llms.txt: The $0 File That Might Matter (And Might Not). Here's What the Data Actually Says.

The Part of This That's Real, and the Part That's Theater

Let's start with the awkward truth: there's a file called llms.txt that the internet has decided matters. Adoption among top-million domains climbed from 0.015% at the start of 2025 to 10.13% by early 2026 — nearly a 700x jump in twelve months. Anthropic publishes one. Vercel publishes one. Stripe, Cloudflare, and Mintlify all publish one.

Also true: Google's John Mueller said in mid-2025, on the record, "No AI system currently uses llms.txt." A Search Engine Land study tested nine sites that implemented it — eight saw no measurable traffic change. No major LLM provider has publicly confirmed real-time ingestion of the file.

So which is it? Essential infrastructure for the AI era, or cargo-cult SEO?

Here's the honest answer after reading the research, auditing the public examples, and running tests on roast.page's own properties: llms.txt is a low-cost, low-risk, high-optionality bet. It won't move your traffic next week. It probably won't move it next quarter. But the people betting against it are the same people who bet against robots.txt in 1994 and schema.org in 2011. Both turned out to be infrastructure.

This post walks through what llms.txt actually is, what the public examples look like, the template that works in 2026, and the decisions you need to make before publishing one.

What llms.txt Actually Is (30 Seconds)

llms.txt is a plain Markdown file that lives at yourdomain.com/llms.txt. It's a sibling to robots.txt and sitemap.xml, with a different audience: large language models.

The structure is minimal. The first line is an H1 with your site or product name. The second line, usually a blockquote, is a one-line description of what you do. Then a short "About" section. Then H2 sections listing your most important pages, each with a Markdown link and an optional one-line summary.

Here's a miniature example:

That's it. No JSON. No schema. No XML. Just Markdown a human can read and a machine can trivially parse. The design is intentionally dumb — that's what gives it staying power.

Why the Big Players Are Publishing It Anyway

If no AI system officially ingests it, why do Anthropic, Vercel, and Stripe all have one?

Three reasons, in order of importance:

1. It's a forcing function for information architecture

Writing llms.txt is a brutally honest exercise in asking "what are the ten pages on my site that actually matter." Most teams can't answer this without an argument. The process of writing llms.txt is more valuable than the file itself: it surfaces stale content, reveals navigation gaps, and forces a team to agree on canonical URLs.

Anthropic's public llms.txt, for instance, is a clean map of their API docs and product pages. Writing it meant their docs team had to stop and decide which guides were load-bearing vs. nice-to-have. That decision propagates into nav, into internal linking, into ads. The file itself is a byproduct.

2. It's a signal of technical maturity

Developers check llms.txt the way they check robots.txt. It tells them something about the seriousness of the team. A developer-facing company without one in 2026 looks dated the same way a site without a sitemap looked dated in 2011. This isn't a ranking factor — it's a credibility factor.

3. It's pre-positioning for inevitable standardization

The robots.txt convention took two years to spread and another five to become universal. schema.org was a 2011 announcement that took six years to become table stakes. llms.txt is in year two. Anthropic, OpenAI, and Google have all published AI crawler specifications. It's near-certain that some version of llms.txt becomes part of the standard — either as-is or with formal extensions.

The teams publishing today are paying a one-hour cost for a multi-year option. That math is hard to argue with.

Teardown: How Anthropic, Vercel, and Stripe Structure Theirs

The three best public llms.txt examples right now are from Anthropic, Vercel, and Stripe. Each takes a slightly different approach. Studying them is faster than reading the spec.

Anthropic: the pared-down docs map

Anthropic's llms.txt (and its companion llms-full.txt) focuses on their developer documentation. The structure is lean:

• H1: "Claude Docs"
• Blockquote: One line describing the API.
• H2 sections: "Getting started", "Build with Claude", "API reference", "Agents and tools", "Policies".
• Each H2 contains 3–8 Markdown links, each with a one-line summary.

