Its own documentation calls it a proposal and says the specification is open for community input. That is not a criticism of the idea; it simply means nothing is obliged to read one, and advice presenting it as settled has gone further than its authors have.

Do we need one to appear in AI answers?

One major search provider says explicitly that no new machine-readable files or AI text files are needed for its AI features. Other systems are a separate question, and any answer about them should come with a source rather than an assumption.

Should we add one anyway?

It is an hour of work that cannot hurt, so treat it as a cheap option on a convention that may or may not settle. What it should not be is a project with a claimed return, because nobody can currently evidence what that return would be.

Glossary

llms.txt

A proposed file at a site's root that points AI systems at its most useful content, written in plain Markdown rather than a formal format.

In plain terms

A file you can put at the top level of a site saying, in plain language, which pages are worth reading and what they are for. It is a suggestion somebody made that others may or may not follow, not a rule anybody set.

Why it matters

Because it is being recommended as though it were established, and its own site describes it as a proposal open for community input. Knowing which of those you are looking at decides whether adding one is a sensible small experiment or a task somebody should be paid for.

How it works

It is a proposal to standardise on a file that helps these systems use a website at the moment they are answering. That framing matters: it is about being useful to something reading your site now, rather than about how anything is indexed.

It sits at the root path of a site, optionally in a subpath, which is the same convention other machine-readable files at a site's root have used for decades. The location is the least novel thing about it.

It uses Markdown rather than a formal structured format, which its own documentation calls unusual. That makes it readable and writable by a person without tooling, and it is a deliberate choice rather than an oversight.

Its status is a proposal open for community input, in its own words. Nothing about that makes it a bad idea; it does mean that describing it as a standard, or implying that anything is obliged to read it, goes further than its authors do.

One major search provider states plainly that you do not need to create new machine-readable files or AI text files to appear in its AI features. That is a documented statement from one side rather than a verdict on the proposal, and it is the single most useful thing to know before spending time on this.

Which makes it a cheap experiment rather than a project. Writing a short file listing your genuinely useful pages costs an hour, cannot hurt, and should be treated as an option on a convention that may or may not settle, rather than as work that has a known return.

How it is described, by its authors and by advice about it

The pattern is worth recognising beyond this one file, because it recurs whenever a sensible idea appears in a fast-moving area. A proposal is published with careful language about its status, it is written about by people summarising it, the qualifications are the first thing lost in summary, and within months it is being recommended as established practice by people who have never read the original and have no reason to doubt what they were told. Nobody in that chain is being dishonest. The defence is the same one that applies to any claim in this field and costs almost nothing: find what the thing says about itself. Here that took a single visit to the proposal's own site, where the word proposal appears throughout and the specification is described as open for community input. Five minutes of reading turns a decision about whether to comply with a standard into a decision about whether to spend an hour on an experiment, which is a much easier decision to make correctly.

Seen in the wild

A short file at a site's root listing the pages worth reading and why.
Perplexity
A documentation site pointing at its own reference pages in plain Markdown.
Hugging Face
Advice to add one presented as a requirement rather than as a proposal.
Writer

Common misconceptions

People assume

It is a standard that AI systems follow.

In fact

Its own site describes it as a proposal, open for community input. That is a description of where it is in its life rather than a criticism, and it is different from a convention anything is obliged to honour.

People assume

You need one to appear in AI answers.

In fact

One major search provider states that no new machine-readable files or AI text files are needed to appear in its AI features. That is one side rather than all of them, and it is a documented statement rather than an inference.

Questions

Is it a standard?: Its own documentation calls it a proposal and says the specification is open for community input. That is not a criticism of the idea; it simply means nothing is obliged to read one, and advice presenting it as settled has gone further than its authors have.
Do we need one to appear in AI answers?: One major search provider says explicitly that no new machine-readable files or AI text files are needed for its AI features. Other systems are a separate question, and any answer about them should come with a source rather than an assumption.
Should we add one anyway?: It is an hour of work that cannot hurt, so treat it as a cheap option on a convention that may or may not settle. What it should not be is a project with a claimed return, because nobody can currently evidence what that return would be.