The Fifth Score Nobody Expected
If you’ve opened Google PageSpeed Insights recently, you might’ve noticed something new sitting next to the familiar Performance, Accessibility, Best Practices, and SEO scores. It’s called Agentic Browsing and unlike the others, it doesn’t show a number out of 100. It shows a ratio: 2/3, 3/3, or similar.
This wasn’t a big announcement. Google quietly shipped it in Lighthouse 13.3.0 on May 7, 2026, and PageSpeed Insights inherited it shortly after. No blog post. No countdown. It just appeared and now every site audit includes it.
What Is Agentic Browsing?
Agentic Browsing measures how well your website works for AI agents not human visitors.
In 2026, it’s not just people navigating websites. AI systems like Perplexity, SearchGPT, and smart assistants are increasingly browsing the web autonomously: reading content, clicking buttons, filling forms, and making decisions on behalf of users. This is called “agentic browsing.”
The problem? Most websites weren’t built with machines in mind. Buttons with no accessible names. Broken ARIA roles. Layout that shifts as the page loads. These things confuse AI agents the same way they confuse screen readers and now Google is measuring it.
The Agentic Browsing category is currently marked “under development,” which means it’s not a hard ranking factor yet. But Google introduced it as a scoreboard, not a warning. The direction is clear.
What Does Agentic Browsing Actually Check?
The audit currently has up to 6 checks, depending on your site setup. The three that apply to every site:
1. Accessibility Tree Integrity
This is the most important one. An AI agent doesn’t “see” your page visually it reads the accessibility tree, the same semantic structure that screen readers use. Headings, buttons, links, form fields all derived from your HTML and ARIA attributes.
If that tree has errors invalid roles, missing labels, broken parent-child relationships the agent can’t navigate the page reliably. Google’s own documentation calls the accessibility tree the agent’s “primary data model.”
Common issues that break it:
- Invalid or inappropriate ARIA role values
- Interactive elements with no accessible name
- Elements that are visually present but hidden from the tree
- Nested roles that violate ARIA spec
2. Cumulative Layout Shift (CLS)
You already know CLS from Core Web Vitals. It measures how much your layout jumps around as the page loads.
For AI agents, this is particularly disruptive. An agent may try to click something the moment it appears before the layout stabilizes. If the element shifts, the action fails or hits the wrong target. A CLS score under 0.1 is what you’re aiming for.
3. llms.txt File
This is a newer concept. llms.txt is a plain-text file placed at your domain root (like robots.txt) that gives AI crawlers structured information about your site: what it does, what pages exist, what’s off-limits.
Lighthouse checks whether the file exists and whether it’s correctly formatted has an H1 header, sufficient content, and working links. If your site doesn’t have one, this check fails.
Why You Should Fix It Now (Even If It’s “Experimental”)
A few practical reasons:
AI traffic is already real. Perplexity, Claude, ChatGPT, Gemini they’re all crawling and reading sites. Whether they can understand and interact with yours affects how they represent your content to users.
Accessibility fixes help everyone. The same issues that trip up AI agents also hurt screen reader users, keyboard-only users, and search engine crawlers. Fixing them is never wasted effort.
Google doesn’t add categories to Lighthouse by accident. CLS started as an experiment in 2020. It became a Core Web Vital. The pattern repeats.
Your competitors aren’t paying attention yet. Most site owners haven’t noticed this audit exists. Getting your site to 3/3 while the space is quiet is straightforward.
Real-World Fix: Elementor and Invalid ARIA Roles
Here’s a concrete case from a recent project a WordPress site built with Elementor.
PageSpeed Insights flagged an invalid ARIA role error on the homepage. Elementor was outputting something like this in the HTML:
<article class="elementor-post" role="article">...</article>
<div class="elementor-posts-container" role="list">...</div>
The problem: the role attribute values being applied by Elementor in this context were either redundant (an element already has an implicit article role) or incorrect for the element type both of which pollute the accessibility tree and trigger audit failures.
The fix was straightforward. Since modifying Elementor’s core output directly isn’t practical, a small snippet in functions.php removes the offending attributes on page load:
add_action('wp_footer', function() {
if (!is_front_page()) return;
?>
<script>
document.querySelectorAll('article.elementor-post').forEach(function(el) {
el.removeAttribute('role');
});
document.querySelectorAll('div.elementor-posts-container').forEach(function(el) {
el.removeAttribute('role');
});
</script>
<?php
});
This targets only the front page (where the posts grid lives) and strips the attributes before the browser finishes painting so Lighthouse sees a clean accessibility tree.
Result: The invalid role audit cleared. The Accessibility score improved. The Agentic Browsing check for accessibility tree integrity passed.
One thing worth noting: this type of issue is extremely common in page builder-based sites. Elementor, Divi, WPBakery they all generate ARIA attributes in ways that made sense years ago but now conflict with current accessibility specs. If your PSI Accessibility score is stuck below 90, invalid roles are often part of the reason.
Quick Checklist: Getting to 3/3 on Agentic Browsing
- Run PageSpeed Insights on your key pages and check the Agentic Browsing section
- Fix accessibility tree errors invalid roles, missing labels, broken ARIA nesting
- Check your CLS score anything above 0.1 needs attention; lazy-loaded images and late-injected ads are common culprits
- Add an llms.txt file to your domain root even a basic one will pass the check
- Don’t use redundant ARIA roles on semantic HTML elements (e.g., role=”article” on an tag)
Need Help With PageSpeed Insights?
If you’re running into issues like these invalid ARIA roles, broken accessibility tree, low Agentic Browsing scores I can help diagnose and fix them.
I work with Performance, Accessibility, Best Practices and SEO issues in Google PageSpeed Insights across all types of WordPress sites, including Elementor, WooCommerce and custom themes. Whether it’s a single technical SEO audit or a full audit pass to push scores into the 90s, feel free to reach out.




Leave a Reply