitivity, location context. Isolating one cause inside that motion is beyond what the available data supports, and an industry that pretends otherwise is selling certainty it doesn’t have. We sell in this market too, so read this essay as the standard we’re asking you to hold us to.
Why the claim outruns the proof
Proving “our work caused your mention” requires holding everything else still: the model, its sources, the question phrasing, the competitors, the week. None of those hold still. So the honest ceiling for anyone’s claim is correlation: the work happened, the mentions moved, here are both with dates, and here’s why we think they’re related. Anyone claiming more than that is claiming a controlled experiment they didn’t run.
None of this means the work does nothing. It means the connection between work and result lives at the level of pattern and plausibility, not proof, and honest reporting says so in those words.
What honest attribution looks like
Before-and-after answers with dates on both sides. A log of the work performed, laid against the timeline of what moved. The word correlation, in the report, where the connection is drawn. And a refusal to celebrate single results, because answers move week to week on their own and a one-week bump is weather, not climate. The record-keeping half of this discipline is its own practice: the receipts are what make the honest version possible at all.
The branded-search shadow
One asymmetry deserves stating because it cuts in the work’s favor while remaining unprovable: some real effects are invisible to attribution by their nature. A customer hears a business named in an answer, searches the name, and calls. The mention worked, and nothing in any analytics ties the call to it. Framed properly, as hypothesis: the shadow means honest measurement probably undercounts what mentions do, rather than overcounting. We hold that belief loosely and label it what it is, because the alternative, quietly crediting ourselves with everything unmeasurable, is the exact move this essay exists to refuse.
How to buy anything in an unprovable-attribution market
The practical payoff, and it applies to tools, agencies, and us equally: judge vendors on their receipts, not their outcome promises. A vendor who shows dated answers, publishes their method, labels their inferences, and says “correlation” out loud is giving you the maximum honesty the field permits. A vendor promising attributable outcomes is promising past the ceiling, and the questions that expose the difference take five minutes to ask. This is also the honest frame for reading your own dashboard: it shows you what moved, with dates; the story of why is built from the log, carefully, or not at all.
Our stake, stated once
We make an AI visibility dashboard and offer an expert managed service, so we’re on both sides of this comparison, and this essay is not neutral ground. The standard binds us anyway, and there’s a selfish reason we keep it: in a market where nobody can prove causation, the only durable trust asset is a record of never having claimed it. Our opinion, held all the way through this essay: honesty about attribution isn’t a marketing style. It’s the only measurement position that survives a smart customer’s second look.