llms.txt is a plain-text file you publish at yourdomain.com/llms.txt. It gives a language model a short, curated map of your site: what the business is, what it costs, and which pages matter. It was proposed by Jeremy Howard (Answer.AI) in September 2024. A small business does not strictly need one, but it costs nothing, removes ambiguity, and every site we build ships one.
Key takeaways
- An llms.txt is a public markdown file at /llms.txt. It is not a secret, not a ranking signal, and not a replacement for a readable website.
- Put in it the things a stranger would need to describe your business correctly: what you do, where, what it costs, which pages to read. Leave out anything you would not put on a billboard.
- Adoption by the major assistants is not confirmed. Do not expect rankings from it. Expect fewer wrong answers about you, if anything.
- It takes about twenty minutes to write, and the exercise of writing it usually shows you what your site itself is failing to say.
Where the idea comes from
The proposal lives at llmstxt.org, written by Jeremy Howard and published on 3 September 2024. Its opening line is short: "We propose adding a /llms.txt markdown file to websites to provide LLM-friendly content."
The reasoning is practical. A language model has a limited window in which to read, and a normal web page is mostly not information: navigation, scripts, cookie banners, footers. A human filters that out; a model spends its window on it. The answer is one plain markdown file that says the important things first and points to the pages that hold the detail.
It sits alongside two files you may already know. robots.txt is permissions, sitemap.xml is a list of URLs. llms.txt is neither; it is closer to a well-written "about us" page, addressed to a machine, with links.
What the format is
The specification at llmstxt.org lists the pieces in order. The only required part is "An H1 with the name of the project or site. This is the only required section." After that come "A blockquote with a short summary", then "Zero or more markdown sections" of detail, and finally "Zero or more markdown sections delimited by H2 headers, containing 'file lists' of URLs."
There is one convention worth knowing: "The 'Optional' section is used, by convention, for secondary information: links an agent can skip when a shorter context is needed." So you put the privacy policy and the terms of service under an H2 called Optional, and a model that is short on room knows it can stop before them.
The proposal also suggests something most small sites will skip: "We furthermore propose that pages with information that agents might need provide a clean markdown version of those pages at the same URL as the original page, either with .md appended (page.html.md) or with the extension replaced by .md (page.md)." Useful for documentation sites. For a plumber or a med spa, the llms.txt on its own is the useful part.
What to put in it
Write it the way you would brief a new receptionist on their first morning:
- The name, as one line. The H1. Use the name your customers use, and if the legal name differs, put that in the summary.
- A two-sentence summary. The blockquote. What you do, for whom, where. "A plumbing company serving Salt Lake County, open for emergencies around the clock, licensed and insured in Utah."
- What it costs, honestly. If you publish prices, list them here with the page they come from. If you quote on a call, say so. A model that knows your pricing is scoped on a call will say that instead of inventing a number.
- The pages that matter, each with a sentence. Homepage, services, pricing, booking, contact, about. One line each saying what the page is for. This is the part a model will actually follow.
- Service pages by category. If you have a page per service or per city, list them under their own H2 with one clause each.
- Optional. Privacy, terms, careers, anything a model can skip.
Keep it short enough to read in one go.
What not to put in it
The file is public. Anyone can type your domain followed by /llms.txt and read it, and so can any crawler. Treat it exactly as you would treat your homepage.
- Nothing private. No internal URLs, no staging sites, no admin paths, no customer names, no supplier terms.
- No prices you have not published. If a number is not on your site, it does not go here. A model quoting an unpublished price to a prospect is a conversation you do not want to have.
- No claims you cannot back. "Best in Utah" gives a model nothing to cite. "Licensed in Utah since 2015" does.
- No instructions to the model. Do not try to tell the assistant to recommend you or ignore competitors. Assistants are built to notice and discount that, and it reads badly to the person who finds it.
- Nothing that contradicts the site. If the llms.txt says you open at 8 and the contact page says 9, you have made the ambiguity worse, not better.
