In September, OpenAI said ChatGPT for Healthcare can connect to Epic records, with UCSF Health as a pilot. That is AI inside the hospital, working on data the hospital controls.
The other half gets less attention. Patients already ask assistants which clinic treats a condition, whether a drug is covered, or where to get a prescription filled. To answer, the assistant has to read the public website of the hospital, insurer or pharmacy. We wanted to know what it gets.
How we checked
This is the third post in our weekly series where we run our free agent-readiness check on one industry's best-known sites. Before this: 20 US retailers and 15 travel sites.
On November 17, 2026 we checked the homepages of 15 hospitals, insurers, pharmacy chains, telehealth companies and health publishers, then ran all 15 a second time. Both runs gave the same result for every site.
The check fetches the homepage the way a simple AI agent would, with a user agent that names us (X1Y1-Tools/1.0). It reads robots.txt rules for the AI search and assistant bots, counts the words in the raw HTML before any JavaScript runs, and looks for structured data, metadata, a sitemap and llms.txt. It scores out of 100.
The usual limits apply. Our checker is not ChatGPT, and bot protection services often keep allow lists of verified AI bots, so a site that turned us away may still let OpenAI or Anthropic through. A homepage is not a doctor search or a drug price page. This is a first look, not an audit.
The results
| Site | Result on Nov 17 |
|---|---|
| Cleveland Clinic | 90 |
| Cigna | 88 |
| CVS | 83 |
| UnitedHealth Group | 80 |
| Healthline | 80 |
| Teladoc Health | 75 |
| WebMD | 70 |
| Aetna | 68 |
| Walgreens | 68 |
| Mayo Clinic, Johns Hopkins Medicine, Zocdoc, GoodRx, athenahealth | Bot-protection response (HTTP 403), both runs |
| Kaiser Permanente (kp.org) | Our checker could not connect, both runs |
6 of the 15 never got measured. Kaiser needs a footnote: from a different network later the same day, kp.org answered with a redirect to kaiserpermanente.org. So we count it as a failure somewhere between our checker and their site, not as a block.
The opposite of travel
Last week, 4 of the 6 travel sites we could read had zero words in their homepage HTML. Booking.com did not even have a title.
Healthcare is different. All 9 sites we could read serve real text before any JavaScript runs:
| Site | Words in the raw HTML |
|---|---|
| CVS | 94,682 |
| Aetna | 4,737 |
| Cigna | 1,801 |
| Teladoc Health | 1,160 |
| Walgreens | 1,126 |
| Cleveland Clinic | 1,018 |
| WebMD | 902 |
| Healthline | 886 |
| UnitedHealth Group | 874 |
All 9 also allow every AI assistant bot we check for in robots.txt: OAI-SearchBot, ChatGPT-User, Claude-SearchBot, Claude-User, PerplexityBot and Perplexity-User. Teladoc Health has no robots.txt at all, which allows everything by default.
So the split in healthcare is not between readable and blank pages. It is between sites that let an unknown agent in and sites that don't. Mayo Clinic, Johns Hopkins, Zocdoc and GoodRx are exactly the sites a patient's assistant would want to read, and all four sent our checker a bot-protection page.
What separates 90 from 68
Cleveland Clinic and Cigna are the only two with a working llms.txt, and both have structured data on the homepage that says what they are: MedicalOrganization for Cleveland Clinic, InsuranceAgency for Cigna.
CVS has an llms.txt too, but it does not start with a Markdown title the way the format expects. Walgreens, UnitedHealth Group and Teladoc Health answer /llms.txt with an HTML page instead of a text file, which usually means a catch-all route serving the site's 404 page with a 200 status. An agent that asks for the file gets a web page back.
Aetna and Walgreens tied lowest at 68. Both have readable HTML and open robots.txt rules, and neither has llms.txt or any JSON-LD on the homepage. WebMD has no homepage structured data either.
None of the 15 returned Markdown when asked for it with Accept: text/markdown, and none had a Content-Signal line in robots.txt. As with travel, those are newer signals, and we would fix the basics long before either.
Two publishers draw a line on training
WebMD and Healthline both block AI training crawlers while allowing the assistant bots. WebMD blocks GPTBot, ClaudeBot and CCBot. Healthline blocks those three plus Applebot-Extended and meta-externalagent. Both still allow Google-Extended.
That is the same split Airbnb made in our travel check, and it makes sense for publishers whose content is the product: an assistant may read a page to answer a user, a crawler may not take the archive for training. The other 7 scored sites allow every training crawler we check for.
Where this is overhyped
What we would not claim from this data:
- That Mayo Clinic or Zocdoc "block AI". Our checker got a bot-protection response. Their policy for verified AI bots may differ, and for a hospital, keeping scrapers off appointment and patient pages is a fair call.
- That a 90 means better care or more patients. It means an AI agent can read the homepage and knows what the organization is.
- That llms.txt moves rankings. Nobody has shown that yet. It is cheap, and it gives an agent a clean map of the site, which is why we count it.
What a clinic or health startup can do with this
A large hospital system is in every model's training data. A specialist clinic, a telehealth startup or an independent pharmacy is not. If an assistant cannot read your site, it describes someone it can read.
What we would check this week:
- Bot protection rules. If you use Cloudflare, Akamai or similar, check whether verified AI bots (ChatGPT-User, Claude-User, Perplexity-User) get through on public pages, while patient portals and booking stay protected.
- Structured data. MedicalClinic, Physician or Pharmacy JSON-LD with your address, specialties and hours.
-
llms.txt. A short Markdown file at
/llms.txtthat lists your services, locations and key pages. Make sure your router returns a 404 for missing files, not your homepage. - robots.txt. Decide on training crawlers separately from assistant bots, the way WebMD and Healthline did.
- One test. Ask ChatGPT and Perplexity what your clinic treats and where it is, and see whether the answer is right.
You can run the first three on your own site with our agent-readiness check. It is free and takes about ten seconds.
Next week
Banking is next. If there is a site you want in the set, tell us.
Top comments (0)