Halcy Research

When a patient asks AI for a doctor, which practices get named?

Usually, no independent one. We sampled 200 independent practices at random and ran nearly 5,000 patient-realistic prompts through ChatGPT. It named zero of them.

This is the running record of what large language models actually say when people look for care – which practices surface, which sources the models trust, and what the visible practices do differently. Built from our own measurement corpus, shared openly.

Practices
200
Prompts
~4,950
Specialties
18
Metros
20
Model
GPT-5.5
Run
May 2026

ChatGPT mentioned and cited 0 of 200 randomly-sampled independent practices.

How we measured it

Every figure on this site comes from one measurement corpus, not estimates.

We drew a stratified random sample of independent practices from the NPPES registry across 18 specialties and 20 metros, then ran a fixed battery of patient-realistic prompts – scoped to each practice’s specialty and metro – through GPT-5.5 via the OpenAI API, with multiple runs per query to absorb variance. We separate a mention (your name in the answer a patient reads) from a citation (a URL the model pulls from), and anchor on mentions wherever the data supports it. Single-platform, US-only, correlational. Behavior on Perplexity, Gemini, and Google AI Overviews may differ.

Disclosure. Halcy is a commercial service; the practices we work with pay for these deployments. We ran this corpus for our own product development and publish it openly so administrators can audit their own AI presence regardless of any vendor relationship.