llms.txt: what it is, what it's really for, and how to generate it automatically
llms.txt is a file proposed as a standard to give AI models a curated map of a site's content, in markdown, at the root of the domain. What it contains, how it is structured, what can realistically be expected by those who adopt it today — no engine has officially confirmed its use — and how AraLith COS generates it automatically from governed content, for every language of the site.
llms.txt: what it is, what it's really for and how to generate it automatically
The honest guide to a file everyone is talking about: what it promises, what is confirmed and what is not.
llms.txt is a markdown-formatted text file, published at the root of the site, that offers language models a curated map of the content: who you are, what you talk about, which pages matter. It is a proposed standard (llmstxt.org), not a requirement: no major AI engine has officially confirmed its use. It is still worth understanding it, and if it costs nothing to generate it, publish it.
In summary: what is llms.txt?
llms.txt is a markdown file published at /llms.txt of your domain, designed to give language models (ChatGPT, Claude, Perplexity and the like) a curated, readable version of your site: a description of the business, a reasoned list of important pages, and essential context. The idea stems from a public proposal in 2024 (llmstxt.org) modeled after robots.txt and sitemap.xml: a conventional entry point, but for AIs instead of crawlers.
- What it is: a curated map of content in markdown, at the root of the site, with an extended variant per language or per section.
- What it is NOT: an officially adopted standard. No major AI engine has publicly confirmed using it for retrieval, and that needs to be said before anyone sells you anything.
- Why publish it anyway: the cost is zero if generated automatically, the risk is zero, and if adoption comes you are already ready. It's the same logic as structured data ten years ago.
- The real value today: in the relationship between llms.txt and SEO, the exercise of curating the map — deciding what your site says it is — is entity SEO work that pays off even if no engine ever reads the file.
What does an llms.txt file look like?
If you are wondering how to create the llms.txt file, the proposed structure is intentionally simple: a title with the site name, a one-line summary, a short description, and link sections with one line of context each. A minimal example:
# AraLith
> Italian software house, since 2007: custom development and AraLith COS.
## Services
- [Custom web development](https://www.aralith.com/sviluppo-sito-web-personalizzato): tailor-made websites and web apps
- [Website quote](https://www.aralith.com/preventivo-sito-web): how the reasoned quote works
## Guides
- [AI Overview Optimization](https://www.aralith.com/guide/ottimizzazione-ai-overview-guida-2027): operational guide 2027
- [GEO vs SEO](https://www.aralith.com/semantic-ia-seo/geo-vs-seo-differenze): the differences that matterThe rules that matter when you write it (or generate it):
- Curation, not completeness: it's not a sitemap. It includes the pages that define who you are and what you want to be cited for — not every URL.
- One line of context per link: that's what the model reads to understand whether the page answers the user's question.
- Consistency with the actual site: if the file says something different from the pages, you've created noise, not signal. It must be updated whenever the content changes.
- Multilingual: a site in multiple languages needs a manifest per language, with the correct canonical URLs — the same language-prefix principle that applies to all internal links.
Is it really needed? The state of adoption, without hype
Most articles you find out there are misleading, so let's lay out the facts:
What is true: the interest is real — the search "llms.txt" records around 33,000 monthly searches globally with very low competition (0.12; Ubersuggest data, August 2026 survey) — and many technical sites have published it.
What has NOT been confirmed: no major AI engine has officially stated that it uses llms.txt to select the sources for its responses. Google's public representatives have repeatedly downplayed it. Anyone who promises you that publishing it "gets you cited by AI" is selling a hope as if it were a mechanism.
The reasonable reading: it's a zero-cost bet. If generating it is automatic, there are few reasons not to publish it; reasons to expect miracles, none. The work that actually moves AI citations today is the one on structure and entities, and llms.txt is at most a complement to it.
How AraLith COS generates it (without writing it by hand)
In AraLith COS the llms.txt file is not written: it is generated from governed content. The system automatically builds the manifests — the main one and one for each active language of the site — from three already curated sources:
- the brand profile (who you are, what you do, with what communication constraints);
- the approved knowledge items from the knowledge base: only verified information, with the selection you decide;
- the published content and the topic coverage calculated on the template of your business type.
The advantage is not the file itself: it is consistency. The manifest says exactly what the site and the knowledge base say, in the right languages with the right canonical URLs, and updates when content changes — without manual maintenance.
Frequently asked questions about llms.txt
No: it is a public proposal (llmstxt.org, 2024) modelled after robots.txt. There is no body that has ratified it, nor any AI engine that has officially confirmed its use in retrieval. It is an emerging convention, with spontaneous adoption by many technical websites.
No one can promise that, and anyone who does should raise your suspicions. The documented mechanisms that influence AI citations are different: answer-first structure, recognisable entities, verifiable data, authority. llms.txt is a zero-cost complement, not a primary lever.
robots.txt tells crawlers what they can visit; sitemap.xml lists all URLs for indexing; llms.txt proposes a curated, annotated selection of the content that matters, designed to be read by a language model. The first two are established standards that engines actually use; the third is a reasonable bet on the future.
If you write it by hand, yes — and that is why many llms.txt files out there are already outdated. If it is generated automatically from your content, as in AraLith COS, it stays aligned on its own: that is the difference between a file and a system.
An always up-to-date llms.txt, without ever writing it
AraLith COS generates llms.txt manifests for every language of the site from governed content: brand profile, verified knowledge base and published pages, always consistent.