The discipline is aggressive. Anthropic has hundreds of docs pages. Their llms.txt lists maybe 40. They're picking the entry points, not the long tail. This is the right instinct.

Vercel: the multi-product hierarchy

Vercel sells Next.js hosting, serverless functions, an AI SDK, a CLI, and more. Their llms.txt mirrors that product surface:

• Separate H2 sections for each major product (Next.js, AI SDK, CLI, etc.).
• Top-level entry points only — no deep links.
• A separate llms-full.txt for the full docs content.

This works for companies with multiple products. If that's you, don't flatten everything into one list. Let each H2 map to a product and link only to the canonical overview page for each.

Stripe: the commerce index

Stripe's llms.txt is notable because it covers more than just docs. It indexes:

• Getting-started guides for their top products
• API reference entry points
• Integration partner directories
• The newly public ACP (Agentic Commerce Protocol) spec — which Stripe co-authored with OpenAI

Stripe's file is larger than Anthropic's but still ruthlessly curated. The lesson: llms.txt isn't a dumping ground. It's a whitelist of what you want AI to treat as authoritative.

The Template That Works in 2026

Here's the structure we recommend after auditing roughly 40 public implementations. You can paste this into a file called llms.txt at the root of your site and be 80% of the way to a good implementation in ten minutes.

Four rules that separate a good llms.txt from a bad one

1. Absolute URLs, always. Relative paths work in most parsers, but agents sometimes fetch the file in isolation. Fully qualified URLs remove ambiguity.
2. One-line summaries beat long descriptions. Agents skim. A sharp ten-word description is more valuable than a paragraph of prose.
3. Ten to forty links total. Not 200. Not five. The point is to curate the entry points to your site, not to mirror your sitemap.
4. Update it quarterly, minimum. Stale links are worse than no links. Put it on the same cadence as your sitemap review.

llms.txt vs. llms-full.txt: What's the Difference?

llms-full.txt is a variant developed by Mintlify in collaboration with Anthropic. Where llms.txt is an index of your important pages, llms-full.txt is the full text of those pages compiled into one Markdown file. Developers can paste its URL into ChatGPT, Claude, or any RAG pipeline to load your entire knowledge base as context in one shot.

You should publish llms-full.txt if:

• You run a developer-facing product with extensive docs.
• Your docs get asked about frequently in AI chat (you can check by asking ChatGPT/Claude "how do I use [your product]?" and seeing whether they cite you).
• You want developers to be able to spin up context quickly when working with your API.

You probably don't need llms-full.txt if:

• You're a pure marketing site with a blog and a few product pages.
• Your total site content is under ~50,000 words. At that size, llms.txt alone is enough.
• Your content is heavily visual (images, videos) and doesn't compile cleanly to Markdown.

For roast.page, we publish both: llms.txt points to our canonical pages and blog posts; llms-full.txt compiles our research articles and methodology so developers building analysis tools can load our methodology into a prompt in one line.

The 15-Minute Implementation

Here's the practical walkthrough, start to finish:

1. Open a blank file. Call it llms.txt. It's just Markdown.
2. Write your H1 and blockquote. Your site name, then a one-sentence description.
3. List your top pages. Group them under H2s — About, Product, Docs, Blog, Policies. Include 10–40 links. Every link gets a one-line summary.
4. Save it at the root of your site. yourdomain.com/llms.txt must return plain text with content-type text/plain or text/markdown. Most static hosts (Vercel, Netlify, Cloudflare Pages) handle this automatically.
5. Add a link to it from your robots.txt (optional but useful): Sitemap: https://yourdomain.com/llms.txt (or as a separate line). This is non-standard but costs nothing and helps some crawlers discover it.
6. Test retrieval. Fetch curl https://yourdomain.com/llms.txt. Confirm it returns clean Markdown with the expected content-type.
7. Set a quarterly review reminder. Broken links in llms.txt are worse than no llms.txt. Keep it fresh.

