Compare
ARD vs llms.txt
Filed together because both are files on your domain aimed at machines. That is where the similarity ends.
They are not alternatives. llms.txt is unschematised Markdown suggesting which pages a model should read, with no query interface and, as of 2026, no published evidence that any major system uses it. ard.json is typed JSON describing callable capabilities, backed by a required search API and live registries that crawl it today. Publish both if you like; only one of them currently does something you can measure.
They are not the same kind of thing
Both are files you put on your domain to help machines understand you. That is the entire similarity, and it is doing a lot of work in most comparisons.
ARD (ard.json) | llms.txt | |
|---|---|---|
| Format | Typed JSON with a published schema | Markdown, no schema |
| What it describes | Callable capabilities and how to reach them | Which of your pages an LLM should read |
| Consumers | Registries and ARD clients, several live today | None demonstrated at measurable scale |
| Query interface | Required: POST /search | None |
| Validation | Conformance tool, errors and warnings | Nothing to validate against |
| Backing | Google, Microsoft, Hugging Face, GoDaddy working group | A community proposal |
| Evidence it changes anything | Measurable: entries in an index get returned in searches | None found by the three studies that looked |
What the evidence actually says about llms.txt
This deserves stating plainly because a lot of advice does not. Independent investigations by SE
Ranking, Contentful and Mintlify all reached the same place: no widely validated evidence that
llms.txt reliably improves visibility in AI answers. Google has said it is not needed.
Our own prior look across a large sample of bot traffic found no measurable effect either.
So should you skip it? None of that means do not publish one. It costs ten minutes and it might become useful. It means do not spend a week on it, and do not let anyone sell it to you as an AI visibility strategy. This site publishes an llms.txt on exactly that basis.
The difference that actually matters
llms.txt is a hint to a crawler about content. ard.json is an entry in
a searchable index of things that can be invoked. One tries to influence what a model reads about
you; the other makes your software callable by a machine that had never heard of you thirty seconds
earlier.
If your goal is being cited in AI answers, llms.txt is the wrong tool and so is
ARD. That is a content and off-site problem. If your goal is having your MCP server or agent found
and used at runtime, ARD is the mechanism that exists.
Questions
Does llms.txt work?
There is no published evidence that it does. SE Ranking, Contentful and Mintlify have each investigated and reported no measurable effect on AI visibility, and Google has publicly said you do not need one. That is not proof it can never work, but it is the state of the evidence in 2026.
Should I publish both?
Sure, they cost nothing and they do not conflict. Just do not treat them as equivalent. One is read by working registries today; the other is a proposal nobody has demonstrated a consumer for.
Is ARD going to end up like llms.txt?
It might. The honest difference right now is that ARD has named implementers, a conformance tool, a required API and several live registries, where llms.txt has a convention and a lot of blog posts. Watch whether the registries keep crawling.
Last reviewed 2026-09-04. Checked against ARD v0.91 (Proposal, 2026-08-26).