Why ChatGPT doesn’t mention your practice
If ChatGPT has never named your practice, you are the norm, not the exception.
We sampled 200 independent practices at random and asked ChatGPT the questions their patients ask. It named none of them. Working backward from the practices it does name, three things separate them – and almost no marketing budget is aimed at any of the three.
- Sampled
- 200
- Named
- 0
- Prompts
- ~4,950
- Model
- GPT-5.5
Audit your own practice in 10 minutes
Open ChatGPT and ask these four questions. Count how many name your practice.
- What’s the best [your specialty] practice in [your metro]?
- I have [a common condition you treat] – who should I see in [your metro]?
- Where can I get [a top procedure you do] in [your metro]?
- Who are the top [your specialty]s in [your metro]?
Does ChatGPT cite your practice?
Run a free scan to see your A–F visibility grade and how often ChatGPT cites your practice.
Five on-site signals, and almost nobody has them
Of 516 practices ChatGPT cited, two in three had none of the five on-site features that distinguish the practices it quotes from their own websites.
Schema markup and content depth tell an AI search engine what a page represents – which physician, which procedure, which credentials. Pages without them read as undifferentiated prose. Here is how rare each signal is among already-cited practices.
Adoption among the 516 cited practices with a clean page scrape. Source: Halcy measurement corpus, May 2026.
Different questions pull from different source pools
There is no single “AI visibility.” Each kind of patient question draws on a different pool of sources – and independent practices are never the largest share of any of them.
Reputation queries route to the regional magazine’s “Top Doctors” pipeline. This is the one lane an independent practice can actively enter.
Practice-org share of named entities, by question type. Hospital-affiliated physicians are the largest bucket in four of the nine question patterns we tested (41–55%); independent practices sit at 7–17% across every one.
Your ceiling is your specialty and your metro
Before you benchmark yourself, know your ceiling: practice-org mention share varies about 2× across specialties and 3× across metros.
By specialty – mention share
By metro – mention share
Metros with dense academic medical centers (Boston, San Francisco, Chicago) crowd out independent practices. The two endpoints disclosed here bracket a ~3× range across the 20 metros sampled.
~3× spreadA Boston cardiology practice at 7% is doing well for its market; a Charlotte ENT practice at 7% is underperforming. Benchmark against your cell, not the average.
Questions about this research
Short answers to how the study was run and what it does and doesn’t claim.
Does “0 of 200” mean ChatGPT never names any practice?
No. It means that across a random sample of 200 independent practices, none were named in the patient-realistic prompts we ran. Working backward from the practices ChatGPT does name shows a clear set of shared on-site signals – the point is that they are rare, not impossible.
What is the difference between a mention and a citation?
A mention is your practice’s name appearing in the answer a patient reads. A citation is a URL ChatGPT links as a source. Mentions are what convert to bookings, so we anchor on mentions wherever the data supports it.
What are the five on-site signals?
Long-form content (at least 500 words on entity pages), on-page Q&A blocks, FAQPage schema, a hero-intro paragraph naming the physician and credentials, and Physician schema. Among 516 cited practices, 66.1% had none of them.
Which AI model and time period does this cover?
All prompts ran against GPT-5.5 via the OpenAI API in May 2026, US-only. Behavior on Perplexity, Gemini, and Google AI Overviews may differ, and model updates over time are a confounder.
Can I improve where my practice lands?
The contestable ground is procedure queries, where the source pool fragments and commercial intent is highest. Practices that combine content depth with schema markup appear in cohort-specific answers more often than either factor alone predicts.
Methodology & caveats
The random-sample experiment used 200 NPPES-sampled practices and ~4,950 queries; the cohort and lever analysis draws on the broader citation corpus. The five-signal incidence is computed over the 516 cited practices with a clean page scrape. All queries ran against GPT-5.5 via the OpenAI API in May 2026. Cohort data is survivor-biased; lever analysis is correlational – no causal claims. Read the off-site companion at the ChatGPT citation map.