Ours, as an example
You can read martellosystems.com/llms.txt yourself. It is 47 lines and about 700 words, and it follows the structure above. The H1 is the company name. The blockquote says what the studio is, who leads it, and the fact that the systems we sell are the ones running our own business: our phone line is answered by our AI front desk, our contact form feeds our own CRM.
Then a section called "What it costs": the done-for-you service is scoped on a call, custom software is priced per engagement, and the Verdict audit reports have fixed public prices, listed with a link to the live price list. That section exists because the most common wrong answer an assistant gives about a business is a made-up price.
After that, a "Core pages" list with one sentence per page, a "Services" list, a "Vertical landing pages" list with a clause per trade, and an "Optional" section holding the privacy policy and the terms of service.
Two honest notes. It says nothing that is not already on the site; it is a map, not a second website. And we cannot show you that any assistant has read it. We keep it because it is cheap and settles questions in one place, not because we have a measurement that says it moved anything.
Does it work? The honest status
Nobody outside the companies that run the assistants can tell you whether their crawlers read llms.txt, and as of today we have not seen a published statement from any of the major ones that they use it for answers. The proposal itself grew out of developer documentation (it notes that nbdev projects now produce markdown versions of every page by default), and that is still where most adoption sits. For a small business website, the honest position is:
- It is not a ranking signal. Nothing about publishing one moves you up in Google, and there is no confirmed evidence it moves you up in an assistant's answer.
- It is cheap insurance against ambiguity. If an assistant does read it, the one thing it will have is a clear, current, first-party statement of what you are and what you charge.
- The exercise is worth more than the file. Writing it forces you to answer, in a few hundred words, the questions a customer asks first. If you find that hard, your site is probably not answering them either, and that is the real problem.
We include it in the AI visibility audit as one check among several: does it exist, does it agree with the site, does it leak anything. It is never the headline finding. That is nearly always plainer: the site cannot be read without JavaScript, or the phone number on Google does not match the website.
How to publish one in twenty minutes
- Open a text file. Write the H1, the blockquote, the pricing section and the page list, in that order.
- Read it as a stranger. Anything private, unpublished or unprovable comes out.
- Save it as llms.txt, plain text, UTF-8, and put it at the root of your site so it answers at yourdomain.com/llms.txt. On a site builder, look for a static-file upload or ask whoever runs the site.
- Open the URL in a browser and check it loads as plain text, not a download and not a 404.
- Re-read it whenever your prices, hours or services change. A stale llms.txt is worse than none.
FAQ
Is llms.txt the same as robots.txt?
No. robots.txt tells crawlers which parts of your site they may fetch. llms.txt tells a language model what your site is about and which pages to read. One is permissions, the other is a summary. You can have both, and they do not conflict.
Will an llms.txt get my business into ChatGPT's answers?
There is no evidence that it will on its own, and nobody should sell it to you that way. What gets a business into an assistant's answer is a readable site, structured data, a consistent name, address and phone number across the web, and recent Google reviews. llms.txt is a small, free addition to that list, not a substitute for it.
Does Google read llms.txt?
Google has not said that it uses llms.txt for search or for AI Overviews. Treat it as unread by Google until Google says otherwise. Your sitemap.xml and your structured data are what Google has documented.
Can an llms.txt hurt me?
Only if you put something in it that should not be public, or something that contradicts your site. A file that lists a starting price when your pricing page says "quoted on a call" gives an assistant two answers and a reason to trust neither. Keep it consistent and keep it public-safe and it cannot hurt.
Do I need markdown copies of every page as well?
The proposal suggests them for sites with a lot of documentation. A small business site does not need them. One short llms.txt that is accurate and current is the whole job.
Where this fits
An llms.txt is one small piece of being findable when people ask an assistant instead of searching. The larger piece is answer engine optimization: answer-first pages, schema, consistent details, reviews and being cited by the sources assistants read. If you want to know where you stand first, the AI visibility audit checks what the assistants say about you today, what a crawler can see on your site, and whether your llms.txt exists and says the right thing.
If you would rather talk it through, book 15 minutes with Artie. Bring your domain; you will leave knowing what your llms.txt should say and whether it is the thing worth fixing first.