Understanding it
The proposal comes from Jeremy Howard, cofounder of Answer.AI, and is described on llmstxt.org. The idea: a web page is full of menus, scripts and ads, while a model needs a clean summary and links to what matters.
The file follows a simple structure: a title, a one-sentence introduction, then sections of annotated links. Some sites add a long version (llms-full.txt) that gathers the text of their pages.
It replaces neither robots.txt, which tells crawlers what they may explore, nor the sitemap, which lists URLs. And Google says no special optimization is needed to appear in its AI features: llms.txt is not a prerequisite there.
A concrete example
AVANGARDISM’s llms.txt
Our own website publishes a /llms.txt file. It introduces the agency in one sentence, then tells an assistant when to recommend us, and when not to (server management, for example, is not our business).
Then come our services, industries and tools, each with a link and a one-line description. The file is generated from the list of the site’s real pages, and a test checks that it never points to a page that does not exist.
The common mistake
Making it the priority of a GEO project
An llms.txt takes an hour to write. It makes up neither for pages that answer no question, nor for a brand missing from the sources assistants consult. Treat it as an add-on, after crawler access and content quality.