Having an llms.txt file is step one. Having a good one is what makes it worth the bother. A thin, auto-generated link dump tells a model nothing it could not have guessed. A curated file with clear descriptions does real work.
This check looks at substance, not just presence.
What “good” looks like
The whole point of llms.txt is curation. You are choosing, on the model’s behalf, the small set of pages that explain your business, and telling it why each one matters. That is exactly the judgment an auto-generated file skips.
Curate, do not dump. Link to your most important pages: what you do, pricing, proof, how to reach you. Not your entire sitemap. If everything is listed, nothing is prioritized, and you have just handed the model your navigation menu in a different format.
Write a real description after every link. The format is [Page title](url): one-line description. That description is the part a model uses to decide whether the page is worth fetching. “Pricing: plans and what each tier includes” earns its place. A bare link does not.
Lead with a clear summary. The blockquote and opening lines should state, in plain language, who you are, what you do, and who you do it for. If a model read only those two sentences, it should still place you correctly.
Use the ## Optional section on purpose. Put genuinely secondary links there. It tells a short-on-context model exactly what it can drop first, which means your essential links survive.
Point to clean versions where you can. If you serve Markdown variants of your key pages, link to those, so the destination is as readable as the index.
What “thin” looks like
These are the patterns this check flags:
- An auto-generated file that mirrors your full sitemap with no descriptions.
- Links with no notes, so the model cannot tell a pricing page from a press release.
- A vague or missing summary, so your identity has to be inferred from URLs.
- Everything marked essential, nothing marked optional.
A quick rule of thumb
Read your llms.txt as if you were the model. From the summary and the link descriptions alone, could you answer “who is this, what do they sell, and where is the proof?” If yes, it is doing its job. If you have to open the links to find out, it is too thin.
One caveat carried over from the llms.txt guide: no major AI engine has confirmed it reads these files in production yet. Quality here is about being ready and being clear, not a guaranteed ranking lever. The same curation work pays off anywhere a model reads your site, so it is not wasted either way.