llms.txt is a dud, and that's fine

If a vendor is selling you an llms.txt audit, save your money. The verdict is in, and it’s not close: Google’s John Mueller and Gary Illyes have both said the search team doesn’t use the file and has no plans to, and the bot logs back them up. In one analysis of more than half a billion LLM-bot traffic events, the share of requests touching /llms.txt was statistically negligible. The file is, as one write-up put it, almost untouched by the bots that would matter.

Why it fizzled

The pitch sounded reasonable: a tidy Markdown file telling AI crawlers what your site is about and which pages matter. The problem is that it asks the engines to trust a self-description they can’t verify — and they already learned that lesson. Mueller compared llms.txt to the old keywords meta tag, the one search engines abandoned because it was trivially gamed: declare whatever you want, show the bot one thing and humans another. That’s not a standard, that’s a cloaking vector with good manners.

Engines don’t need you to summarize your site. They can read it. A file that adds a manipulation surface without adding verifiable information was never going to get adopted by the parties it was meant to influence.

“But my llms.txt got picked up!”

You’ll find the occasional post claiming a freshly submitted llms.txt started “powering AI answers” days later. Be skeptical of the causation. If the same content is on your normal pages — and it should be — the engines got it from there. Correlating a citation with the file you happened to add the same week is exactly the kind of mistake proper measurement exists to catch. One screenshot is not a result.

What to do with the ten minutes you saved

Adding llms.txt is harmless — we put one on client sites because it costs nothing and signals care, like a tidy robots.txt. Just don’t mistake hygiene for leverage. The things that actually move AI visibility are less novel and more work:

• Answer-first formatting — lead with the answer, add a table, be specific enough to quote.
• Structured data that matches the page — the one place you state facts in a machine format, validated and consistent with your visible content.
• Freshness — engines visibly prefer recently updated sources on anything time-sensitive.
• Corroboration — the same claims echoed on sources the engines already trust, which is what turns your assertion into a fact.

The thing worth watching instead

Mueller has said he prefers WebMCP, a Google-backed approach for sites to expose structured actions and data to agents. Whether it goes anywhere is an open question — but it’s a more serious proposal than a text file the engines can’t trust. Keep an eye on standards the platforms actually back. Ignore the ones being sold as a shortcut. In AI search, the shortcut is almost always the tell.