Google’s New Agentic Browsing Score: What Every Website Owner Needs to Know in 2026

My website scored 0/3 on Google's newest Lighthouse test. Here's what the Agentic Browsing score actually checks, and exactly how I fixed it.

Google’s Lighthouse — the tool behind PageSpeed Insights that every SEO professional relies on — now scores your website on something entirely new: how well AI agents can browse and interact with it.

A new Agentic Browsing category has quietly appeared alongside the familiar Performance, Accessibility, Best Practices, and SEO categories. Unlike those, this one isn’t about human visitors at all. It’s about whether AI agents — the kind that browse, click, and fill out forms on a user’s behalf — can effectively navigate, understand, and interact with your website.

I discovered this the hard way when I ran a PageSpeed audit on sadhyam.com and saw a score of 0/3. Here’s everything I learned fixing it.

What the Agentic Browsing Audit Checks

The Agentic Browsing category uses a pass/fail ratio — for example, 2/3 means two out of three audit groups passed. Here are the three groups it checks.

1. Discoverability — llms.txt

The audit checks for an llms.txt file at your domain root. Think of it as robots.txt for AI — a machine-readable summary that tells AI systems who you are, what your site covers, and where to find your key content.

A good llms.txt includes:

• Site name and description
• Author information (name, role, expertise, social profiles)
• Key pages (home, about, contact, blog)
• Topics covered, with article counts
• Recent articles with URLs and dates

2. Agent-Centric Accessibility

AI agents navigate your website using the accessibility tree — the same structure screen readers rely on. Lighthouse checks three things here:

• Names and labels — every interactive element must have a programmatic name, typically via aria-label
• Tree integrity — roles and parent-child relationships must be valid
• Visibility — elements must not be hidden from the accessibility tree while remaining interactive

The most common failure: “ARIA hidden element must not be focusable or contain focusable elements.” This happens when a container uses aria-hidden="true" while it still contains clickable links inside it — a very common pattern in mobile navigation menus.

The fix is simple: use CSS display:none instead of aria-hidden. When an element has display:none, it’s automatically removed from the accessibility tree, so there’s no conflict for an agent — or a screen reader — to trip over.

3. Layout Stability (CLS)

For AI agents, layout shifts are catastrophic. An agent that identifies a button’s coordinates and then the layout shifts beneath it will miss the target entirely — there’s no human eye to notice and re-click. Common CLS culprits include:

• Images without explicit width and height attributes
• Fonts loading late and causing text reflow (fixed with font-display: swap)
• Dynamic content injection that pushes elements around after the initial render

My Experience: Taking sadhyam.com From 0/3 to Passing

Fix 1: Created llms.txt

I added a dynamic llms.txt endpoint to the WordPress theme that automatically generates a comprehensive site summary — author information, every published article with its URL, topic categories, and technical metadata. I also added a corresponding link tag in the HTML head for discoverability.

Fix 2: Fixed the Accessibility Tree

The failing element was the mobile navigation, which used aria-hidden="true" while still containing focusable links. I removed aria-hidden entirely and switched to CSS display:none to hide the mobile nav by default, toggling to display:block when the menu button is clicked.

Fix 3: Eliminated Layout Shifts

I added explicit width and height to every image, set a min-height on the header, ticker, and nav, preloaded critical fonts with font-display: swap, and added aspect-ratio CSS to every image container.

What About WebMCP?

You may notice up to three “Not Applicable” audits related to WebMCP integration. These are experimental audits that require Chrome 150+ and an origin trial registration. For most websites today, these aren’t required to pass — focus on the three core audits: llms.txt, accessibility, and layout stability. We cover WebMCP itself, and how to prepare your site for it, in a dedicated guide.

Why This Matters for Your Business

The agentic web isn’t a future prediction — it’s happening now. AI agents are already shopping, comparing, and transacting on behalf of users. Google adding an Agentic Browsing audit to Lighthouse is a clear signal: optimizing for AI agents is becoming as important as optimizing for human visitors, the same way Bing’s new AI visibility reports show search engines are already tracking how often AI systems cite your content.

Action Checklist

1. Run your site through PageSpeed Insights and check the Agentic Browsing section
2. Create an llms.txt file at your domain root
3. Audit your HTML for aria-hidden="true" on elements containing focusable children — replace with CSS display:none
4. Ensure all buttons, links, and inputs have accessible names
5. Add explicit width and height to all images
6. Use font-display: swap for web fonts
7. Re-run the audit to verify you pass

Written by mootkur9

17 years in digital marketing. Founder of Sadhyam. Writing about AI search, AEO/GEO, and the agentic web. More about me →