Lighthouse now measures if your site is ready for AI agents. What it means and why not to panic

Imagine an AI agent that a customer asks: "find winter tyres 205/55 R16 in shop X and add them to the cart". The agent opens the site, takes a screenshot, tries to guess where the search box is, clicks. Sometimes it hits. Sometimes it clicks an ad, because the layout shifted half a second earlier. This is what most of these attempts look like today. Clumsy and blind.

Google decided to measure it. In May it quietly added a new tab called "Agentic Browsing" to Lighthouse, its site audit tool. It checks one thing: how ready your site is to be used not by a human, but by a program. And before someone tells you that you now have to rebuild everything because "Google says so", read to the end. Because Google contradicts itself on this.

What actually happened

The new category showed up in Lighthouse 13.3, released on 7 May 2026. It works from Chrome version 150 and up, with no extra setup - it is simply there in DevTools. Status: experimental, based on standards that are only now being written. That is important, I will come back to it.

The audit looks at four things: WebMCP, the accessibility tree, the llms.txt file and layout stability. It sounds technical. Let me break it down in plain words.

The four things it checks

The accessibility tree is the invisible skeleton of a page that describes what each element is. This field is the "Send" button. This is a heading. This is the link to the cart. People do not see it, but screen readers for blind users have relied on it for years. It turns out that AI agents use the same skeleton as their main source of information about the page - it is more convenient for them than raw code or a screenshot. So the same thing you do for accessibility helps machines. We showed how to check this on your own site in a separate post about website accessibility testing.

Layout stability is the familiar CLS metric, that is, whether the page does not "jump" while loading. You know that feeling when you want to click a button and at the last moment a banner pops in and you click something completely different. An agent taking screenshots gets lost exactly the same way you do. Google has measured CLS since 2020, we described it when writing about PageSpeed Insights. Here it just got a new use.

llms.txt is a text file in the root folder of the site, something like a map for AI models. It tells them where the most important content is, so they do not have to dig through the whole site. Lighthouse checks whether you have it.

WebMCP is the newest and least mature part. It is a way for the site to tell the agent itself: "here is a ready tool called search_product, give it a keyword and I will return results". Instead of guessing from an image, the agent gets a clear instruction. In code it looks roughly like this:

navigator.modelContext.registerTool({
  name: "addTodo",
  description: "Adds a task to the list. Use when the user asks to remember a task.",
  inputSchema: {
    type: "object",
    properties: { text: { type: "string" } },
    required: ["text"]
  },
  execute: ({ text }) => { todoApp.addItem({ text }); }
});

WebMCP was announced at Google I/O 2026, and Google and Microsoft develop it together. It sounds promising, but this is where it gets harder. For Lighthouse to test WebMCP at all, you have to turn on a so-called origin trial, that is, sign your site up for a closed trial. Without it, this part of the audit simply does not run. WebMCP also needs an open browser tab and, on bigger sites, a fair amount of code rework. This is not something you turn on with one click.

Here starts the part you need to understand

First, about the score itself. This category does not give a 0 to 100 score like the rest of Lighthouse. It gives a fraction: how many tests passed out of how many possible, plus a pass or fail status on each. Google explains it plainly: the standards for an agent-driven web are only being built, so for now it is about collecting data and giving hints, not about handing out a final grade.

In practice this means one thing: you cannot "fail" this audit. A normal, simple site gets a passing result. If you do not have an llms.txt file, the audit shows "not applicable", not an error - because the file is optional for now. Nobody will give you a red cross for not implementing an experimental AI feature.

And now the crux. Google adds a check for the llms.txt file to Lighthouse. And at the same time, in its own guide on optimising for AI search, in the myth-busting section, it writes: you do not need to create new machine-readable files, AI files or markup to appear in generative search. So one Google tool checks for something that another Google document says is unnecessary.

John Mueller from Google clears it up like this: llms.txt is not built for search. It is useful to AI systems for code and documentation, it saves tokens, but for a normal company site he calls it a "temporary crutch" and says plainly that he does not see much sense in it. It is a tool mainly for developer sites.

From this comes a conclusion I want you to remember. Lighthouse is not Google ranking. Passing the Agentic Browsing category will not move you up in search results by a single place. It is a thermometer for agent readiness, not a ranking factor. So if someone now sells you a package "we will prepare your site for AI, because Google says so" and scares you with a drop in positions, you can calmly put your wallet away.

So what is actually worth doing

If you should not run off and generate llms.txt in a panic, then what? Boring fundamentals. The same ones that help people, search engines and agents at once, no matter which trend comes next.

A clean accessibility tree. Every button, link and field described so it is clear what it is. The same thing you do for people using screen readers, which a machine reads as the site's instruction manual. If accessibility is a new topic for you, start with the post on what WCAG actually is.

A stable layout. Low CLS, that is, a page that does not jump while loading. Good for the customer's nerves, good for the PageSpeed score, good for an agent taking screenshots.

Structured data. Schema.org in JSON-LD format, that is, the invisible piece of code that tells the machine plainly: this is a company, it is called this, this is an article, this person wrote it. We wrote about this in more detail when covering the website as the single source of truth about your brand. Structured data was important before anyone said the word "agent", and it will be important when this trend passes.

See the pattern? None of these things is new. None of them needs an origin trial, experimental APIs or a prayer that the standard catches on. It is simply solid work that pays off anyway.

WebMCP and llms.txt are worth watching. If you run a shop or an app where customers really will let agents perform tasks, in a year or two it may matter. But today it is a construction site, not a finished product. Invest in the fundamentals, track the experiments.

For us this is standard, not news

dede.agency has an llms.txt file, but that is the least important piece of the puzzle. The real work is elsewhere. We implement structured data and a proper accessibility tree on every project as part of the craft, not as a paid add-on. Every page, offer and post gets its schema.org markup and a structure machines can read. We did this before Lighthouse added a tab for agents, and we will do it when the next standard arrives.

That is the difference between keeping up with a trend and staying ahead of it. When the foundation is solid, you meet new requirements before they even become loud. When it is missing, every new fad means catching up.

Want to know how your site does in the new Lighthouse category and whether the fundamentals are in place? Write to us, we will take a look and tell you plainly what is worth fixing and what you can safely leave alone.

Sources

• Lighthouse Agentic Browsing - scoring (Chrome for Developers)
• llms.txt audit (Chrome for Developers)
• WebMCP (Chrome for Developers)
• Technical analysis of the category (DebugBear)
• Google's contradiction on llms.txt (Search Engine Land)
• Google's inconsistent guidance (Search Engine Journal)
• Google adds a category to Lighthouse (Semrush)
• WebMCP as a web standard (InfoQ)