llms.txt is a Markdown file at the root of your site that gives AI agents a short map of your practice: what you treat, who you serve, which pages answer which questions, how to reach you. No engine has confirmed it as a ranking signal. Publish it, but treat it as a courtesy for agents, not a lever.
What llms.txt Actually Is
A website publishes a plain text file called llms.txt in the root of its domain, written in Markdown, and that file tells a language model what the site is and which pages matter. The proposal and the format are published at llmstxt.org, where the file is described as a voluntary convention rather than a standard any engine must follow.
(file not found)
# Your Practice Name > Family dentistry in Your City. ## Key pages - [Dental implants](/dental-implants): cost, process, recovery - [Emergency care](/emergency-dentist): hours and what to do
The file has three parts: a title, a short summary of the business, and a list of links with one line of description under each. That last part is why it exists. Without it, an agent has to fetch your homepage, work out that your treatment pages live under a menu called Services, and guess that cost-of-implants is the page a patient asking about price wants.
The convention also allows a second file, llms-full.txt, holding the complete text of the pages you want considered, for agents that would rather read the content than browse the map.
What It Is Not
Three things it is not, and the confusion about each is why practices publish it with the wrong expectations.
It Is Not a robots.txt Replacement
robots.txt decides whether a crawler may fetch a URL. llms.txt is a document a crawler may choose to read once it has permission. One is a gate, the other is a leaflet inside the gate. If you have never looked at which crawlers are welcome, start with which AI crawlers to allow in robots.txt, because a permission mistake costs more than this file ever could.
It Is Not a Ranking Factor
A ranking factor is something a system uses to order results. llms.txt appears in no engine's published ranking systems, and no vendor has documented one that includes it. The difference matters when you decide where to spend an afternoon: a convention tells a tool how you would like to be read, a signal tells it how to behave. You can follow every line of the convention and still be absent from the answer.
It Is Not a Promise Of Citation
Nothing inside the file makes an engine quote you, and no file could. The document has no opinion about your prices, your reviews or whether you suit the patient who asked. It is an address book, and whether the engine reads it is not something a website can direct.
What The Evidence Actually Says
Here is the documented position, stated plainly rather than dressed up. Google's AI features optimization guide (Google's AI features optimization guide) says there is no special markup or file required for AI features: the fundamentals that make a page eligible and readable for search are the ones that make it eligible and readable for AI answers. That guide does not mention llms.txt, and it does not list it as an input to anything.
What the other vendors document is equally ordinary. They crawl, they read structured data such as the Dentist type described in our guide to dental schema markup, they read your pages, and they cite what they find. What no vendor has published is a statement saying it reads llms.txt as a signal.
So the honest position has two halves. There is no evidence that llms.txt changes how often you are recommended, because no vendor has confirmed reading it. There is also little evidence that it does harm, because a file nobody fetches does nothing. It sits in a small and useful category: cheap to publish, unproven to help, almost certainly harmless.
How To Write One For a Dental Practice
Keep it to a single page. The published specification uses second-level Markdown headings for its link lists; bold labels do the same job and are easier to scan. Here is a complete example you can copy and fill in.
# [Practice Name]
``
> [One paragraph: what the practice is, who it treats, and where it is.]
``
Treatments
``
- [Treatment one], [what it involves and who it suits]
- [Treatment two], [what it involves and who it suits]
- [Treatment three], [what it involves and who it suits]
``
Who we treat
``
[Families, children, older patients, nervous patients. Be specific.]
``
Key pages
``
- About The Practice, team, training and appointments
- Costs and Payment, what patients pay and how to pay
- Insurance We Accept, the plans we take
- Dental Emergency, what to do when the pain is bad
- Reviews, what patients actually wrote
``
Facts an assistant may quote
``
- [Practice name], [street], [city], [state], [ZIP code].
- Phone [number]. Email [address].
- Open [days and times]. Walk-ins [policy].
- [Number] dentists, [Number] hygienists.
- Languages spoken: [list].
- New patients: [accepting, waitlist, or the one sentence to say].
``
Contact
``
[Booking link]. [Phone number]. [Full address].
Two details go wrong most often. The first is listing bare links: a row of URLs tells a model nothing except that those pages exist, so every line needs a sentence describing what a patient would find there. The second is the facts block, which should be written so each line can be quoted word for word, because that is the section an assistant is most likely to lift. Only list pages you would be happy to see quoted.
Where To Publish It and What To Link
Two paths. The short file sits at /llms.txt on the root domain and the long one at /llms-full.txt, both served as plain text. Then point at them from your /sitemap.xml, the footer of every page, and the full file back to the short one.
Linking is not decoration. An agent that has never heard of the file will not guess the path, and the sitemap is the discovery route that already works.
Swipe the table sideways to see every column.
| File | What It Holds | Who Reads It | When You Update It |
|---|---|---|---|
/llms.txt | A title, a one-line summary and a list of the pages that matter, each with a line of description | Agents that fetch the file by convention | When a treatment, a page or a contact detail changes |
/llms-full.txt | The complete text of the pages listed in the short file | Agents that want the content rather than the map | Whenever you publish or retire a page |
/robots.txt | Rules about which crawlers may fetch which paths | Every crawler that reads it, including the AI ones | When a vendor changes what its crawler does |
/sitemap.xml | The list of real, indexable addresses on the site | Search engines, and agents that read sitemaps | After every publish, rename or removal |
Keep it in sync. A stale map is worse than no map, because an agent may quote a page whose hours changed or which no longer offers the treatment you listed. Put a recurring reminder in the practice manager's calendar and treat a changed service list as a reason to edit the file that day.
Where a file like this fits: how to structure a dental site so each question has one page decides what goes on the map, and how to link a dental site so each page owns a question decides whether an agent can reach it.
Is It Worth The Ten Minutes
Yes, with conditions attached. The cost is one file and one reminder a month. The benefit is a small but real chance that an agent which prefers a curated map gets a better one than it would have assembled from your homepage and your menu. That chance is not zero.
It matters most in two situations: a site big enough that a model cannot work out what matters from the navigation alone, and a service people research across several pages before booking, such as clear aligners or an implant consultation where cost and candidacy both have to be understood.
It matters least when your pages are not being read at all. If a crawler is turned away by your rules, or the fetch fails, a map pointing at those pages changes nothing.
We keep one for our own site at drsmiles.ai, and the reason it is worth the hour is the discipline it forces: you have to decide which pages are the important ones and write that decision down. If you would rather see how your pages perform first, the snapshot runs ten questions across three engines in about sixty seconds and the method is described in Dental AEO. Pricing and terms are answered on our pricing and terms FAQ.