Where llms.txt comes from
The format was proposed in 2024 as a convention for helping language models use websites at inference time, when an assistant reads pages to answer a question. It is written in Markdown: a title, a one-paragraph summary, then sections of links with short descriptions.
It is not an official web standard, and AI providers do not all say whether or how they read it. Treat it as a low-cost signal, not a ranking switch.
What a good llms.txt contains
Start with the company name and a two or three sentence description that matches the one on your about page word for word in substance. Add the facts a model gets wrong most often: what you offer, where you work, which languages you serve and how to contact you.
Then list your key pages with one line each: services, products, case studies, pricing or process pages and your most useful articles. Some sites also publish an llms-full.txt with longer context. Ours, for example, includes a product catalog and notes on what should not be inferred.
What it does not do
It does not override what your pages say. If the file claims something your site does not support, a careful model will trust the page. It does not unblock crawlers either; that is the job of robots.txt.
And it will not fix a thin website. A model that finds a tidy summary pointing to pages with nothing in them has little to cite.
How to publish it
Serve it at yourdomain.com/llms.txt as plain text, keep it short and update it whenever your services or key pages change. If your site is built with a framework such as Next.js, generate it from the same data as your sitemap so the two never drift apart.
Where it fits in GEO
llms.txt is one item on a longer list: consistent entity facts, structured data, answer-ready pages and crawler access. Our GEO guide covers the rest, and our technical SEO and GEO service puts it into practice.