Publish llms.txt only if you run docs, and expect nothing from search

Publish llms.txt for docs, and expect nothing from AI search.

What to do: If you have developer docs, auto-generate an llms.txt file, which platforms like Mintlify produce for free. If you do not have docs, skip it and move on.

Why it works: It helps coding assistants and people who paste your docs into a model. It does not get you cited in AI search, no matter what the GEO crowd sells.

Example: As of July 2026 I checked, and Anthropic, Stripe, Vercel, Cursor, and Bun all serve a real /llms.txt. But Ahrefs found almost none of these files ever get fetched by an AI crawler, and Google's own team has publicly called it speculative and unused. Publish it as cheap hygiene, not as a growth play.

Walk it through

I ran this against three real docs sites and one real AI query in July 2026. Here is exactly what came back.

1. Check that the file is real, and weigh it.

curl -s https://stripe.com/llms.txt | wc -c
curl -s https://vercel.com/llms.txt | wc -c
curl -s -L https://docs.anthropic.com/llms.txt | wc -c

Stripe's file is 64,229 bytes. Vercel's is 193,824. Anthropic's runs 202,909, and only after a redirect, because docs.anthropic.com now forwards straight to platform.claude.com. Anthropic moved its dev docs to a new domain at some point this year, and the redirect chain is the only place that shows up in public. Three companies, three file sizes, no shared template. Nobody is grading these against a spec, because there isn't one that anyone enforces.

2. Open one and read what's actually inside.

curl -s https://stripe.com/llms.txt | head -6

It is a plain markdown index. A one-line company description, a link to the full-docs version, then section headers with a bullet per page, each one a link plus a one-sentence summary. No keywords stuffed in, no schema, nothing built to game a ranking. It reads like a table of contents because that is exactly what it is, a table of contents built for a model to fetch instead of crawl.

3. Ask an AI search engine if any of this buys you a citation.

I put the question straight to Perplexity: does llms.txt help my site get cited by ChatGPT or Perplexity.

Its own answer: "maybe a little for Perplexity, but not reliably for ChatGPT," pointing to one study that found higher citation rates after adding the file and another analysis of 300,000+ domains that found no statistically significant lift. That is the AI search engine grading its own file format, live, and it lands exactly where the verified numbers in this chapter's card already put it.

The read

• No two files agree on size or structure. 64,229 bytes to 202,909, because there is no enforced spec, just a convention a few docs platforms adopted on their own.
• Every publisher I found is a developer tool. Stripe, Vercel, Anthropic, this is a docs-team habit, not a marketing move, and the audience is coding assistants pulling in context, not a shopper asking Perplexity where to buy something.
• Perplexity hedges on its own file format. "Maybe a little, not reliably" is the honest state of the evidence in July 2026, not a marketer's spin, and it matches what Ahrefs and Google's own team already said.

Steal it

If you already maintain developer docs, generate the file this afternoon. Mintlify and a few other docs platforms will build it straight from your existing page tree, or you write a short script that dumps every doc URL, its title, and its first sentence into one markdown file at /llms.txt. It costs nothing, it helps Claude Code and Cursor and anyone who pastes your API reference into a model, and it will never hurt you.

If you do not run docs, do not build any just to have the file. And do not let anyone sell you an "llms.txt optimization" package. There is no ranking signal inside the file to optimize. It is a static list a crawler either fetches or ignores, and right now almost none of them fetch it.

Gotchas

• Auto-generated files rot fast. Wire the generator into your deploy pipeline, not a quarterly task, or it goes stale the moment you ship a new doc section.
• "llms.txt SEO" is not a real service. If someone pitches you an audit or a retainer for this file, you are paying for a checkbox, not traffic.
• This could change, honestly. The evidence in July 2026 says do not expect a citation lift from it. If AI crawlers start actually consuming these files at scale, that verdict has an expiration date, so recheck before you build a whole content strategy on ignoring it.