What is llms.txt? A plain explanation of the file written for AI agents

Your website now has two kinds of visitor. The first is a person, who looks at the page, reads the headline and scrolls. The second is a piece of software, an AI agent, sent by a person to find out about you before a call, a booking or a quote. It does not look at anything. It reads, fast, and it wants the facts without the furniture. llms.txt is a small file written for that second visitor. This post explains what it is, in the order someone who has never heard of it needs.

What it is, in one paragraph

It is a plain text file, named llms.txt, that sits at the top of your website at yourname.com/llms.txt, the same place robots.txt has lived for thirty years. It contains a short description of what the site is and who it is about, followed by a list of the pages that matter and a line on what each one holds. It is written in a simple format that both people and software can read, and it is there so that an AI system arriving at your site can understand it in one read instead of crawling every page and guessing which parts are important.

Where it came from

The idea was proposed in September 2024 by Jeremy Howard, a well known AI researcher, as a convention rather than a rule: a suggested place and format for a file that gives language models a clean summary of a site. It borrows the thinking behind robots.txt (one agreed location, one simple format) and applies it to the new kind of reader. Since then a fair number of companies have adopted it, particularly for documentation sites, and some AI tooling reads it when it is present. It is not an official standard and no search engine or assistant is obliged to use it.

That last point matters and this post will not pretend otherwise. As of now there is no public confirmation that the major assistants fetch llms.txt automatically when someone asks about you. What is true is that the file costs nothing, cannot hurt, is read by a growing set of agents and tools, and is the clearest single summary of your site that exists, which is useful to a human reader too. Think of it as a sign on the door rather than a lock: it does not force anyone to read it, but anyone who does gets the right answer.

What it is not

It is not robots.txt, which tells crawlers what they may and may not fetch. llms.txt tells readers what is there and what matters; it grants no permissions and blocks nothing.

It is not a sitemap, which lists every page for a search engine. llms.txt is curated: the handful of pages that matter, with a sentence on each.

It is not structured data. The hidden layer inside your pages is machine readable facts in a formal vocabulary. llms.txt is a human readable summary in plain words. They do different jobs and a good site has both.

It is not a place for marketing. An agent reading it wants to know who you are, what you do, where the facts live and how to contact you. Adjectives slow it down.

What goes in one for a person

For a company, llms.txt usually points at product documentation. For a person with a record, it points at the record. The useful contents, in order:

A one line statement of who you are, using your one exact name form, your role and your company, and the address of your canonical page. This is the line an agent will lift if it lifts anything.

A short paragraph on what the site is: your personal record, the official source of facts about you, kept current.

Where each kind of fact lives: the home page for biography and career, the media kit for booking and appearances, the press page for coverage, the writing page for books and newsletters. One line each, with the address.

Your identifier, the address of your Person node, so an agent that understands structured data knows which entity you are.

How to contact you, and how your name is pronounced if that is a common error.

What not to include: anything you would not want quoted, anything the pages do not state, numbers that cannot be checked, and any instruction to the agent about how to behave. Files that try to tell an AI what to say are treated with suspicion by the systems that read them, and rightly so.

An example you can copy

Fourteen lines. Replace every value with your own, keep the shape, save it as llms.txt, and put it at the top of the site. If your site is on a builder, ask whether it lets you upload a file to the root; some do, some do not, and that is one more entry in the list of things a platform makes harder than it should be.

Does it help with Google, ChatGPT and the rest?

Honestly: a little, indirectly, and more over time. The things that decide what the assistants say about you are the canonical page, the hidden layer, the profiles agreeing and the third party corroboration, in that order; how AI search decides who you are covers the mechanism. llms.txt does not replace any of them. What it does is make the site legible in one read to any agent that looks for it, which a growing number do, and it is the cheapest insurance against the next reader that arrives before anyone has written the standard for it. We include one with every record for the same reason we include a sitemap: it is ten minutes of work and it will never be the thing that was missing.

How to check yours is there

Type your address into a browser with /llms.txt on the end, for example yourname.com/llms.txt. If the file appears as plain text, it is live and any agent that looks will find it. If you get a not found page, it is either missing or not at the top of the site. Then read it as a stranger would: does the first line say who you are, does every address in it open, does anything in it disagree with the page it points at? That is the whole test, and it takes two minutes.

Keeping it true

Because it is a summary, it goes stale the moment the thing it summarizes changes. Update it whenever you change a role, add a page or move the media kit. A file that says you are at a company you left is a stale source you wrote yourself, and the whole point of the record is to stop those existing. Check it in the same monthly ten minutes as Search Console.

A checklist