Does llms.txt do anything? Three studies, and 584 crawler reads of our own

Search for llms.txt and you will find a dozen free generators and validators, each of them a lead magnet for a tool that wants your email. Almost all of them take it as settled that the file helps you get cited by AI answer engines. It is not settled. Three independent studies have looked for an effect and none of them found one, Google has said outright that it does not support the file, and our own server log — every request to this site for the last thirty days — says AI crawlers almost never ask for it.

What llms.txt is meant to do

The proposal is simple and reasonable on its face: publish a markdown file at /llms.txt listing your important pages with a line of description each, so that a language model reading your site gets a clean index instead of having to infer one from your navigation. A companion file, llms-full.txt, holds the whole text of those pages. Think of it as a sitemap written for a reader rather than a crawler.

The idea is good. The question is whether anything actually reads it, and whether publishing one changes whether an engine names you. Those are different questions, and the second is the one you are being sold.

What the three studies found

Three large-scale looks at this have been published, and they disagree with the marketing.

• Ahrefs, across 137,000 sites. Ninety-seven per cent of the llms.txt files they found had never been fetched by anything. The file existed; nothing read it.
• SE Ranking, across roughly 300,000 domains. No statistically significant correlation between having the file and being cited in AI answers.
• Otterly, from ninety days of crawler logs. One tenth of one per cent of AI crawler requests touched llms.txt.

On top of that, Google’s Gary Illyes has said that Google does not support llms.txt and has no plans to. Google is the engine behind AI Overviews and AI Mode, which is to say a large share of the AI answers an Indian buyer actually sees.

The case for it, tested against our own log

There is a case for the file, and it deserves stating properly rather than being waved away. Perplexity is reported to retrieve llms.txt, and Anthropic references it in its guidance on writing for agents. So the claim is not that nobody reads it — it is that two engines might.

We can test the strongest half of that claim directly, because this site publishes an llms.txt and runs a sensor that records every crawler that touches it. Over the thirty days to 18 September 2026, AI crawlers made 584 requests to oftheard.in:

• OpenAI — GPTBot 149, OAI-SearchBot 79, ChatGPT-User 69. One request for llms.txt, on 7 September, and never again.
• Perplexity — PerplexityBot 80, Perplexity-User 43. Not one request for llms.txt.
• Anthropic — ClaudeBot 44. Not one request for llms.txt.
• Others — Applebot 87, Amazonbot 32, Google-Extended 1. Applebot and Amazonbot fetched llms.txt once each.

The engine most often named as the reason to publish an llms.txt made 123 requests to this site and did not once ask for the file.oftheard.in server log, 19 August to 18 September 2026

Three fetches out of 584 requests is half of one per cent, which lands close enough to Otterly’s tenth of a per cent to be the same finding twice. And the three that did fetch it were Applebot, Amazonbot and one visit from GPTBot — not the engines whose answers a brand in India is trying to appear in.

What this is not

One site over thirty days is not a study, and we would rather say so than dress 584 requests up as one. This is a small site. A documentation-heavy developer product with far more traffic might see a different pattern, and an engine may fetch the file in a mode our log cannot distinguish. What this log can tell you is narrow and worth exactly what it is: on one real site that publishes the file, the crawlers of the engines that matter went thirty days without reading it.

It is also not an argument that the file is harmful. It is a few kilobytes and it costs nothing to serve.

What we actually do about it

We publish an llms.txt on this site. It took about twenty minutes, it doubles as a human-readable index of the pages we care about, and if Perplexity or Anthropic start using it in earnest we are already there. That is the honest reason to have one: it is cheap and it might pay off later.

Our own site audit treats it accordingly. llms.txt is checked, reported and not scored. A site without one loses no points, because we cannot show you evidence that it costs you anything. Every check we do score is one where the evidence exists: whether the AI crawlers are allowed in, whether your pages can be fetched and parsed, whether the answer to a buyer’s question is on the page in a form that can be quoted.

What we would push back on is paying an agency for it. It is a short markdown list of your own URLs. If someone quotes you for llms.txt work as a line item in an AI visibility retainer, you are being charged for twenty minutes and a claim nobody has evidenced.

What to do with the twenty minutes instead

If you want the file, write it and move on. The work that has evidence behind it is duller and further up the stack:

• Let the crawlers in. A Disallow aimed at GPTBot, ClaudeBot or PerplexityBot in robots.txt ends the conversation before it starts, and plenty of sites have one by accident. Google-Extended is worth checking too: it governs grounding at answer time in Gemini, not only training.
• Be fetchable. If the answer to a question only appears after JavaScript runs, most of these crawlers will not see it.
• Answer the question on the page. Engines quote passages. A page that states the answer in a sentence gets lifted; a page that builds to it over eight paragraphs does not.
• Be where they already look. Answers cite directories, comparison pages and forums as often as they cite brands. Being absent from the sources an engine already trusts is a bigger hole than a missing text file.