Does llms.txt actually work? What the server logs say in 2026

llms.txt does not currently affect whether you appear in ChatGPT, Perplexity, or Google's AI Overviews. Google has said its Search systems do not use the file, and the largest public log study found the overwhelming majority of llms.txt files were never requested at all. It does reach one real audience, which is coding agents fetching a site's docs on purpose.

I serve one on this site, and I built the free llms.txt generator that writes them. So I have every reason to want the answer to be yes. It is worth being clear about what the evidence actually supports, because the cost of getting this wrong is not the ten minutes. It is the month somebody spends on machine-readable files while the thing that does move citations sits untouched.

Does llms.txt help you show up in ChatGPT, Perplexity, or AI Overviews?

Not in any way anyone has measured. The biggest dataset so far is an Ahrefs study of 137,210 domains, published on June 15, 2026. In May 2026, 97% of the llms.txt files it tracked received zero requests of any kind. Of the requests that did arrive, only 19.5% came from AI bots.

The search engines say the same thing. Google's AI features guide states that you do not need new machine-readable files, AI text files, or markup to appear in its AI features. In June 2025, Google's John Mueller wrote on Bluesky that "no AI system currently uses llms.txt", and added that your server logs make it obvious.

What is llms.txt, and who proposed it?

llms.txt is a Markdown file at the root of a site that gives language models a short, curated map: what the site is, which pages matter, and where the plain-text versions live. Jeremy Howard published the proposal at llmstxt.org on September 3, 2024. It is a proposal, not a standard. No standards body has adopted it, and no search engine has promised to read it.

The idea is sound. HTML is noisy, context windows are finite, and a site owner knows better than a crawler which ten pages matter. The problem is not the idea. The systems people want citations from have simply not opted in.

Who actually reads llms.txt?

Mostly bots that are not AI tools. In the Ahrefs data, 77% of the bots fetching llms.txt were SEO audit tools, tech profilers, and other crawlers. Among AI tools, agents lead: aside from GPTBot, at 4.51% of requests, and one agent-infrastructure bot, Claude Code out-fetched every AI retrieval bot, every AI assistant, and every AI training crawler.

That pattern fits how people use these tools. A developer tells Claude Code or Cursor to read a library's docs, and the agent looks for the curated version first. Anthropic serves one for its own developer docs for exactly this reason. If your buyers are developers working inside an AI coding tool, llms.txt reaches them. If your buyers ask ChatGPT for a recommendation, it probably does not.

So is llms.txt worth shipping?

Yes as a courtesy, no as a visibility tactic. Here is how I weigh it:

The one real risk is drift. An llms.txt that lists old prices or a dead refund policy hands every agent that does read it the wrong facts. If you ship one, keep it next to the pages it describes and update both in the same change. The line-by-line walkthrough of ours shows what goes in it.

How to check whether yours is being read

You do not have to take a study's word for it. The file lives at a fixed path, so its traffic is trivially separable in your own logs, and this is the rare SEO question you can settle with evidence from your own server in ten minutes.

Filter your access logs for requests to /llms.txt, then group by user agent. Three things are worth counting:

1. Total requests per month. If this is zero, everything below is moot and you have your answer.
2. The share from AI user agents rather than SEO crawlers and tech profilers. In the Ahrefs data that share was small, and your own ratio is what matters for your site.
3. Whether the same agents also fetched real pages. An agent that reads llms.txt and then fetches three linked pages is using the map. One that reads the file and leaves was checking whether it exists.

On a Vercel or Cloudflare deployment this is a log query rather than a project. If you cannot see logs at all, that is a more useful thing to fix than the file itself, because every other question in this space is also a log question.

The mistake this file encourages

Here is the part I find more interesting than the verdict.

llms.txt is appealing because it converts a hard, open-ended problem into a small, closed, satisfying one. Getting cited by AI systems is uncomfortable work: it means writing passages that answer a question cleanly, attaching real numbers, and being an entity a model has encountered before. Writing a markdown file and committing it is none of those things, and it produces the feeling of having done something.

That is a category of error worth naming, because it is not specific to this file. Machine-readable artifacts feel like infrastructure, and infrastructure feels like leverage. But a map is only valuable to a reader who wanted to go somewhere. Publishing one to an audience that has not shown up yet is a way of avoiding the question of why they have not.

The test I use before shipping anything in this category: can I name the specific system that reads it, and can I verify in a log that it did? For llms.txt the honest answer is yes to coding agents and no to answer engines, and that is a perfectly good reason to ship it, as long as you do not then bill it to the wrong budget.

What moves AI citations instead?

The things AI engines demonstrably fetch: your actual pages. That means a robots.txt that lets the search bots in (OAI-SearchBot, PerplexityBot, Googlebot), content that renders without JavaScript, and passages that answer one question cleanly. The AI crawler checker tests the first part in a few seconds. The GEO guide for small shops covers the rest.

I am keeping my llms.txt. It is cheap, it helps developers who point agents at this site, and it will be ready if a search engine ever opts in. I just do not count it as a reason anyone finds us.