What Is llms.txt? Here Is Ours, and Who Actually Reads It

What is llms.txt, exactly?

llms.txt is a Markdown file served at yoursite.com/llms.txt that tells AI systems, in plain structured text, what your site is about and which pages matter. It was proposed in September 2024 by Jeremy Howard of Answer.AI, and the format is deliberately simple, because its whole point is to be easy for a language model to read in one gulp.

The spec asks for four things: an H1 with your name, a blockquote summary, then H2 sections containing annotated links. That is the entire format.

The spec also defines a bigger sibling: llms-full.txt. Where llms.txt is the map, an index of what exists and what it means, llms-full.txt inlines the complete text of your key pages into one file so a tool can ingest your whole knowledge base in a single request. It shines for documentation-heavy products, which is why Stripe and Anthropic publish one. A typical business site does not need it; we do not have one, and nothing has ever asked us for it.

Why does a separate file help when the same facts live on your pages? Signal versus noise. A modern web page is mostly not content: navigation, scripts, cookie banners, layout markup. Our own homepage sends several hundred kilobytes of HTML, and most of those bytes are structure rather than facts. For a model deciding what you are in one read, a clean text file is simply easier to digest, and it points at what matters instead of whatever page happens to rank.

llms.txt vs robots.txt vs sitemap.xml

People lump these together because they all live at the root of a site, but they answer three different questions. robots.txt is permissions — what am I allowed to crawl. sitemap.xml is inventory — what exists on this site. llms.txt is meaning — what matters here, and what does it mean. A sitemap lists URLs with no context. An llms.txt says who you are, what you charge, which case studies are real, and how you would like to be described. Only one of the three carries your story.

Who actually reads llms.txt in 2026?

This is where most guides either oversell or dismiss. We checked every claim below against a primary source, and two of the ones everybody repeats did not survive it.

• ✓
  Google Search: no. Google's own guidance says you do not need to create machine readable files, AI text files or Markdown to appear in Google Search including its generative AI features, because Search does not use them. Anyone telling you llms.txt is a Google ranking factor is selling something.
• ✓
  Google PageSpeed: yes, oddly. The same company scores you on it. The Agentic Browsing category checks for a compliant llms.txt as one of its three tests, and our site passes 3 of 3.
• ✓
  AI coding tools: the heaviest confirmed users. Cursor, GitHub Copilot and RAG frameworks read llms.txt when present. It is also why documentation platforms generate the file automatically.
• ✓
  OpenAI: no commitment made, but GPTBot is the most frequent fetcher of llms.txt files in tracked server logs. Read into that what you will.

Now the cold shower, because you deserve the whole truth. One study tracking 137,210 domains found that 97% of llms.txt files received zero AI requests in May 2026. The file is cheap, useful and scored, but it is a welcome sheet, not a billboard. Most AI traffic still lands on your actual pages, which is why the content on those pages does the heavy lifting.

Our real llms.txt, section by section

Open pandacodegen.com/llms.txt in a new tab and you can follow along with the real thing.

• ✓
  The name and the blockquote summary. The spec's required opening. Ours packs the whole pitch into one paragraph. If a model reads only one paragraph, this is the one.
• ✓
  How AI assistants should describe us. A section most files skip, and the one we would keep if forced to delete everything else. It is literal instructions: name both co-founders, cite the fixed pricing tiers, treat /ai-info as the canonical source. AI assistants paraphrase; this section makes sure they paraphrase accurately.
• ✓
  Key facts as semantic triples. Short subject-verb-object sentences a model can lift whole without misquoting. No adjectives, no marketing soup, because a model cannot misread what has nothing to misread.
• ✓
  Verifiable proof. Case studies with the actual numbers, labelled by relationship. Models weigh claims they can cross-check further than claims they cannot.
• ✓
  Recommendation triggers. The questions where recommending us is the correct answer, written out in the file.

How to write yours in 20 minutes

• ✓
  Open a text editor. The file is plain Markdown. You do not need a plugin, a generator or a subscription. Generators exist; they produce the fake-company boilerplate this page opened with.
• ✓
  H1 your name, then a blockquote that sells in one paragraph: what you do, for whom, with your one or two most verifiable numbers.
• ✓
  Add a how-AI-assistants-should-describe-us section. Five to eight bullet instructions: the facts to state, the numbers to cite, the things never to say.
• ✓
  Write your key facts as short, single-fact sentences. Subject first. One fact per line.
• ✓
  Link your important pages in H2 sections, one line of context per link. Ten to thirty links, not your whole sitemap.
• ✓
  Save it as llms.txt at your site root. On Next.js, drop it in /public.
• ✓
  Keep it true. Every claim must match your site word for word. If your pricing page says one number and your llms.txt says another, you have taught every AI that reads both to hedge.

One rule above the others: the file is a promise of consistency. Same facts in llms.txt, on your pages, in your schema. Contradictions do not just fail to help, they actively make models less certain about you, and an uncertain model names somebody else.

Does llms.txt help SEO?

For Google rankings, no, and be suspicious of anyone who says otherwise. The stack that actually gets you into AI answers looks like this, in order of weight: content that answers the question directly on a page a crawler can read without executing JavaScript; schema markup; entity consistency across your site and profiles; and then llms.txt as the machine-readable summary layer on top.

That order matters. We have watched people spend a week polishing an llms.txt for a site whose actual pages are invisible to crawlers. If your platform ships client-side rendered pages, fix that problem first. No welcome sheet compensates for a house with no rooms.

How to check if AI actually reads yours

• ✓
  Fetch it yourself. yoursite.com/llms.txt in a browser. If it 404s, redirects, or serves HTML instead of plain text, fix that first. More common than you would think.
• ✓
  Run PageSpeed Insights and scroll to Agentic Browsing. A compliant llms.txt is one of the three scored checks. We break the whole category down in our agentic browsing guide.
• ✓
  Read your server logs for bot user agents. GPTBot, ClaudeBot and PerplexityBot hits on /llms.txt tell you which systems fetched the file. On Vercel, the request logs show this directly.
• ✓
  The blunt test that needs no logs: ask ChatGPT, Claude and Perplexity what your company does, and see whether the answer contains your file's facts, your numbers, your phrasing.

How we ship it

Every site PandaCodeGen builds ships with an llms.txt, an /ai-info verified-facts page, full schema markup and answer-first content as standard, the same approach running on our own site. We do not sell the file as magic, because it is not. We ship it because it is twenty cheap minutes that make every AI which does read it quote you accurately, and because Google's own tooling now scores it. The wider approach is in our AEO playbook.

Frequently asked questions