Most explanations of llms.txt stop at the spec. That tells you the format, but not what teams actually ship. So instead of theorizing, we fetched the live /llms.txt files from eight production sites — the day this post was written — and measured them.

Short version: adoption among developer-facing companies is real, the files range from 648 bytes to nearly 200 KB, and almost every one of them follows one of four patterns. If you just want to generate a spec-correct file for your own site, our free llms.txt generator does it in your browser. If you want to understand what everyone else is shipping first, read on.

What llms.txt is, in one paragraph

llms.txt is a proposed convention: a Markdown file at the root of your site that gives AI assistants a curated map of your content — a title, a one-line summary, and sections of links with short descriptions. Where robots.txt tells crawlers what they may not read, llms.txt tells them what they should read, in a format language models parse natively. It is a proposal, not a standard — more on that honest caveat at the end.

The 8 real files, measured

Eight files, three orders of magnitude in size — and every one of them is a legitimate reading of the same spec. The size tells you the strategy: small files curate, huge files enumerate.

The four patterns that repeat

1. The minimal card (llmstxt.org)

The spec site's own file is 648 bytes: an # H1, a > blockquote summary, and a single ## Docs section with three described links. That is the whole file. For a small site — a portfolio, a SaaS landing page, a local business — this is genuinely all you need, and it takes ten minutes.

2. The flat docs index (Supabase, Next.js, Perplexity, Stripe)

Documentation sites list every guide as a link with a one-line description. Two details worth stealing:

• Supabase opens with a pointer to a separate llms-full.txt that contains the complete documentation in a single file — the index for browsing, the full file for ingestion. Vercel does the same, in the very first line.
• Next.js and Supabase link .md mirrors of their pages (/docs/guides/auth.md instead of the HTML page). The model gets clean Markdown instead of a JavaScript-rendered page it has to strip. If your framework can render a Markdown twin of each route, this is the highest-value trick on this list.

3. The federated tree (Cloudflare)

Cloudflare's root file doesn't try to index hundreds of products in one file. Each product line links to its own per-product llms.txt, and the root file explicitly recommends drilling into those. If your site is really several sites in a trenchcoat, this decomposition keeps every file small enough to fit in a context window — which is the point of the whole exercise.

4. The marketing-site pattern (ours)

Developer docs are the easy case — the content is already structured. A services business is messier, and that is the case most agencies and studios actually face. Our own techpotions.com/llms.txt maps the site the way we'd brief a human: what we do (services with one-line descriptions), free tools, real case studies, and recent writing. It is generated from the same content that drives the site, so it never goes stale — we wrote up the 100-line Next.js route handler behind it in this article.

How to create your llms.txt

The anatomy every example above shares:

1. # Site name — one H1, first line.
2. > One-sentence summary — a blockquote directly under it. This is the line assistants quote back, so write it like your elevator pitch.
3. Optional context paragraph — a few sentences of plain prose about what the site is.
4. ## Sections of links — each entry as `- Title: description`. The description matters more than the title; it is what lets a model decide whether to fetch the page.

Three ways to produce it, by effort:

• Hand-write it. Under ~30 pages, this is the honest answer. Copy the structure above.
• Use our generator. Free, runs in your browser, outputs a spec-correct file you can paste at your web root.
• Generate it from your router or CMS. The production answer for sites that change weekly. Ours is a Next.js route handler that reads the same data as our pages, so deploys update it automatically.

The honest caveat

Nobody outside the AI labs can prove today how much llms.txt moves the needle, and skeptics are right that no crawler guarantees it reads the file. What we can verify: companies with the most sophisticated developer-audience SEO on the planet — Stripe, Vercel, Cloudflare — are shipping and maintaining these files in 2026, and AI assistants increasingly cite the sites that make themselves easy to read. The cost of adding one is an afternoon. That asymmetry is the whole argument.

If you want your site legible to the assistants your customers are already asking, generate your llms.txt — and if you want the deeper version wired into your stack, that's literally what we do.