A hint file in the llmstxt.org shape, worth twenty minutes of your time. It is not a ranking lever, and anyone selling it as one is selling you something.
/llms.txt is a Markdown file at the root of your site that tells a language model, in plain text, what your site is and which pages are worth reading. It was proposed by Jeremy Howard in 2024, it takes twenty minutes to write, and it is surrounded by more nonsense than anything else in this field.
So let us do the honest version.
What it looks like
# Northstar Analytics
> Product analytics for B2B SaaS teams, focused on activation and retention.
> Self-serve from $49/month, SOC 2 Type II, EU data residency available.
Northstar is used by product and growth teams to see which features drive
retention. It is not a general BI tool and does not replace a warehouse.
## Docs
- [Quickstart](https://northstar.example/docs/quickstart): install and first event in ten minutes
- [Event schema](https://northstar.example/docs/events): how to structure events you send
- [API reference](https://northstar.example/docs/api): REST and webhook endpoints
## Product
- [Pricing](https://northstar.example/pricing): plans, limits and what counts as an event
- [Security](https://northstar.example/security): SOC 2, subprocessors, data residency
## Optional
- [Changelog](https://northstar.example/changelog): release history
The shape from the llmstxt.org spec: an H1 with your name, a blockquote summary, optional prose, then H2 sections of annotated links. An ## Optional section at the end marks links that can be skipped when context is tight.
What it is genuinely good for
Being a clean, unambiguous definition of your company in one place. No navigation, no cookie banner, no hero animation. If something reads it, it gets the facts with no extraction step. That is worth having regardless of who reads it.
Curation. You are stating which of your four hundred pages actually matter. That is information nobody else can supply, and it is more useful than a sitemap, which lists everything with no ranking.
Being cheap. Twenty minutes, one file, no risk. Even at a low probability of paying off, the expected value is fine.
What it is not
It is not a ranking signal. No assistant has committed to reading it. Adding the file does not make ChatGPT recommend you, and no plan you buy from anyone — including us — changes that. We say this in the product, in the FAQ, and here, because the alternative is selling a lie.
It is not a substitute for the actual work. If your robots.txt blocks the crawler, or your homepage renders empty without JavaScript, or you have no Organization schema, llms.txt fixes none of that. It is the last twenty minutes of the job, not the first.
It is not a control surface. Nothing in the file constrains what a model does with your content. It is a hint, in the same family as a sitemap: helpful to whoever chooses to use it, ignored by everyone else.
It is not robots.txt. Different file, different purpose, no access control whatsoever. Do not put anything in it you would not publish.
Writing a good one
Write the summary as a definition. The blockquote is the highest-value part of the file. "Product analytics for B2B SaaS teams" is a definition. "The leading platform empowering modern teams to unlock insights" is not, and a model asked what you do will produce something equally vague.
Annotate every link. A bare URL list is a worse sitemap. The sentence after the colon is what makes the link useful — it tells a reader whether to spend a fetch on it.
Include the awkward facts. Pricing, limits, what you do not do. "Not a general BI tool" is one of the most useful sentences in the example above, because it stops you being recommended for the wrong thing and then getting a churned customer.
Keep it under a couple of hundred lines. This is a table of contents, not a knowledge base. Twenty to forty well-chosen links beats four hundred.
Serve it as text/plain or text/markdown at /llms.txt, with a 200. A 404 is the common failure; a redirect to your homepage is the second most common.
The companion file
The spec also describes llms-full.txt — the full text of your key pages concatenated into one document, so a model can ingest everything without following links. It makes sense for documentation sites and rarely for anyone else. If you generate it, generate it from source at build time rather than maintaining it by hand; a stale full-text dump is worse than none.
Should you bother?
Yes, once the rest is done. Rank it below crawler access, below Organization schema, below a readable About page, and above almost everything else — because those three take an afternoon each and this takes twenty minutes.
Our llms.txt generator will read your sitemap and navigation, rank the pages worth listing, and hand you a draft to edit. It works with no AI key configured at all, because the ranking is deterministic. Edit what it gives you — particularly the summary. That sentence is yours to write.