Is there a standard ai.txt file?

No. At least three unrelated proposals use that name or something close to it, including an individual IETF draft and a vendor describing its own product as an industry standard. There is no ratified specification behind any of them, so publishing one is a bet rather than a compliance step.

Should I publish every discovery file I can find?

No. A file at a well-known address makes a claim about your site, and one that nothing reads is a claim nothing backs. Publish the ones whose purpose you can state in a sentence, and leave the rest until something changes about who honours them.

Agent discovery files

Agent discovery files are the files a website publishes at a fixed address to tell AI agents what it is and what may be done with it — and only some of them are backed by anything that deserves the word standard.

This is the most confused corner of the field. Several proposals share similar names, at least one product calls itself an industry standard, and the difference between them is not obvious from the outside. Here is what each one actually is.

Which of these are settled?

Two, and only two, rest on a finished specification. RFC 8615 is what makes `/.well-known/` a legitimate place to publish a declaration at all, and everything below either follows it or ignores it.

`/.well-known/agent-card.json` — defined by the A2A specification, now governed by the Agentic AI Foundation. Settled.
`/.well-known/tdmrep.json` — defined by a W3C Community Group Final Report, which states it is not a W3C Standard. Finished, but narrower in authority than it is usually given credit for.
`/llms.txt` — a proposed convention with no engine publicly committed to reading it. Widely published, not ratified.

Which of these are not what they appear?

There is an IETF Internet-Draft proposing `/.well-known/ai.txt`, authored by one individual. Every Internet-Draft carries the same warning, and it applies here: «It is inappropriate to use Internet-Drafts as reference material or to cite them other than as work in progress».

There are also sites and vendors publishing their own `ai.txt` or discovery endpoint and describing it as a universal standard or an industry standard. Calling something a standard is not a process, and in at least one case the proposal comes bundled with the product that implements it.

One detail settles it: every Internet-Draft expires after 6 months unless somebody renews it. That is the shelf life of the document being cited in places as though it were a specification to build against.

So what should a website actually publish?

Publish the ones whose purpose you can state in a sentence. A robots.txt with explicit per-agent rules, because permission is real. A tdmrep.json if you have a position on mining. An agent card only if you operate an agent. An llms.txt because it costs an hour and forces a useful exercise.

Skip the rest until something changes. A file published at a well-known address that no engine reads is not neutral: it is a claim about your site that nothing backs, and this field already has enough of those.

Where does this come from?

The standard that governs the address, the draft that shows what a draft looks like, and the most recent serious proposal. Read together they make the gradient visible.

RFC 8615, Well-Known Uniform Resource Identifiers
The standard that reserves `/.well-known/`. It is the line between a considered declaration and a file invented at a domain root.
An IETF Internet-Draft proposing ai.txt
Dated 12 June 2026, authored by one individual. Quoted above for the warning every Internet-Draft carries about not being cited as reference material.
Agentic Resource Discovery, Google
The most recent and best-backed proposal in this area, announced in 2026 and using `/.well-known/ai-catalog.json`. New enough that its adoption is not yet a settled fact.
The llms.txt convention
The proposal that started most of this. A de facto convention rather than a ratified specification, and it says so itself.

Frequently asked questions about Agent discovery files

Is there a standard ai.txt file?: No. At least three unrelated proposals use that name or something close to it, including an individual IETF draft and a vendor describing its own product as an industry standard. There is no ratified specification behind any of them, so publishing one is a bet rather than a compliance step.
Should I publish every discovery file I can find?: No. A file at a well-known address makes a claim about your site, and one that nothing reads is a claim nothing backs. Publish the ones whose purpose you can state in a sentence, and leave the rest until something changes about who honours them.

Citability Index