We have a free llms.txt generator, and we still wrote this. That should tell you we are not here to hype the file. We are here to tell you what it does, what it does not, and why it is still worth ten minutes.
Quick recap, since we covered the basics elsewhere. An llms.txt is a plain-text file at your site root that names your site, summarizes what you do, and links the pages you most want models to read. Think of it as a polite table of contents aimed at AI. The idea is reasonable. The question is whether anything actually reads it.
What the evidence says
Not much reads it, and Google is now explicit. The honest state of play in 2026:
- ✓Ahrefs studied 137,000 sites and found roughly 97% of llms.txt files are never fetched by an AI crawler, on the order of one fetch in a thousand.
- ✓Google's current generative AI optimization guide says Google Search ignores llms.txt files.
- ✓No major engine, OpenAI, Anthropic, Google, or Perplexity, has confirmed it uses the file to decide what to recommend.
- ✓Confusingly, tooling has started adding llms.txt checks anyway, which makes teams think it is a ranking factor when the platforms say it is not.
An emerging convention is not the same as a ranking signal. llms.txt is the former, and the marketing around it keeps pretending it is the latter.
So should you ship one?
Only if a tool you use consumes it, you want to run a documented experiment, or the summary has another real operational use. Skipping it is reasonable. A file with stale claims creates maintenance work without a confirmed search benefit.
The tell for an overhyped AEO tactic is a confident promise with no platform behind it. llms.txt is optional, and Ron no longer includes its presence in the readiness score.
What to do instead
If your goal is to become easier to find and cite, follow the fundamentals current platform guidance supports: useful and unique pages, clear visible facts, crawlable links, a healthy sitemap, current evidence, consistent entity information, and structured data only when it matches the page.
Use crawler and rendering tools to find real access failures, then measure recommendations and citations directly. Permission and markup are inputs, not proof of an outcome.