Health Biz Scale asked an AI model 2,292 patient-style questions about 764 health practices between May 11 and Aug 20, 2026. Only 6 of the 764 practices (0.8%) were named in its answers. When the model gave website addresses, about 1 in 3 mentions pointed to a domain that did not resolve.
The model answered from training memory with no web search. So this study measures what AI "knows" without looking anything up, which is a different thing from what a patient sees in ChatGPT with search on.
Key findings
- 6 of 764 practices (0.8%) were named at all. Across 2,292 answers, 10 named the practice being checked (0.4%).
- 1,159 of 3,562 website mentions (32.5%) did not resolve. This check covers 1,233 answers about 411 of the practices.
- 745 of 1,449 unique domains (51.4%) did not resolve. Counted once each, about half of the addresses the model offered had no working domain behind them.
- All 6 named practices came from questions tied to a real city. None of the 155 practices with no city on file were named.
Results by specialty
| Specialty | Practices | Named | Practices with domain check | Website mentions | Did not resolve |
|---|---|---|---|---|---|
| Functional medicine | 378 | 2 | 233 | 2,039 | 660 (32.4%) |
| Wellness | 69 | 0 | 0 | n/a | n/a |
| Naturopath | 59 | 2 | 24 | 213 | 56 (26.3%) |
| Chiropractic | 55 | 0 | 30 | 263 | 84 (31.9%) |
| Acupuncture | 47 | 0 | 28 | 237 | 90 (38.0%) |
| Weight loss | 31 | 0 | 23 | 189 | 60 (31.7%) |
| Integrative medicine | 29 | 1 | 14 | 117 | 44 (37.6%) |
| Primary care | 23 | 0 | 19 | 163 | 52 (31.9%) |
| Other | 73 | 1 | 40 | 341 | 113 (33.1%) |
| Total | 764 | 6 | 411 | 3,562 | 1,159 (32.5%) |
Specialty here is the search term used to set up each check. It defaults to functional medicine, so treat the groups as rough. Small groups can swing on one practice, so don't read the differences between them as real gaps.
Download the aggregate data (CSV). It holds totals only, with no practice or city names.
What this means for a practice
When an AI model has no strong, consistent signal about who is real in a local market, it fills the gap with plausible guesses. A third of the website mentions pointed nowhere, so the model was offering patients addresses that were wrong, made up or no longer live.
The established practices we checked were almost absent from the model's memory. Being good at what you do and having a website did not get a practice remembered.
Live search narrows the guessing, because the model can look things up. It looks in the same places, though: your website, your listings, the directories and the sites that mention you. The practices that get named are the ones those places describe clearly and consistently. Our AI SEO guide covers how to build that.
Method
- Model: GPT-4o-mini through OpenAI's API, answering from training memory with no web search.
- Dates: May 11 to Aug 20, 2026.
- Practices: 764 health practices, one check per practice (latest run kept), test runs excluded. They are practices we researched as potential clients, so this is a selected list, not a random sample of the market.
- Questions: 3 patient-style questions per practice, such as "best acupuncture doctor in [city]" and "who treats hormone imbalance and fatigue in [city]." 1,574 of the 2,292 questions (68.7%) named a real city. The rest used "near me" or "in your area."
- Named: an answer counted when it mentioned the practice being checked.
- Website check: a domain lookup at scan time, on up to 6 addresses per answer. It shows whether the domain exists, which is a narrower test than loading a page. It ran on August scans only, so it covers 411 practices and 1,233 answers.
Limitations
- This is one small model with no web search. The ChatGPT app searches live and can answer very differently.
- The 764 practices are a selected list, and half are tagged functional medicine.
- Mentions are counted per appearance, so a domain repeated across answers counts each time. The unique-domain figure above corrects for that.
- Named websites include some directories and health systems, not only practices.
- We named no practice or city here, and we will not.
How to cite this study
Health Biz Scale. "What AI Remembers About Health Practices: A 2026 Study of 764 Practices." October 2026. https://www.healthbizscale.com/seo-aeo/ai-memory-study-health-practices-2026
Updated Oct 31, 2026. This version corrects two earlier totals, from 3,554 and 1,153 to 3,562 and 1,159, and states the coverage of the website check.