You Don't Need an llms.txt File. Here's What AI Actually Reads.

September 14, 2026

An advance directive only works if the people treating you actually open the chart and read it.

You can file the most carefully worded document in the world, and if nobody on the code team checks for it in the moment that matters, nothing about what happens next changes.

What protects you is the habit of reading the chart. The document sitting in the file does nothing on its own.

A vendor is probably pitching you something with the same structure right now, except it’s aimed at your website instead of your chart.

The product is a file called llms.txt, sitting in your site’s root directory, that supposedly tells AI systems like ChatGPT or Google’s AI Overviews how to read and use your practice’s content.

Add the file, the pitch goes, and the AI answering a patient’s question about your practice will read it the way you intended.

The file is real. The instruction inside it is optional, and no AI vendor has confirmed that it follows the file.

What the file actually is

llms.txt is modeled on robots.txt, the decades-old text file that tells search engine crawlers like Googlebot which parts of a site to crawl and which to skip.

robots.txt works because every major search engine built compliance into its crawler as a matter of long-standing convention. Ignore it and you get flagged as a bad actor.

llms.txt has no such convention behind it. It’s a proposed format, put forward by developers who wanted an equivalent tool for the new class of AI systems that read websites to generate answers.

And the idea is reasonable on its face.

You point the file at your plainest, most structured content – your services, your credentials, your locations – so an AI summarizing your practice pulls from the right material and doesn’t have to guess.

The trouble is the second half of that. An AI has to be built to look for the file, and then it has to choose to follow what the file says.

Nothing requires either step. Whether it gets honored is up to each AI vendor.

The checkable claim, and the sold claim

You can verify this part yourself. No major AI provider has published confirmation that its systems check for llms.txt or change their output based on what it says, and that includes OpenAI, Google, and the major crawlers behind AI Overviews.

When Google’s own search advocate was asked about it directly, he said he wasn’t aware of any AI product actually using it.

Track the format’s real adoption instead of its marketing and you land in the same place. There’s no evidence any major model reliably honors the directives inside it.

The sold claim is bigger.

Add this one file and you gain control over how AI represents your practice. That’s what shows up on the invoice line, dressed up next to schema markup and other technical items that sound like the same category of thing.

The honest version of the pitch is much smaller. Adding the file costs you almost nothing, and there’s a chance some future AI system reads it. Fine thing to have.

Getting billed for it like it moves the needle today is where the pitch goes wrong.

You’d ask the same thing of a drug rep claiming a new formulation “helps absorption” – helps it compared to what, and where’s the study.

So ask whoever is selling llms.txt as an AI-visibility strategy which system has confirmed reading the file and changed its answer because of it. If you get silence back, you know what you’re actually buying.

Where the same hour of work goes instead

Adding llms.txt costs almost nothing, and there’s no harm in having it.

The real cost is the hour a vendor spends generating that file rather than spending it on the things AI systems demonstrably do pull from when a patient asks something like “OBGYN near me who takes new patients” or “dermatologist for eczema in [town].”

What AI Overviews and chat-based answer engines are documented to lean on:

• Structured data (schema markup) that names your practice, your services, and your physicians in a format machines can parse directly, so they aren’t inferring it from paragraph text.
• A named, credentialed author on your content. A real physician bio, with a name and credentials on it, attached to the pages that answer patient questions. “Our Team” doesn’t count.
• Condition-specific and service-specific pages, each answering one real question in plain language, instead of one general “Services” page trying to cover everything.
• Content that changes. New FAQ entries, updated hours, current staff. Freshness is one of the few freely available signals a system has that your practice is still operating and still being maintained.

A two-physician family practice that spends an hour writing a real FAQ page on “do you accept new patients” and “what insurance do you take,” with the physician’s name and credentials attached and basic schema markup underneath it, has done more for AI visibility in that hour than the same hour spent generating an llms.txt file nobody has confirmed reading.

A dermatology clinic that adds a genuine physician bio page with board certification listed in structured data has given every current AI system something concrete to summarize.

And none of it requires an AI vendor to have agreed to anything. It runs off signals these systems already use for ordinary search, which is what the answer engines are largely built on top of.

So what do you tell the vendor?

Before you approve any line item for llms.txt, ask for the same thing you’d want from a rep pitching a new device claim. A named AI system, and a documented example of it changing its output because the file was present.

If they can’t produce one, decline the line item and put the hour toward one FAQ page, written in your own words, naming the physician who answers it.

Add basic schema markup underneath it if your site supports it. That’s the version of this work an AI system can act on today.

Photograph: Daniil Komov / Pexels