Does llms.txt work?

Last reviewed: 22 October 2026. Author: Camillo Dantas. Sources were opened on that date.

What is llms.txt?

llms.txt is a proposed convention for a Markdown file, placed at /llms.txt, that gives AI agents a short overview of a website and links to the pages most useful to them. Jeremy Howard proposed it on 3 September 2024; the specification page at llmstxt.org was at version 2, last modified on 10 August 2026, when it was opened for this article.

The format is plain. The file starts with an H1 holding the site or project name, which is the only required element. A blockquote with a summary follows, then optional free text, then H2 sections that contain lists of links in the form [name](url), with a short note after a colon if needed. A section called "Optional" holds links an agent can skip when it needs less context.

It is not a standard issued by a standards body. It is a public proposal, and each system decides whether to read it.

Who reads it?

The honest answer is that nobody outside the systems themselves can say. The llmstxt.org page, written by the proposal's author, says that thousands of sites publish the file, that documentation platforms such as Mintlify generate it automatically, that Chrome's Lighthouse audits for it, and that OpenAI, Anthropic and Google's Gemini publish one for their developer documentation. It also expects agents to use it mainly when answering a request, not for training. Those are claims on the proposal's own page; this article did not verify them one by one. The same page names no consumer AI product that reads the file when answering users.

Publishing a file and reading a file are different facts. A company that publishes an llms.txt for its documentation shows that it considers the format useful for developers. It does not show that its search or chat product reads other people's files.

What does Google say?

Google addresses it in two documents. The page on AI features and your website, updated on 10 December 2025, says you do not need new machine-readable files, AI text files or markup to appear in AI Overviews and AI Mode, and that no special schema.org structured data is required. A later guide on optimizing for generative AI features, updated on 10 July 2026, is more direct: Google Search does not use these files, creating them is optional, and maintaining them for other AI services neither helps nor hurts visibility in Google Search, because Search ignores them. The same guide says you do not need to split content into small pieces for AI and that there is no ideal page length.

Two limits apply. What Google says covers Google Search and its AI features. It says nothing about ChatGPT, Claude, Perplexity or Copilot. And "Search ignores it" is a statement about one company's product today, not a proof that no agent anywhere reads the file.

What did the one large study find?

SE Ranking, a vendor of SEO tools, analyzed nearly 300,000 domains. About 10.13 percent had an llms.txt file. Using Spearman correlation and a regression model with contribution analysis, the authors found no sign that having the file relates to how often a domain appears as a source in sampled LLM responses. Removing the variable made their model more accurate, which they read as noise. The page shows the date 15 October 2026 when opened; another source consulted for the Portuguese edition of this site dated the same study 7 November 2025, so the date on the page may reflect an update.

Read it with care:

• It is a vendor study, and the vendor sells tools in this field.
• It is correlational. It does not prove the file is useless everywhere, only that no association appeared in that dataset.
• The authors write that the results are context-dependent and tied to the model and data used.
• It measures citation frequency in sampled answers, which depends on the questions and platforms chosen. It does not measure use by coding agents that fetch documentation on demand.

What is known and what is not?

When is it worth creating?

It is worth it when the cost is low and part of your audience reads documentation through agents: developer documentation, libraries, API references, large help centers. In those cases a curated index of the best pages is a reasonable convenience.

It is not worth treating as the main citation strategy. If your time is short, fixing facts on your key pages, making sure crawlers can reach them, and keeping outside profiles consistent have a better documented return. This site publishes an /llms.txt generated from its pages as a low-cost experiment, with no expectation of effect.

How do you create one, if you decide to?

1. 01List 10 to 30 pages that answer the central questions about your site. Success criterion: each page is public, indexable and answers a question on its own.
2. 02Write the H1 with the official brand name and a one- or two-sentence summary without sales adjectives. Success criterion: the summary says what the site is and who it serves.
3. 03Group links into sections with a one-line description of what each page answers. Success criterion: no link is broken and no description is promotional.
4. 04Publish the file at the root of the domain as plain text, and review it whenever pages change. Success criterion: the list matches the site this month.

Example for a fictional company:

# Model Plumbing

> Residential and commercial plumbing services in Belo Horizonte, Brazil.

## Services
- [Sewer unclogging](https://example.com/unclogging/): scope, response times and what the price includes.
- [Leak detection](https://example.com/leaks/): inspection method and when it is indicated.

## Company
- [About us](https://example.com/about/): technicians, licensing and service areas.
- [FAQ](https://example.com/faq/): response times, payment and warranty.

Do not use the file to give orders to models or to include content that is not visible on the site. If a visitor cannot see it on a page, it does not belong in the file. To build and validate one in your browser, use the llms.txt generator and validator. The decision that actually changes crawler access is the one in robots.txt, covered in the Portuguese guide on robots.txt for AI crawlers.

How would you know whether it made a difference?

Check your server logs for requests to /llms.txt and see which user agents made them. That shows reading, not effect. To look for an effect, run a fixed set of questions before and after, and write down everything else you changed in the period. With a single site you cannot isolate the file from other changes, so avoid building expectations on it. The full method is in the six-step LLM Seeding method, and the Portuguese article llms.txt funciona? carries the same analysis.

Frequently asked questions

Sources opened on 22 October 2026. The company in the example is fictional.