Why the practice down the street gets a better ChatGPT description than you

Ask ChatGPT where to get a mole removed in Phoenix and it may name several dermatology practices. In our study, two of them in the same metro got very different descriptions. One was described as having “fellowship-trained Mohs surgeons,” including a named physician who was “board-certified and fellowship-trained in Mohs and dermatologic oncology.” The other was called “a reasonable option if location is convenient.” ChatGPT cited each practice’s own website. One site handed it detailed credentials to repeat; the other gave it little beyond an address. ChatGPT was not judging the quality of care. It was describing each practice from the information it could find.

The first article in this series showed that when patients ask AI for a doctor, the names that come back lean hospital-affiliated, because AI assembles its answers from the sources it can find and hospitals own the biggest ones. This piece is about what happens a step later, when an independent practice does make the list. Getting named turns out to be a different problem from controlling what gets said about you.

For the average practice ChatGPT named, only about 2.1% of the citations describing it pointed back to its own website. The rest came from somewhere else: a hospital staff roster, a specialty-association directory, a state medical board profile, a city magazine’s “Top Doctors” issue. One Phoenix dermatology practice was named 21 times across our queries, and not one of the citations behind those descriptions pointed to its own site. Every source ChatGPT drew on to describe it was a third party.

The result is a kind of ventriloquism: the practice’s name is in the answer, but a third party supplies the description. When that third party is a current hospital roster, the description is polished. When it is a stale directory still listing an address from a prior employer, ChatGPT repeats it just as confidently.

Those citations are not spread evenly. Across the whole study, practice websites received about 26% of all citations, but those citations are concentrated among a small number of practices. When we instead average the own-site share across named practices, it falls to 2.1%. A few practices account for most of the practice-owned citations; the average named practice supplies almost none of its own description.

The own-site share varies sharply by specialty. In dermatology and plastic surgery, practice websites account for roughly three in ten citations, so a practice with detailed pages has meaningful influence over how it gets described. In cardiology and family medicine, that share drops to single digits, and the description comes mostly from sources outside the practice’s control. A dermatologist may have real leverage over the description through the practice website; a cardiologist is more likely to be described by outside institutions, which makes those external profiles the more important place to focus. Metro shifts the mix too.

Why the practice website still matters

A 26% share can make the practice website look secondary. But when ChatGPT cites your site, the description is only as specific as what you published. That is what separated the two Phoenix practices: one gave it detailed credentials, the other little beyond an address.

The website may also shape third-party sources. Hospital rosters, association directories, and local profiles often carry information supplied by the physician or practice. We did not trace that pathway in this study. The data establishes only that ChatGPT cites practice websites and uses information published on them.

So, the takeaway is narrow and usable. You cannot control every source ChatGPT draws on, but you can control your own pages, and you can audit the handful of third-party profiles that carry weight in your specialty: the roster, the board, the association directory. Which of them matters most depends on your specialty and your metro. The final article in this series will examine what the practices with specific, credible descriptions do differently, and how much of that work is routine maintenance rather than marketing.

Methodology

The description-source findings draw on roughly 33,000 ChatGPT citation events from about 4,950 patient-realistic queries, spanning 200 independent practices selected at random from the NPPES registry across 20 metros and multiple specialties. Each practice was tested with a fixed prompt battery scoped to its specialty and metro. Each prompt was run multiple times to reduce dependence on any single response, and every citation event was grouped by source type and analyzed both in aggregate and by specialty. The 2.1% own-site figure is a per-practice average across the practices ChatGPT named; the 26% is the pooled share of all citations landing on any practice’s website. All queries ran against GPT-5.5 through the OpenAI API in May 2026. The work is single-platform and US-only; behavior on Gemini and Google’s AI Overviews may differ. The analysis is correlational, and the description-source figures cover only the practices ChatGPT actually named, which introduces a selection bias toward practices it was already more likely to surface.

Nathan Woo is the founder of Halcy (halcy.ai), an AI growth platform for independent healthcare practices. The research described here was conducted by Halcy and is shared so practice owners can audit their own AI presence regardless of any vendor relationship.

Companion: For the on-site signals a practice controls, and why its specialty and metro set the ceiling, see our on-site findings.