WebMCP Has a Chrome Origin Trial. llms.txt Has a Study.

Two answers to the same question landed within a week of each other this year, pointing in opposite directions. On 9 June 2026 Chrome opened an origin trial for WebMCP in Chrome 149, giving a page a way to hand an agent a defined set of actions. Six days later a log study put a number on how often anyone actually fetches an llms.txt file. The question underneath both is how a site makes itself legible to software that arrives to do something rather than to read.

What a page can declare through WebMCP

The Chrome announcement, published 9 June 2026 under the title “Join the WebMCP origin trial”, frames the API as a way to “provide rules for interaction between web applications and agents”. The problem it targets is the one every agentic browsing demo hits: an agent arriving at a page infers what each control does from the DOM, and infers wrong. Its two worked examples are filling complex structured forms where field requirements differ, and exposing developer tooling a human interface would hide.

The structural difference from scraping is the point. Under WebMCP the page declares what it can do and the agent calls it; the agent is not guessing at a button. The explainer sits in the Web Machine Learning Community Group repository on GitHub, which is where the specification work is happening rather than in any shipped standard. It is the same problem covered in what happens when AI clicks instead of you, from the site’s side of the transaction instead of the agent’s.

A browser API with an implementation against a convention with none

An origin trial is a real, if provisional, thing. Chromium has built the feature, registration is open, and sites can run it on live traffic behind a token. That is a different category of existence from a naming convention anyone can adopt and nobody is obliged to consume. The two have been discussed for a year as competing options, when only one has a consumer.

The evidence against llms.txt, stated once

Google’s AI features optimisation guide, last updated 10 July 2026, is unambiguous. The page states that you do not need AI text files to appear in Google Search: “You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn’t use them.” It goes further on the question of harm, saying that publishing them “will neither harm nor help your site’s visibility or rankings in Google Search, as Google Search ignores them.”

Ahrefs published a server-log study on 15 June 2026 covering 137,210 domains using its Web Analytics product, with the traffic data drawn from May 2026. Three figures matter:

• 28% of those domains publish an llms.txt file.
• 97% of those files received zero requests in May 2026. The authors’ phrasing is “Nothing fetched them at all.”
• The study’s own ceiling: Ahrefs notes its customers “skew more technical and SEO-aware than the web at large”, and that a fetch is not the same as consumption, so “every figure in this study is therefore a ceiling on actual llms.txt consumption”.

That last caveat is the one worth carrying. The 3% that were fetched at all cannot be assumed to have been read, parsed or acted on. The mechanics of the file itself, and why the idea was reasonable when proposed, are covered in what llms.txt actually does.

Shopify serves the file already, and that is not evidence

The counter-argument usually points at Shopify. It is true: a changelog entry dated 28 May 2026 describes a Shopify-managed agents.md with customisable templates, and states that “the paths /llms.txt and /llms-full.txt also point to this content by default”. Merchants override each path with templates/agents.md.liquid, templates/llms.txt.liquid or templates/llms-full.txt.liquid, falling back to the agents.md template and then to Shopify’s default.

That is a platform deciding to serve a file cheaply on every store. It is not evidence that anything consumes it. The default costs Shopify almost nothing and settles an argument with merchants; it says nothing about the demand side.

What an origin trial does not mean

WebMCP is not shipped. An origin trial is time-limited, registration-gated and Chromium-only, and features have been withdrawn from trials before. Shopify’s WebMCP support for Liquid and Hydrogen storefronts, dated 5 August 2026 in its changelog, runs under that same trial rather than on a stable API. Treating it as a build requirement for the next quarter would repeat the mistake llms.txt advocacy made, with a better underlying idea.

The defensible position is narrow. Watch the Chrome Status entry and the community group repository, register for the trial only if you run a storefront with complex forms and have engineering time for a throwaway integration, and otherwise keep the HTML semantic, the forms labelled and the important actions reachable without JavaScript. That set is what an agent, a crawler and an assistive technology can all use today. It is also the part of whether generative engine optimization is real that survives contact with primary sources.