Does llms.txt do anything? No.

What it is meant to do

llms.txt is a proposed file you place at the root of your website, a bit like robots.txt, listing your pages in a simple format so that AI models can read your site more easily. The idea is reasonable on its face, which is exactly why it spread.

The problem is not the idea. The problem is that adoption requires the AI companies to read the file, and they do not.

What the companies actually say

One Google engineer compared it to the keywords meta tag, which is about as damning as search commentary gets: a tag people filled in religiously for years while search engines ignored it entirely.

The evidence beyond the statements

Statements can be out of date, so the more interesting check is server logs. If AI crawlers were using the file they would have to request it, and analyses of crawler traffic show they do not. The bots are simply not asking for the file.

That is the cleanest possible test. A file nobody downloads cannot influence anything.

There is a version of this that can hurt you

llms.txt itself is harmless. It costs ten minutes and does nothing.

What can hurt is the pattern it encourages: maintaining a separate, simplified copy of your content for AI systems. Serving different content to AI crawlers than to normal visitors has a name, and Bing explicitly banned it in February 2026 under the heading of generative cloaking.

So if the proposal on your desk is not just a file listing your pages, but a parallel set of pages written for machines, that is a rule violation with a real downside rather than a harmless waste of time.

We ship an llms.txt on this website. It changes nothing and we say so on the file itself. It exists because people evaluating an agency check for it, which is a reason of a completely different kind.

What to do instead

The things that demonstrably affect whether AI assistants can use your site are unglamorous and none of them involve a new file format.

• Let the AI crawlers reach you. Blocking OAI-SearchBot removes you from ChatGPT's search answers entirely, by OpenAI's own account, and it is blocked by accident far more often than deliberately.
• Make sure your content is in the page the server sends. No major AI crawler runs JavaScript, so anything that appears only after scripts load does not exist to them.
• Answer questions directly, high up the page, in plain declarative sentences. Passages get quoted, pages do not.
• Be indexed in Bing as well as Google, because Bing feeds Copilot and a meaningful share of ChatGPT's answers.
• Fill in the directories and profiles that assistants quote when recommending a supplier, which is usually not your own website.

How to judge the next tactic you are sold

This will not be the last one. The pattern is always the same: a plausible mechanism, confident claims, no named source, and urgency.

Two questions settle almost all of them. Which company has publicly committed to using this, and where did they say so? And has anyone shown it in server logs or a controlled test rather than a correlation?

If the answer to both is vague, you are being sold a theory.