What a medical website needs before AI can cite it
A checklist from our own sites: letting the right crawlers in, putting facts in readable text, and why llms.txt is a courtesy rather than a lever.
Before an AI engine can name you, it has to be able to read you. Most medical websites we look at fall down here first, before content even comes into it. This is the checklist we use on our own sites.
1. Let the right crawlers in
AI companies run more than one crawler, for different jobs. OpenAI, for example, separates OAI-SearchBot, used for ChatGPT search, from GPTBot, used for training. Perplexity runs PerplexityBot. Google’s AI features in Search use the normal Googlebot, while a separate robots.txt token, Google-Extended, controls use in Gemini models and doesn’t affect Search.
Decide on each one deliberately. Then check your firewall and CDN: bot protection often blocks these crawlers even when robots.txt lets them in. Our own sites allow all of them.
2. Put the facts in the page text
AI reads text. Prices in an image, timetables in a PDF, doctor profiles that load only after a click, or pages built entirely in JavaScript are all easy to miss. Server-rendered HTML with the facts in plain sentences and tables is the safest format.
3. One page per real question
A page called “Services” that lists everything gives AI nothing specific to cite. A page that answers “How many visits does an implant take?” does. Build pages around the questions patients actually ask, and keep their addresses stable.
4. Structured data that matches the page
Schema markup helps engines understand who you are and what a page is about. It should describe what is visible on the page, and nothing more. Google says its AI features need no special markup: normal good practice applies.
5. Every language on its own address
If you serve patients in several languages, give each language its own URL and link the versions with hreflang. Text translated by a script after the page loads is easy for crawlers to miss.
6. Keep facts the same everywhere
If your website, your maps listing and a directory profile give three different opening hours, an engine has no reason to trust any of them. Fix the source of truth first, then update the copies.
7. llms.txt: fine, but not a lever
An llms.txt file lists your key pages for language models. It is cheap, and we publish one on our own sites. Don’t expect it to change much: Google says it doesn’t use llms.txt for visibility in Search. Treat it as a courtesy, not a strategy.
8. Measure answers, not proxies
Crawler logs show that a bot visited. They don’t show what an engine said. The only way to know whether this work pays off is to ask the questions and read the answers, the same way, over time.