What the measurement can see and how it can be applied
Two questions decide whether an AI visibility finding is worth anything. Whether there is a record to check the answer against. And whether the analysis can tell you why the answer came out that way, rather than only what it was. This page answers both, so the answer is the same whoever is reading it.
The measurement only works where there is a record we did not write.
A registered adviser files a Form ADV. A firm performing federal audit work is named in a mandatory filing. A campaign that produced a named thing — an act, a coalition, a registered day — left something in the world that can be checked.
A positioning statement is not that. Neither is a measurement of a set of brand attributes, because the client wrote them, and measuring a machine against a client's own claims produces a marketing finding that wants a marketing fix. Where no external record exists, we can still measure stability and retrieval composition — whether the machine says the same thing twice, and what it is reading. That is a narrower product and it is priced as one.
This is the first thing we establish on a call, and it is the reason some inquiries end there.
What the citation record shows that a score cannot.
A visibility score reports whether a firm appeared. It cannot report how, or why, or what the platform read to get there. Those distinctions change what a response to the finding should be — and they are only visible in the citation record.
There are three ways into an AI answer, and they need different responses.
A firm can be named because a platform read its website, because a directory or association listing carried it, or because the model already held the firm in memory and cited nothing at all. These look identical in a mention count. They are not the same finding, and work that addresses one does nothing for the others.
Being read and being recommended are close to independent.
A firm's site can be consulted repeatedly across a study and the firm never named. Another can be named repeatedly with its website never read. A single visibility score reports those two situations identically, and they call for opposite conclusions.
Sometimes the firm was never in the retrievable record at all.
There is a difference between a platform reading a firm's material and passing over it, and a platform never encountering the firm. The second is not a positioning problem. No amount of positioning work reaches a system that has never retrieved the firm — and knowing which situation you are in before the budget is committed is worth more than the budget.
The question decides which competitive field appears.
A generic category question and a specific buyer situation summon different sets of firms. Some questions produce no local recommendation at all, because the platform routes them to a different profession entirely. The result depends on what was asked, which is why the question has to be traceable to something outside the measurement.
This is the level at which strategy becomes possible. Not “the score went down,” but “the system never retrieved the firm,” “the evidence did not support a recommendation,” or “the question activated the wrong competitive category.” Findings from AIWorthy's published studies appear on the Studies page as they are released.
What the method will not do.
NO COMPOSITE SCORE
We report counts and the conditions under which they were collected. We do not compress them into a single number out of a hundred. A composite hides which of the distinctions above produced it, which is the only part worth acting on.
A REFUSAL IS NOT AN ABSENCE
When a platform declines to answer, we record a refusal. Scoring that as a zero would make a platform's safety behavior look like a firm's invisibility.
ONE RUN IS NOT A MEASUREMENT
Answers move between runs, and local results move by location. Every question is asked repeatedly, from a stated location, on a stated date. A measurement that does not say where it was taken can hide that error indefinitely.
SOME FIXES REACH SOME PLATFORMS
Published changes reach the platforms that search the live web within weeks. They do not reach the platforms answering from stored knowledge until those are retrained, and nobody can say when. We will tell you which of the two you are dealing with, and we will not give you a timeline for the second.
A credible measurement sometimes ends with ‘we do not know yet.’ We would rather say that than produce a number that survives a meeting and not a year.
The record behind all of it.
Panels, screens and collection conditions for every study are published before the first question is asked. Findings publish in full, including the corrections we issue against our own work.