What exactly goes in it?
A Markdown file with a conventional structure: the site name as the title, a paragraph summarizing what the site is and who it serves, and sections of annotated links to the pages that matter.
What separates it from a sitemap is the curation. A sitemap lists everything without telling the essential from the incidental; llms.txt is a selection with context — closer to explaining your site to someone than to handing them an inventory.
How is it different from robots.txt and the sitemap?
| File | Answers | Format |
|---|---|---|
| robots.txt | What are you allowed to access? | Permission rules |
| sitemap.xml | Which URLs exist? | Exhaustive list |
| llms.txt | What is this site, and what matters here? | Curated Markdown |
The three complement each other and none replaces another. Having an llms.txt without a sitemap or structured data is starting from the end.
Is there evidence that it works?
Nothing conclusive, and that is worth saying plainly. It is a community proposal that has gained traction among technical sites, but the major model providers have not publicly confirmed that they use it as a crawling or comprehension signal.
Anyone telling you llms.txt improves your visibility in ChatGPT is claiming something that cannot be demonstrated today. The honest argument is a different one: it takes twenty minutes, it breaks nothing, and if the convention does settle, you already have it.
When does it make sense to implement?
Once the things that are demonstrated are already done: content that is reachable without JavaScript, crawlers allowed in, consistent structured data, and content that answers real questions.
If any of those is missing, our advice is to leave it for later. It is the last 2% of the AEO work, not the first, and mistaking it for a solution is the most common error among people who have just discovered it.
How do you keep it from going stale?
By generating it rather than writing it by hand. An llms.txt written manually is out of date by the first batch of new content, and a site map that lies is worse than no map at all.
The one on this site is generated from the same data that builds the site, so every publication updates it without anyone stepping in. It is the only way a file like this survives longer than three months.
And if what you are missing are the things that are demonstrated, that is what we look at first in a Digital Audit.