A patient with knee pain used to type ‘orthopedic near me’ and scroll. Increasingly they ask an assistant: who is good, who takes my insurance, who can see me this week. The assistant answers from the pages it is allowed to read.
If your host, CDN, or ‘security’ plugin added Disallow: / for GPTBot, ClaudeBot, PerplexityBot, or Google-Extended, you are not in that answer. Your Google rank can look healthy while AI search treats the practice as a blank lot.
The failure is usually one file
robots.txt is a public instruction manual. Most practice owners have never opened it. Agencies copy a boilerplate that blocks ‘bad bots’ and accidentally includes the assistants sending you patients.
- Allow OAI-SearchBot (and ChatGPT-User) if you want ChatGPT search. GPTBot is training — a separate policy choice.
- Allow Claude-SearchBot and Claude-User. ClaudeBot is training.
- Allow PerplexityBot — Perplexity cites sources; blocked sites do not get the citation.
- Do not block Googlebot. That is Search, including AI Overviews. Google-Extended is a different control for other Google AI systems.
- An llms.txt is optional. Google Search does not use it. We still publish one as a Markdown index and say so.
Access is not the same as being chosen
Opening the door is step one. Assistants still need extractable facts: specialty, locations, providers with NPI, conditions treated, hours, and plain-language service pages. A five-page brochure site with stock photos and no schema is readable and still useless.
If search cannot find you, the work goes somewhere else.
Run the free AI visibility checker on your domain. If a crawler is blocked, fix that today — before you spend another month on blog posts nobody’s model is allowed to read.