What Is llms.txt? Setup Guide and an Honest Verdict (2026)

llms.txt is the rare GEO topic where most of what's written is more excited than the evidence. This page covers what the file is, how to build a good one, and what the data actually says about whether it matters, including why we maintain one ourselves despite that data.

What it is

Proposed in 2024 as a convention, llms.txt is a markdown file served at /llms.txt, robots.txt's editorial cousin. Where robots.txt tells crawlers what they may fetch, llms.txt tells language models what your site means: a summary of the business, your key claims and numbers, and an annotated list of your most important pages. Some sites add an llms-full.txt with expanded page content for models with big context windows.

How to build a good one

• Open with an H1 and a blockquote summary: who you are, for whom, in two or three quotable sentences.
• State your facts with numbers: pricing, founding year, what you measure, your research findings, exactly the details you want models to repeat correctly.
• List key pages with one-line annotations, canonical URLs only (watch the www and trailing-slash forms).
• Keep it truthful and current. A stale llms.txt is worse than none: ours is generated at build time from the same data that renders the site, after we caught an old hand-maintained version quoting pricing that was five times out of date.
• Don't stuff it. It's a manifest, not a landing page; marketing superlatives make the whole file less credible to a model reading it.

The honest verdict: does it drive citations?

Not measurably, as of 2026. An SE Ranking analysis across roughly 300,000 domains found no significant correlation between having llms.txt and being cited by AI engines, and almost none of the most-cited domains on the web bother with one. Anyone selling llms.txt as a citation lever is selling ahead of the evidence.

Why we keep one anyway

Three reasons. It costs nearly nothing to maintain once generated from a single source of truth. It's insurance for the agent layer: AI agents that visit sites to act on a user's behalf benefit from a machine-readable summary, and that traffic class is growing regardless of whether citation engines ever read the file. And it forces a useful discipline, writing down your canonical facts in one place tends to expose the inconsistencies (pricing, positioning, URLs) that genuinely do hurt you across the surfaces that actually drive citations. File it under hygiene, spend an hour, then put your real effort into structure, freshness, and off-site corroboration. You can read ours at ezgeo.ai/llms.txt.