Your Clients Can Now See a Score You Can’t Explain

For years, PageSpeed Insights was a shared language. Your client ran it, saw four numbers, and you explained them. Performance, Accessibility, Best Practices, SEO. You knew what moved each one and roughly what it cost to move.

In May, Google added a fifth category to the default Lighthouse configuration, and PageSpeed Insights inherited it within a fortnight. Your clients can now see an Agentic Browsing section on the same report they have been running for a decade.

It doesn’t produce a number. It produces a fraction. It says “experimental” on it. And it tests four things, only two of which you can currently do anything about.

You are going to get asked about this. This is what to say.

What actually shipped, and when

Lighthouse 13.3 added the Agentic Browsing category to the default configuration in May 2026. Testing it requires Chrome 150 or later. PageSpeed Insights picked it up shortly after, which is why clients who have never opened DevTools in their lives are now seeing it.

Google’s own documentation describes the category and the WebMCP support underneath it as experimental and based on proposed standards. That is not a hedge you should skip past when you’re explaining it — it’s the most useful thing in the sentence.

Why it reports a ratio rather than a score. Every other Lighthouse category gives you a weighted average out of 100. This one gives you a pass ratio — how many agentic readiness checks your site passes out of how many were run. Google’s stated reason is that the standards for the agentic web are still emerging, so the current aim is to gather data and give actionable signals rather than rank anyone.

That is the sentence to give a client who wants a number. There isn’t one, deliberately, and the absence is a feature. A site scoring 2/4 is not “50% agent-ready” in any meaningful sense, and you should resist letting anyone in the room describe it that way.

The four things it tests, and what you can actually do about each

WebMCP — real standard, not yet a deliverable

Lighthouse checks whether your site registers WebMCP tools, whether your forms are missing declarative WebMCP, and whether your WebMCP schemas are valid. It monitors registration through the Chrome DevTools Protocol, covering both declarative tools defined in HTML and imperative ones defined in JavaScript.

WebMCP lets a site hand an agent a structured list of what it can do and the exact parameters each action needs, rather than making the agent screenshot the page and guess where to click. It is a genuinely good idea. We wrote an explainer for site owners: WebMCP explained, for people who own websites.

Where it stands: published as a W3C Draft Community Group Report in February 2026, currently in a public origin trial in Chrome. Firefox, Safari and Edge are involved in the working group discussions and have shipped nothing. Testing the WebMCP audits in Lighthouse requires registering for the origin trial.

What to tell a client: this is the part of the score that is a standards bet, not a task. Implementing it today means writing against a draft spec, for one browser, behind an origin trial. There are businesses for which that is a reasonable investment. Most of your clients are not among them.

What not to do: quote for “WebMCP implementation” as a line item on a small business website. If it moves, you will be rewriting it, and the client will remember who sold it to them.

llms.txt — cheap, and near-worthless

Lighthouse checks for a machine-readable summary file at your domain root.

We have been consistent on this and we’re not going to become less so: independent analysis of over 137,000 sites found that 97% of llms.txt files are never requested by an AI system at all, and having one shows no correlation with being cited more often. We covered the evidence in what the data actually says.

What to tell a client: add one, because it takes twenty minutes and Lighthouse looks for it. Do not build anything on it, and be wary of anyone selling llms.txt strategy as a service.

This is a useful trust-building moment in a client conversation, incidentally. Telling someone which part of a new thing is hollow buys you credibility for the parts you say are real.

The accessibility tree — real, implementable today, and you already know how

This is the substantial one, and it is the reason to take the category seriously.

Google’s documentation is explicit that agents rely on the accessibility tree as their primary data model. Lighthouse filters a specific subset of accessibility audits that matter for machine interaction: every interactive element having a programmatic name, roles and parent-child relationships being valid, and content not being hidden from the tree while remaining interactive.

Which is to say: the fashionable new category is largely scoring accessibility work that has been on your clients’ to-do lists for years, and which they already owe under the Equality Act 2010 regardless of what any browser vendor does.

What to tell a client: this is billable, deliverable, standards-stable work that improves two scores and reduces a legal exposure at the same time. If they want to spend money on agent readiness this year, spend it here.

Layout stability — CLS, wearing a different hat

Cumulative Layout Shift appears in the agentic category because an agent identifies an element, then acts on it, and anything that moves in between causes a failed task. A human recovers from a shifting button. An agent clicks the wrong thing.

Google also flags CLS as one of the reasons agentic results fluctuate between runs, along with the timing of dynamic tool registration and variability in how the accessibility tree gets built. Worth knowing before a client shows you two different results and asks why.

What to tell a client: nothing new. You already sell this. It now has a second justification.

The two answers that will make you look good

“Why is there no number?” Because Google chose not to publish one while the standards are still moving, and a pass ratio is an honest reflection of that. Anyone giving you a percentage is making it up.

“Should we be worried?” Two of the four checks are things we should be doing anyway and can start on this quarter. One is a twenty-minute job with little evidence behind it. One is a draft standard in a single browser that we will watch rather than buy. That is a proportionate answer, and it is also true.

One thing to check that isn’t on the report

If a client has asked you to make their site agent-ready, check their CDN before you invoice for it.

Cloudflare is changing its crawler defaults on 15 September 2026, and among the configurable categories is Agent — automation acting in real time on behalf of a waiting human. A meaningful number of site owners have that blocked without having decided to, usually via a legacy one-click “block AI bots” toggle. It is an expensive irony to spend a quarter on agent readiness and then discover the CDN was refusing the traffic all along.

We’ve written the checks up in full: on 15 September, your CDN decides who can read your site.

The thing worth noticing

Every few years the industry produces a new score, and the consultants who do best out of it are not the ones who sell hardest against it. They’re the ones who can say, calmly and specifically, which parts of it are real.

The Agentic Browsing category is a good example because it is honestly constructed. Google published it as experimental, refused to reduce it to a number, and built half of it out of accessibility checks that predate the entire conversation. The people who read it carefully will end up recommending semantic HTML, proper labelling and stable layouts — which is what they would have recommended in 2015, and were probably ignored about.

The new thing rewards the old discipline. It usually does.

Where we fit

We’re a development partner for SEO consultants, designers and agencies. You keep the strategy and you keep the client relationship — we write the code, work to your brand, and stay current on specifications like WebMCP so you don’t have to read the working group minutes.

If a client has started asking about agentic scores and you’d rather have a developer in the room than a position on it, let’s have a partnership discussion — or schedule a call.

Code is invisible. Your reputation isn’t.

Sources

• Chrome for Developers, Lighthouse agentic browsing scoring
• Chrome for Developers, Accessibility for agents
• Chrome for Developers, llms.txt audit
• Cloudflare, Your site, your rules: new AI traffic options for all customers, 1 July 2026