AI search
llms.txt: the honest answer
A file that tells language models what your site is about. Google has declined to support it, the citation studies show nothing, and it is still probably worth the half day it takes.
6 min read
llms.txt is a plain markdown file at the root of a site — /llms.txt — that summarises what the business is, what it offers, and where the substance lives. It exists so a language model does not have to infer any of that from rendered HTML and navigation.
It is being sold hard at the moment, usually with the implication that it improves your standing in AI answers. That claim does not survive contact with the evidence.
What the evidence says
~10%
Adoption across a survey of 300,000 domains
None
Measurable citation lift found in controlled studies
Declined
Google's position on supporting the file
Google has stated it does not support llms.txt and does not plan to, with one of its search staff comparing it to the keywords meta tag — a file publishers fill in about themselves, which is exactly why search engines stopped trusting that one. The major crawlers do not request it in meaningful volume. One modelling study found that removing llms.txt as a variable actually improved the model's ability to predict citations, meaning the file was contributing noise rather than signal.
If someone tells you llms.txt will get you into ChatGPT, they have not read the studies. It is plumbing, not leverage.
Why we ship one anyway
The case for it is not the citation case. It is narrower and more honest.
- Some systems do read it
- Anthropic recommends it in its own guidance for writing for agents, and Perplexity has said it retrieves the file to help prioritise pages. That is not universal support, but it is not nothing.
- The agent web is a different audience
- Coding assistants, MCP servers, and automated tools fetch these files constantly. If any part of your business is likely to be used through an agent rather than a browser, the file has a real reader today.
- It costs half a day
- Generated from the same source as the site, it cannot drift out of date and needs no maintenance. The downside is a few hours; the upside is optionality if adoption shifts.
- Writing it is clarifying
- Compressing a business into one page of plain statements exposes vagueness fast. Several of our clients found the exercise more useful than the file.
What to put in it
- What the business is, in one sentence, at the top.
- What it actually sells, named the way a customer would name it.
- How pricing works — not necessarily the numbers, but the shape.
- Links to the pages that hold the substance, each with a line saying what is there.
- Anything you do not want misstated. Ours says the studio publishes no performance figures, so nothing invents them on our behalf.
The correct posture on llms.txt is mild. Ship one, keep it truthful, generate it from your real content so it stays accurate, and do not expect it to move anything on its own. If adoption changes, you are already there. If it does not, you spent an afternoon.
This is one small part of the AI visibility work, not the whole of it.
How Grow worksFollow-up
Questions this raises.
- Where does llms.txt go?
- At the root of the domain, served as plain text: example.com/llms.txt. It is a markdown file, not a config format, and it is public — anyone can read it, so it should say nothing you would not print on the site itself.
- Does Google use llms.txt?
- No. Google has said it does not support the file and does not plan to, with one of its search staff comparing it to the keywords meta tag — a self-declared file that search engines learned not to trust.
- Does anything read it?
- Some things do. Anthropic recommends it in its guidance for writing for agents, and Perplexity has said it retrieves the file to help prioritise pages. Coding assistants and agent tooling fetch these files routinely. The major web crawlers largely do not.
- Is llms.txt worth doing?
- Probably, on the grounds that it costs about half a day and gives you optionality if adoption shifts. It is not worth doing on the grounds that it will improve citations, because the controlled studies do not show that.