If you've followed agent-readiness guidance for a year, you have a pile of conflicting advice. One source tells you to publish llms.txt. Another says /.well-known/ai-plugin.json. The newest says WebMCP and /.well-known/webmcp. They sound related. They're not interchangeable. They do different jobs, ship in different stages of standardization, and have wildly different actual adoption.
This is the honest field comparison for an SMB site owner deciding which one to ship in 2026.
The three manifests, in one paragraph each
llms.txt is a markdown file at the site root that tells AI crawlers what your site is for and points them at the canonical pages. Anthropic published the spec in late 2024. Format: a single .md file with a title, summary, and bulleted resource list. Consumed by: AI training and citation crawlers, when they choose to fetch it. Status: informal spec, no formal standards body. Adoption: under 11% of public sites as of the May 2026 SERanking 300K-domain study. Citation lift in the same study: not statistically measurable.
/.well-known/ai-plugin.json is the OpenAI ChatGPT plugin manifest from 2023. It declared what a plugin did and where its API was, so ChatGPT could call your endpoints. Status: OpenAI deprecated the plugin platform in early 2024 in favor of GPTs and then function calling. The well-known path still works mechanically, but no major agent platform consumes it as a discovery primitive in 2026. The format influenced later manifests but is itself a dead branch.
/.well-known/webmcp is the WebMCP discovery file from the Chrome / W3C Community Group Draft that shipped in Canary on February 10, 2026. JSON manifest declaring what actions an agent running in the user's browser can take on your site. Consumed by: any WebMCP-aware browser agent. Status: active spec with browser implementation already shipping. Adoption: nascent (the spec is months old), but the distribution path is a default-on browser feature, not a crawler opt-in.
What each one is actually for
The three are not three takes on the same problem. They sit at different layers of the agent stack.
llms.txt is content discovery. "Here's what this site has." It doesn't say what an agent can do. It's the closest equivalent to a sitemap — a hint about what to read.
/.well-known/ai-plugin.json was action declaration via API. "Here's an API endpoint, here's the auth, here's what it returns." That model assumed the agent ran in a vendor's cloud (ChatGPT) and called your server directly. The platform that consumed it doesn't exist anymore in that form.
/.well-known/webmcp is action declaration via browser. "Here's what an agent running in the user's browser can do on this page." The agent already has the page loaded; the manifest tells it what's actionable. No API auth, no server-to-server contract — the agent acts as the user.
The mental model that makes this stop being confusing: llms.txt is for the crawler, ai-plugin.json was for the cloud agent, and webmcp is for the user's local agent. Each was the right answer to a different version of "where does the agent live."
Adoption reality in May 2026
| Manifest | Real adoption | Major consumers in production |
|---|---|---|
llms.txt |
<11% of public sites | None published; some indirect signals |
/.well-known/ai-plugin.json |
Tiny holdover from 2023 plugin era | Effectively zero |
/.well-known/webmcp |
Nascent (months old) | Chrome Canary; agent extensions in preview |
The story the table tells: llms.txt has had over a year of adoption opportunity and most sites still don't ship it because the payoff is unclear. ai-plugin.json's platform died. WebMCP just shipped in a major browser, with cross-vendor governance, and is the only one of the three with a credible path to default-on consumer reach.
Which one to ship if you ship one
If you want to spend an hour adding agent-readable surface to your site this month, the order is:
-
WebMCP first. The browser-native distribution path is the only one of the three with a clear forcing function (Chrome flips the flag, every Chromium browser inherits, and your action surface becomes immediately useful to any user running an agent). Even the minimal version (a
/.well-known/webmcpJSON discovery file plus declarative attributes on one form) puts you ahead of essentially every other SMB site as of writing. -
llms.txt second, only if you have time. The downside is small (one markdown file at the root) and the upside is a small amount of optionality if a major crawler suddenly starts honoring it. But the data after a year of opportunity says: don't expect anything from this. Ship it because it's cheap, not because it's a lever.
-
/.well-known/ai-plugin.jsonnot at all. The platform that consumed it doesn't exist. Don't add it. If you have a leftover from 2023, you can delete it (or leave it; nothing will read it either way).
Why the order matters more than the inventory
The natural temptation reading this list is to ship all three "just in case." That's the wrong instinct. The cost isn't the bytes — it's the maintenance debt and the false sense of having shipped something material.
WebMCP is a real action surface. When it works, an agent can submit your contact form on the user's behalf, navigate your site, run a search. That's a measurable user benefit and a measurable signal to the agent ecosystem that your site is participatory. llms.txt, even at peak adoption, is a hint to a crawler. ai-plugin.json is a relic.
Time and attention are the scarce resource. Spend them on the manifest that has a credible default-on consumer.
A minimal first ship for each
If you do decide to add WebMCP plus llms.txt this month — total effort under an hour — here's the actual file list.
/.well-known/webmcp at the site root:
{
"name": "Your Site",
"description": "What this site does, in one sentence",
"capabilities": []
}
/llms.txt at the site root:
# Your Site
> One-sentence summary of what your site does.
## Key pages
- [Homepage](/): what it covers
- [About](/about/): who you are
- [Services](/services/): what you offer
- [Contact](/contact/): how to reach you
Both files are static, both deploy as plain text or JSON, neither needs server logic, neither breaks anything for human visitors. Run WebMCP Readiness Checker to confirm the WebMCP file is detected, and LLMs.txt Audit to confirm the llms.txt is well-formed.
What's actually changing in 2026
The reason the WebMCP recommendation is different from "just ship all three" is that the agent ecosystem is consolidating around runtime surfaces, not crawler hints. The agent runtime as the new browser layer shift means agents are increasingly local to the user — running in the browser, in a desktop app, or in a CLI on the user's machine. Crawler-side declarations like llms.txt assume an architecture where someone else's bot fetches your file and decides whether to use it. Browser-side declarations like WebMCP assume an architecture where the user's own software is the agent.
The architectural bet is the second one. WebMCP shipping in Chrome Canary in February 2026 was the inflection point. Every additional browser that ships WebMCP-compatible support compounds the value of having declarations in place.
If you're going to ship one manifest in 2026, ship the one that's getting browser distribution.
Related reading
- WebMCP — The First Real Agent API For Your Site — full breakdown of the WebMCP protocol and four declarations
- WebMCP Readiness Checker — audit your site for the four WebMCP surfaces
- Markdown For Agents — What It Is And How To Fix It — the content-retrieval companion
- The Conversation Has Moved Past The Model — why runtime, not model, is the relevant layer
Fact-check notes and sources
- WebMCP Chrome Canary launch (February 10, 2026): developer.chrome.com/blog/webmcp-epp
- llms.txt original spec (Anthropic, 2024): llmstxt.org
- llms.txt 2026 adoption study (May 2026): aeo.press — State of llms.txt in 2026
- OpenAI plugin platform deprecation context (2024): widely reported; OpenAI shifted focus from plugins to GPTs and function calling, leaving
/.well-known/ai-plugin.jsonwithout an active consumer at the platform layer - W3C
.well-knownregistry pattern: defined by RFC 8615, the canonical mechanism for site-level capability discovery files
For the broader operating model behind shipping these declarations as a one-person site owner — including which audits to run, when to ship, and how to measure — The $20 Dollar Agency covers the lean approach to running an SMB site as if it were the agent's preferred interface.
This post is informational, not legal or implementation advice. Mentions of OpenAI, Anthropic, and Chrome are nominative fair use; no affiliation is implied. WebMCP is a W3C Community Group Draft as of writing — surface conventions may change before stable release.
