Google Pagespeed Now Has a Fifth Metric
For years, we’ve had four main metrics to review when running a site through https://pagespeed.web.dev:
- Performance
- Accessibility
- Best Practices
- SEO
We now we have a 5th one with Chrome 149:
- Agentic Browsing
What is Agentic Browsing?
Agentic Browsing refers to an AI’s ability to autonomously navigate the internet and interact with websites to complete multi-step tasks on behalf of a user.
Instead of just searching for information and summarizing it, an AI capable of agentic browsing acts as a “digital proxy.” It can think, click, type, fill out forms, and make decisions to execute complex workflows.
Key Characteristics
- Action-Oriented: It moves beyond reading text to taking physical actions on a webpage (e.g., adding items to a cart, scheduling an appointment).
- Multi-Step Autonomy: You give it a high-level goal (e.g., “Find and book the best hotel in Chicago under $200 for next weekend”), and it manages the entire research and execution process across multiple websites.
- Dynamic Problem Solving: It can handle authentications, navigate dropdown menus, and adapt if a webpage layout changes.
In short: Traditional browsing is you using a tool to look at the web; agentic browsing is the AI using the web as a tool to get things done for you.
TLDR
- Defining Agentic Browsing (above)
- The new pagespeed.web.dev ‘Agentic score’ in Chrome 149
- Existing and new scoring aspects of the New & Not Applicable (yet): a. WebMCP tools (experimental, coming up as Unscored) b. llms.txt (a Table of Contents for the Agents, also Unscored)
- Why WebMCP Tools deployed via GTM probably won’t appear in Pagespeed, but can still check & confirm
OK now that’s out of the way, let’s go to a real site: billshultz.com
This new Agentic Browsing tool was shown at Google I/O back in May. Now it seems to be generally available.
Here’s the process I followed:
- https://pagespeed.web.dev
- Add ‘billshultz.com’ & run it
- Scores come up
Then click on that fifth option: Agentic Browsing
This is of course still in Experimentation Mode
And also my own opinions:
Scroll down to see the screenshot — this is valid, the browsers will see it; but Pagespeed might not
Screenshots:
- Results, I clicked the ‘DESKTOP’ tab:

Passed Audits
- See ‘Agentic Browsing (2/2)’:

- Accessibility tree is well-formed
- CLS: Cumulative Layout Shift = 0
CLS measures how much visible content on your website moves around unexpectedly while the page is loading. The CLS score has been a challenge for many, many websites - but remember, Agents don’t work like people. They are just looking for crisp, quick, accurate, ‘things to execute’ & get done for users.
Plus both of these audit points are existing, long standing technical SEO metrics/gauges etc..
Here’s where the fun stuff begins: ‘NOT APPLICABLE’
*** This translates to ‘Experimental’***
There are two different topics in this section:
- WebMCP Tools
- llms.txt file

- WebMCP form coverage: This is the KEY TOOL that I think all the developers are working for — to set WebMCP Tools up for FORMS, so Agents can do the work for you and submit your information upon your request.
- WebMCP tools registered: The WebMCP tools that are registered — and I’m betting for that page only. WebMCP is browser tab based & well guess what, so is pagespeed.web.dev. That’s why they call it pagespeed, not sitespeed.
- WebMCP schemas are valid: This is the one I’m most interested in for SEO. This might end up being a score about how well your WebMCP Tools match up to your SCHEMA. It may end up being that if you DON’T have well formed schema markup on the page, then the tool will tell you something like this: “Your schema and page structure/IA doesn’t match up with what the WebMCP Tool says the content is supposed to do etc..
- llms.txt: Another ‘oldie but goodie, ***all the way back to September 2004 by Jeremy Howard - ancient. I think this will do precisely what it was designed to do, become a table of contents for LLM’s in markdown format.
Since I don’t have a llms.txt file on this site, it didn’t see it. I’ll build one and try it again…
OK - PASSED
For this file: https://billshultz.com/llms.txt:

Interesting Update: June 30 - July 1
Poking around, I see that these WebMCP Tools may NEVER show up in pagespeed.web.dev IF these are pushed out via GTM.
Plus this is still experimental. Check this one:
- https://googlechromelabs.github.io/webmcp-tools/demos/french-bistro/
- This is one of the DEMO SITES used at Google I/O back in May. Run pagespeed on it & nope, still showing as “Unscored”.

So maybe it WILL work eventually, but who cares — we can still see it in that nice little add-on in Chrome Inspect.
Two things:
-
This site is built in ASTRO, which has more complex stuff going on & doesn’t seem to like GTM very much. Will try to fix that + probably try on another site.
-
It is how pagespeed.web.dev works
From Gemini:
That is incredibly interesting! If you have already gone ahead and implemented WebMCP tools on billshultz.com — either via the Declarative API on your HTML forms or imperatively using navigator.modelContext.registerTool—there are three very specific engineering reasons why the PageSpeed Insights engine marked them as “Not Applicable” during that scan.
- The CSP Hangover
If your WebMCP tools were being registered inside an inline script block — like script is:inline, the PageSpeed Insights browser runner would have hit that exact same firewall violation. The browser blocked the script from running, meaning navigator modelContext registerTool was never executed, and the page reported 0 registered tools to the auditor.
I think Gemini is wrong here, because I can still see the WebMCP Tools in Chrome Inspect
- The Headless vs. Live Tab Isolation (I think it is more of these two instead)
WebMCP is a brand-new browser standard designed for live, user-driven browser tabs. The security model dictates that these tools are strictly ephemeral—they exist to let an AI agent help a real user inside an authenticated session.
PageSpeed Insights uses automated, headless Chromium instances running on remote Google servers. Because headless environments don’t represent a true “live user tab session,” the automated runner sometimes fails to expose or instantiate the experimental navigator.modelContext global API unless the testing suite explicitly forces the runtime flag (—enable-features=WebMCP).
- Site Execution Timing
Because your site is running View Transitions, the timing of when your script registers the tools matters. If your tool registration script runs immediately on a script load but Lighthouse captures the DOM state a fraction of a second too early (or expects them to be bound to a global lifestyle hook like window.onload), it can miss them.
- Still in Experimental Mode
Gemini didn’t type this one. This might be it, considering the demo shown at Google I/O doesn’t show the WebMCP tools as active either LOL. Ok back to Gemini…
How to verify your WebMCP tools right now
Don’t rely on the automated PageSpeed server to test your tools. Since you are using Chrome, you can audit them locally in real-time:
- Open your site on your local dev server or live production tab.
- Open Chrome DevTools (F12) and navigate to the Application tab.
- Look at the left sidebar menu—under the Storage/Background services sections, Chrome has a dedicated WebMCP panel.
- Click it, and you will see a clean, structured printout of every single tool, schema, and action definition currently active in your tab!
AOK, thanks Gemini - understood!
Summary on confirming WebMCP Tools are active
Basically, Pagespeed is headless & will ping faster than the tag loads. Still, can check Inspect and verify.
That Setup Context Tag is still the one IMO. It is a shortcut to the llms.txt - highest priority corrections for hallucinations.
Bill Shultz is an enterprise digital marketing consultant with 20+ years of experience. He operates FixClicks LLC and writes about paid search, analytics, SEO, and the emerging discipline of agent optimization at billshultz.com.