If you use Mintlify, Vercel's AI SDK, or Cloudflare's AI toolkit, there are free generators that will scaffold this for you. For most teams, writing it by hand is faster and produces a sharper result.

What llms.txt Cannot Do (And What to Stop Expecting)

A lot of the hype is junk. Here's what llms.txt does not do, based on public statements from AI providers and current observable behavior:

• It does not force AI systems to cite you. Citation decisions come from content quality, authority signals, and the AI's retrieval ranking — not your llms.txt.
• It does not replace schema.org markup. Structured data still matters, especially for products, reviews, and FAQs. Agents parsing your commerce pages rely on Schema.org Product and Offer, not on your llms.txt.
• It is not read in real-time by any major AI today. ChatGPT, Claude, Gemini, and Perplexity don't fetch llms.txt during a conversation. They may use it during training crawls, but that's months of lag.
• It does not improve your Google rankings. Zero direct ranking impact, per Google.
• It does not fix poor content. If the pages you list are thin, slow, or poorly written, pointing AI at them makes the problem more visible, not less.

If someone tells you llms.txt is a growth hack, ignore them. The teams winning at AI-era visibility aren't winning because of llms.txt — they're winning because their content is clear, their pages are technically clean, and their claims are specific enough for an agent to extract and cite.

The Real Question: Is Your Content Worth Citing?

The conversation about llms.txt is a distraction from the real question: is the content on the pages you'd list actually good enough to be cited?

Our work with thousands of pages through roast.page suggests the answer is usually no. The most common failures we see:

• Vague value propositions. Pages that don't clearly state what the product does, for whom, and at what price. Agents need extractable specifics. See our deep-dive on invisible value propositions.
• No statistics or data. Research by Princeton and Georgia Tech showed that adding statistics to content boosts AI citation rates by 41%. Most marketing pages have zero.
• Missing FAQ sections. Pages with 10+ FAQ questions show 156% higher citation rates. Most landing pages have none, or three pitch-y ones.
• Features buried in images and JS. Agents can't extract what they can't read. If your differentiators are inside a Figma export, they don't exist.
• Stale updates. Pages not updated quarterly are 3x more likely to lose AI citations over time.

Fix these five things and your llms.txt becomes useful — because the pages it points to are worth pointing at. Skip them and you can have the best-structured llms.txt in the industry; no agent will care.

When llms.txt Stops Being Optional

Here's the scenario worth planning for: by late 2026 or early 2027, llms.txt becomes standard infrastructure. We're nearly certain about this because the adoption curve is following a textbook pattern — first the hyperscalers adopt (Anthropic, Vercel), then the developer tools (Stripe, Mintlify), then mainstream SaaS, then everyone.

When that happens, not having one will look like not having a sitemap in 2014 — unusual, slightly negligent, and worth a one-hour fix.

You don't have to publish one today to avoid that outcome. But the teams that publish a good one today will have six months of compounded discipline advantage over the teams that wait. And unlike most "be early" bets, this one has no downside. There's no SEO penalty. There's no cost. There's no reputational risk. It's a file. It sits there.

Your Move

The playbook:

1. Write your llms.txt today. One hour. Use the template above. Publish it at your root. Move on.
2. Then go fix your content. Not because llms.txt demands it — because an agent reading your page won't cite it if your value proposition is muddy, your pricing is hidden, and your claims are generic. That's where the real visibility comes from.
3. Audit in 90 days. Has AI-referred traffic grown? Has your brand shown up in ChatGPT/Claude/Perplexity answers more often? If yes, publish llms-full.txt. If no, focus on the content itself, not the index file.

The best signal that your landing page is AI-ready is simple: run it through roast.page and look at your technical signals score. If your page parses cleanly, states its value in the first 200 words, and carries real structured data, your llms.txt has something worth listing. If it doesn't, fix that first. The file comes last, not first.

Skeptics are right that llms.txt is overhyped. Optimists are right that it's coming. Both can be true. Your job is the boring middle: publish the file, keep it current, and make sure the pages it points to are genuinely worth an AI's attention.