Is llms.txt enough to make a site useful to AI assistants?

No, because llms.txt is an index and not content. It tells a crawler what exists and where, which helps discovery and does nothing for the person who wants your actual thinking inside their own assistant. Anthropic, Cloudflare, Supabase, and Vercel all publish one, and the public directories tracking adoption list over 780 sites doing the same. Publishing the index is now table stakes. The part almost nobody does is publishing the substance: a downloadable corpus of your claims and summaries that someone loads into a project once and works from offline.

Does llms.txt actually affect AI search rankings?

There is no public evidence that any major assistant weights it as a ranking signal, and treating it as one is how it gets oversold. What it reliably does is make a site cheaper to traverse for an agent that has already decided to look, which is a real benefit and a smaller one than the marketing around it suggests. Ship it because it costs an afternoon, not because you expect a rankings lift.

What is the difference between llms.txt and llms-full.txt?

llms.txt is the short index, conventionally under about 80 lines, listing the canonical sections of a site. llms-full.txt is the expanded enumeration of the corpus, every chapter and post with its description. Both are still maps. Neither hands anyone the content in a form they can load and work from, which is the gap a downloadable corpus file fills.

What should a site ship after llms.txt?

Three things, in order of how few sites do them. Markdown at every URL, so any page can be fetched without HTML chrome, which the acceptmarkdown.com convention specifies. A 404 that returns a recovery map instead of a dead end. Then a corpus: one downloadable file carrying your claims and per-piece summaries with canonical links, sized to fit in a context window beside the reader's own work.

Is a corpus file the same as letting AI train on my site?

No, and the license is where you draw that line. The corpus on this site is free for personal and team use with attribution, and explicitly not licensed for republication, resale, or model training. Publishing something loadable is not the same as publishing something absorbable, and saying so in the file rather than only on the page matters, because the file travels without the page.

Is llms.txt Enough for AI Search? No, and Here's Why

The short version

No. llms.txt is an index, not content. It tells a crawler what exists and where, which helps an agent that has already decided to traverse your site and does nothing for the person who wants your actual thinking inside their own assistant. Anthropic, Cloudflare, Supabase, and Vercel all publish one, and the directories tracking adoption list over 780 sites doing the same. That makes it table stakes rather than an edge. The part almost nobody ships is the substance: a downloadable corpus someone loads into a project once and works from offline.

The distinction that matters

There are two different jobs and the discourse keeps merging them.

Discovery is an agent arriving at your site and needing to know what is here. llms.txt, llms-full.txt, a sitemap, and clean markdown representations all serve this. They are maps.

Adoption is a person deciding your thinking should be present in their own assistant, permanently, and needing a file to make that true. Nothing in the llms.txt family serves this. A map is not a thing you load into Claude Projects.

Almost every "AI SEO" checklist covers the first and is silent on the second, which is why so many sites now have an excellent index pointing at content nobody can conveniently take with them.

What this site ships, in order of rarity

Four things, roughly ordered by how few sites do them:

llms.txt and llms-full.txt. 780+ sites. Table stakes.
Markdown at every URL. Append .md to any path, or send Accept: text/markdown, per the acceptmarkdown.com convention. Far rarer, and the single most useful one for an agent mid-task.
A 404 that recovers. A nonexistent path returns a real HTTP 404 whose body carries a short map back to the sitemap, llms.txt, and the six main sections. An agent that guessed a URL wrong finds its way instead of dead-ending.
A corpus. One downloadable markdown file carrying five claims, 41 handbook chapters, 44 agent blueprints, 113 templates, and the whole answer library, each entry a summary with a canonical link. About 194 KB, sized so it fits beside the reader's own work rather than consuming the context window.

The fourth is the one I would argue about. The first three make a site legible to machines passing through. The fourth makes it portable, and portable is what gets cited months later by someone who loaded it once and forgot where it came from, which is exactly why the attribution line lives inside the file and not only on the page.

The honest caveat

None of this is a ranking signal anyone has demonstrated. I ship it because the cost is an afternoon and the failure mode of not shipping it is invisible: you cannot see the assistant that gave up on your site, or the reader who wanted your thinking in their project and had no file to load. Treat the whole category as cheap insurance and an unusually direct way to be useful, not as a growth tactic with a measurable curve.

Sources: llms.txt directory, sites using llms.txt, acceptmarkdown.com, The Falkster Corpus.

Is llms.txt enough to make a site useful to AI assistants?

The short version

The distinction that matters

What this site ships, in order of rarity

The honest caveat

SOURCES

THE LONG VERSION

RELATED ANSWERS