AI visibility graders vs. client-discovery monitoring: what a therapy practice actually needs

Brand-introspection tools send your company name, industry, and location to a model and ask it to characterize you: how recognized you are, how people feel about you, where you sit against competitors. The model answers from whatever it absorbed during training—plus, in some tools, a live search from one of the engines but not the others.

Client-simulation tools do something narrower and more literal. They ask the questions a prospective client would type, record which practices are named, note which public sources the answer cited, and repeat the same questions on a schedule so change is visible. The output is less sweeping, but every line of it corresponds to something that can actually happen to you: a real question, a real answer, a real citation.

When a model is asked to estimate your 'mention count' or 'share of voice' from its training data, it will usually produce a confident figure. That figure is a plausible-sounding guess, not a tally of anything. Models do not keep countable records of brand mentions, and their training data has a cutoff date—sometimes a year or more in the past—so even an honest recollection describes an older internet than the one your next client will search.

This is not a criticism of any one tool; it is a property of the method. The same model that declines to invent an answer when it can search the live web will happily estimate one when asked to introspect. A small practice with a modest public footprint gets the thinnest version of this: there is little in the training data to recall, so the description leans on the practice's name, its category, and general industry patterns.

If a number cannot be traced to a stored answer, a cited source, and a timestamp, treat it as a prompt for curiosity—not as a measurement to act on.

Whatever tool you use, the trustworthy versions have a common shape. They show their work, and they are specific about what was and was not measured.

• The exact questions that were asked, in the words a client would use
• The full answers, stored, with the date and time they were captured
• The sources each answer cited—your website, a directory profile, a Business Profile
• Which model produced each answer, and whether it searched the live web or answered from memory
• A clear rule for what counts: answers backed by live search, not a model's recollection

A free brand grade is not worthless. It can surface how a model categorizes you, name competitors worth a look, and start a useful conversation about public presence. Treated as a snapshot with a wide margin of error, it does no harm.

The trouble starts when a one-time introspection score is treated as a baseline to optimize against. Scores built on model recall can swing between runs for reasons that have nothing to do with your public presence, which makes week-over-week comparison unreliable—and week-over-week comparison is the entire point of measuring.

A national software company plausibly cares how a model describes its brand, because buyers ask exactly that. A therapy practice's next client almost never asks an assistant to describe a practice by name. They describe a situation—a city, a specialty, a payment preference—and the assistant produces a short list. Either your practice is on it, or another practice or a directory is.

That is why the useful check for a practice is built from client-style questions, local surfaces, and the sources doing the talking: the map pack, the Business Profile, the directory profile that outranks your site, the specialty page the answer quoted. Those are things you can inspect, improve, and re-measure. A sentiment estimate is not.