AI visibility for doctors: what being named means, and how to measure it

What counts as being named, for a practice?

Three things, in rising order of value. A mention: your practice appears anywhere in the answer, including in a list of other options or with a caution attached. A recommendation: the assistant puts you forward as a suitable choice for the patient who asked. First place: yours is the first practice name the patient reads. An AI recommendation is the nearest thing an assistant does to a referral; a mention only shows that it knows you exist.

Each grade is relative to a question, and the questions that matter carry more than "a doctor". A patient asks for a family doctor who is accepting new patients and takes their plan, a sports medicine doctor who sees teenagers, a clinic with Saturday hours near their work. If you are named for the plain version of the question and missing from the versions with a filter attached, the filter points at the page to check first. The facts inside those questions are set out in what patients ask before choosing a practice.

Some answers are written from pages fetched with web search on, and some from what a model learned in training, which can include hours that have since changed. The panel records both kinds the same way, and in both it is the stated facts you check. How the two routes work for ChatGPT is in what ChatGPT can read on a practice's site.

How is it different from ranking in search or in a directory?

A results page or an insurer's provider directory lists every practice that matches, and the patient does the choosing. An assistant does part of the choosing for them: it writes a paragraph that names a few practices, often with a line on each, and stops. There is no second page to scroll to, so the thing to measure is not a position among dozens but whether your name is in that paragraph at all.

Who shares the answer with you differs too. Our audit of 9,483 small business websites (40 trades, 90 metros in the US, Canada, Europe and the UAE, late August to early September 2026) included a probe that put one buyer question per business to Claude with web retrieval on, for 9,240 of them. Of the names in those answers, 99.4% were businesses and 0.6% were directories or aggregators. It is a whole-sample figure; no medical-practice cut exists. If it holds for your questions, the names beside yours will be other practices, and their pages are the ones worth reading.

Which patient questions belong in a practice's panel?

The panel is the set of questions you ask every month, written the way your patients type them and then left alone, so that one month can be laid beside the next. For a paying client ours holds twelve questions, set for the practice's specialty and location; six will do for a first pass of your own. The general method, including how to tell a change from noise, is in how to measure AI visibility. What follows is how it applies to a practice.

A practice's panel has to carry the filters patients put in their questions, because that is where practices differ from each other. Between them, the questions should cover each office, each service line you want patients for, and each fact patients filter by: new patients, a plan, a language, a visit type, a price. One question can carry several of those at once. Leave out questions about symptoms and treatment. Whatever an assistant says to "what does this rash mean", it is not choosing between practices, and the panel exists to record that choice.

Keep the questions in a file with the date each was added, typed the way a patient types: lowercase, the town the way locals say it, the plan by the name on the card. When a question turns out to be one no patient asks, retire it with a note and add a new line with its own start date. Rewording a question while keeping its history hands the history to the new wording. The record fields are in setting up prompt monitoring.

Running the panel: four assistants, three runs, one reader

Put every question to ChatGPT, Claude, Gemini and Perplexity, three times each, because answers to the same question vary from run to run. Have one person read every answer, record whether the practice was mentioned, recommended or first and who else was named, and keep the answer verbatim. Our methodology takes a narrower reading for clients: twelve questions, each put to one assistant, Claude, at the baseline, again once the fixes are live and every month with monitoring, and a visibility score from 0 to 100 that is the share of those questions in which the practice was named, with first place and the other names reported beside it.

For a practice, read each verbatim answer a second time, for the facts it states about you. An answer that names you with last year's hours, an office you closed or a physician who left sends a patient to a door that is no longer there. Check every stated fact against what the front desk would say, mark the ones that are wrong, and trace each to the page or listing it came from. It is either a page you can edit or a listing someone else controls, and the fix differs.

Keep a dated log beside the panel of what you changed on the site and in the listings: the week the new patients page went up, the day the second office's Saturday hours changed, the month a physician joined. A shift in the answers means little until you can set it next to a change you made, and the log is what lets you do that three months later.

What a first reading can look like, and how to read it

Whatever the first reading shows, treat it as a starting line. In the same probe, 14.9% of 9,240 businesses were named when one buyer question from their trade and metro went to Claude with web retrieval on; among the 5,928 answers that gave a clearly differentiated list rather than a hedge, 17.7%. One question per business, one assistant, one day, every trade together: a direction rather than a rate, and a Claude figure only. A first panel that names your practice on few of its questions, or on none, would be in line with that probe.

Hedges belong in the record. An answer that offers criteria for choosing a doctor instead of names has ranked nobody, you included, and it is marked as a hedge rather than a miss. The point of the baseline measurement is the comparison it makes possible: the second reading, taken with the same questions, shows which question types moved and which did not. The patterns below are the ones worth reading for.

The second row is the place side of a practice, and it is the subject of getting a practice's addresses and directories right. The fourth is about access, which for ChatGPT is set out in what ChatGPT can read on a practice's site.

What the measurement cannot tell a practice

It does not count patients. A patient who hears your name and calls leaves no trace in the panel, and nothing in it measures bookings, no-shows or revenue. If you want that number, ask new patients how they found you, and write down the words they use.

It does not measure the quality of care. An assistant's answer is not a clinical assessment, and a practice that rises in the panel has become easier to read, not better at medicine. Treat the score as a reading of your pages and listings, and keep it out of anything you say to patients.

It does not see the patient's own conversation. A patient may have told the assistant their town, their plan and three earlier questions before asking for a doctor; the panel asks each question fresh and the same way at every reading, because that is what lets readings compare. And it covers only the assistants the panel is put to, which in our readings is one, Claude, and the questions written into the panel. The map results, an insurer's search tool and a question nobody wrote down are outside it.

Nor does the reading change anything by itself. The companies that build the assistants decide what the answers say; a practice, and anyone it hires, can change only what those answers have to draw on. The measurement shows which patient questions leave your practice out and who is named instead, and the work it leads to is on your own pages, starting from the overview for medical practices.