Less than the enthusiasts claim, more than nothing. That is the verdict on llms.txt, and the details fit in about six hundred words.
What it is
llms.txt is a proposed convention: a plain-text file at the root of your domain that gives language models a curated map of your site. Think of it as a younger cousin of robots.txt, except that instead of telling crawlers what to avoid, it tells models what matters. Who you are, what the site contains, where the important pages live, written in simple markdown. One file, no tooling, no build step. The format is deliberately boring, which is the best thing about it.
What it is not
It is not a ranking lever. No major engine has published a commitment to honor it, adoption across engines is uneven, and evidence that it changes retrieval or citation behavior is thin to absent. Shipping llms.txt does not cause inclusion in AI answers, and anyone selling it as the switch that turns citations on is selling weather control. Treat every breathless llms.txt case study with the suspicion it has earned.
Why we ship it anyway
Because the trade is lopsided. It costs about an hour. It has no downside we have been able to find. And it adds one more location where your company's facts appear in consistent, machine-friendly form, which matters for entity resolution: the process by which an engine concludes that your domain, your brand name, and the name you retired in the rebrand are all the same organization. Engines assemble identity from many weak signals, and this is a cheap one, so we take it. An hour of work for a marginal signal is a good trade. A week of work for a marginal signal would be a hobby.
What goes in it
Keep it short and factual. Four things earn a place:
- Your canonical company name, including former names if you have rebranded, so models connect old references to the current entity.
- Your core pages: what you do, who it is for, and the handful of pages you would want a model to read first.
- Your key proof: the recognition, certifications, or flagship work you want attached to the entity.
- Contact and location facts, stated exactly the way they appear everywhere else.
That last clause is the real instruction. The file is a claim about your identity, and it only helps when it agrees with your schema, your about page, your directory listings, and everything else the engine has already seen. A novel fact that appears only in llms.txt is not a feature. It is a contradiction waiting to be noticed.
The warning
llms.txt without crawler access, schema, and citable content is a menu for a restaurant with no kitchen. If your robots rules block the AI crawlers, the menu never gets read. If the listed pages carry no structured data, the dishes have no recipes. If the content itself is a wall of adjectives with nothing liftable in it, there is nothing to serve. The file is the last five percent of a machine-readable company, not the first five, and the other ninety-five is the part that pays; we laid all of it out in The Machine-Readable Company.
Ours
We run ours at dantontensix.com/llms.txt. It took an afternoon, and most of that was the argument over which proof points made the cut. It has not transformed anything, and we did not expect it to. It is one more brick, laid level, in a wall that only works as a wall.
If you would rather talk through the first ninety-five percent, we are an easy conversation.