If you have run a PageSpeed Insights test recently, you may have noticed a new category next to Performance, Accessibility, Best Practices and SEO — Agentic Browsing.

Google added this category to Lighthouse 13.3 in May 2026 and PageSpeed Insights inherited it shortly after. It is currently marked experimental, but it signals a major shift in how websites will be consumed going forward — not just by humans, but by autonomous AI agents acting on their behalf.

This article explains what agentic browsing is, what Google's new audit checks, how it differs from traditional SEO, and what you need to do to prepare your website.


What is Agentic Browsing?

Agentic browsing is the paradigm where AI agents navigate, read and interact with websites autonomously on behalf of users. Instead of a person clicking through your pages, an AI agent — like Google's Project Mariner, OpenAI Operator, Perplexity or a custom AI agent — visits your site to complete tasks: comparing prices, booking services, filling out forms, extracting information and making decisions.

These agents do not "see" your website the way a human does. They read the page through the accessibility tree — the structured representation of your page's interactive elements, labels and relationships. They interact programmatically with buttons, links and forms, and they return results to the user who may never directly visit your site at all.

How Agentic Browsing Works

👤
User
Gives a task to the AI agent
🤖
AI Agent
Navigates and reads the site autonomously
🌐
Website
Accessibility tree, Schema.org, llms.txt
Result
Task completed, answer returned to user

This is fundamentally different from a search engine crawler that indexes your content for ranking. An AI agent is doing things — filling a contact form, comparing your pricing page to a competitor's, extracting your service offerings and presenting them as structured options to a user.


What Does the Agentic Browsing Audit Check?

Unlike the traditional 0–100 scoring model, the Agentic Browsing category uses a pass ratio — for example, 3/3. There are three default audits:

1. Accessibility Tree Validation

AI agents read your page through the accessibility tree, not visually. This audit checks whether:

  • All interactive elements (buttons, links, form fields) have programmatic labels
  • ARIA roles are valid with correct parent-child relationships
  • Elements are not improperly hidden via aria-hidden
  • Every clickable element has discernible text an agent can reference

A site can score 100 on traditional SEO and still fail this audit. A button like <button class="submit-btn"></button> (no text, no aria-label) is invisible to an agent. This is particularly important for AI chatbot widgets and interactive elements embedded in websites.

2. Cumulative Layout Shift (CLS)

Agents often take screenshots for visual analysis or calculate click targets programmatically. Layout shifts — even a few pixels — can cause an agent's simulated click to land on the wrong element. The target is CLS below 0.1, ideally zero.

Common culprits:

  • Images without explicit width and height attributes
  • Late-loading web fonts
  • Dynamic banner injection
  • Ad slots that load after initial render

3. llms.txt Validation

This checks for the presence and validity of an /llms.txt file at the domain root. Think of it as robots.txt for AI agents — a curated reading list in Markdown format that tells LLMs what your site is about and where to find key information.

A valid llms.txt must contain:

  • An H1 heading (your site/company name)
  • A blockquote summary of what you do
  • Links to your most important pages with descriptions
  • Substantive content that parses against the llmstxt.org reference parser

The Three Agentic Browsing Audits

🌳
Accessibility Tree
Labels on all interactive elements. Valid ARIA roles. Discernible button and link text.
📐
Layout Stability (CLS)
CLS below 0.1. Explicit image dimensions. No late-loading layout shifts.
📄
llms.txt
Valid Markdown at /llms.txt. H1 heading, blockquote summary, links to key pages.

How Is This Different from Traditional SEO?

Traditional SEO is about being found and ranked in search results. Agentic browsing readiness is about whether AI agents can actually use your site once they arrive.

Aspect Traditional SEO Agentic Browsing
AudienceSearch engine crawlersAutonomous AI agents
How it readsHTML parsing, link followingAccessibility tree, screenshots
InteractionRead-only crawlingActive — form filling, button clicking
Success metricRanking position, CTRTask completion, citation frequency
Discovery filesrobots.txt, sitemap.xmlllms.txt, agents.json
Layout toleranceCLS matters for UXCLS is critical — shifts break clicks

A site can score 100 on traditional SEO and 0/3 on Agentic Browsing. This is not a theoretical scenario — it is already happening. Buttons without discernible text, links without labels and images without dimensions all pass SEO checks but fail agent readability.


The Emerging Agentic Web Stack

Beyond the three default audits, Google and the broader ecosystem are developing additional standards. Here's what is coming:

agents.json

A /agents.json manifest at the domain root that declares available actions on your site — search, quote requests, booking, cart additions — with typed parameters. Think of it as an API contract between your website and AI agents.

WebMCP (Web Model Context Protocol)

An experimental protocol that lets websites register agent-callable tools directly in the browser. A site could declare "this form books a consultation" or "this search searches our knowledge base" using structured annotations, making it possible for agents to interact with your site without scraping. This is related to the Model Context Protocol that is already widely used in AI agent development.

agent-instructions.md

A /agent-instructions.md file providing behavioral guidance for agents — how to quote prices, when to defer to canonical sources, how to handle multi-step workflows. This is the natural extension of the prompt engineering that goes into building production AI agents.

The Agentic Web Stack — From Foundation to Frontier

Layer 1 — Foundation Already standard
robots.txt · sitemap.xml · Schema.org JSON-LD · Semantic HTML
Layer 2 — AI Readiness Lighthouse 13.3
Accessibility tree · CLS < 0.1 · llms.txt · FAQ schema
Layer 3 — Agent Integration Emerging
agents.json · agent-instructions.md · WebMCP · Auto-discovery links
Layer 4 — Agentic Commerce Future
AI-to-AI transactions · Autonomous procurement · Agent-native checkout

How Does This Relate to AEO and AIO?

If you are already thinking about Answer Engine Optimization (AEO) — making your content the preferred citation source for AI systems like Google AI Overviews, ChatGPT, Perplexity and Claude — then agentic browsing readiness is the technical infrastructure layer that supports it.

The relationship is hierarchical:

  1. Traditional SEO = foundation (crawlability, indexing, ranking)
  2. AEO / AIO = content strategy layer (being cited by AI, appearing in AI Overviews)
  3. Agentic Browsing = technical readiness layer (can AI agents actually use your site?)

Your content can be authoritative and well-structured — with proper FAQ schema, structured data and topical depth — but if your buttons lack labels, your layout shifts on load and you have no llms.txt, agents cannot mechanically interact with your site.

The convergence point: sites that are both AEO-optimized and agentic-browsing-ready will capture the growing segment of AI-mediated traffic — the searches that now result in zero clicks, where users get everything they need from LLM-powered responses.


Practical Checklist: Make Your Website Agentic-Browsing Ready

Here is what you should do today:

Must-have (to pass the three default audits)

  • Label every interactive element. Every <button>, <a> and <input> must have discernible text — either visible content, an aria-label or an aria-labelledby reference
  • Set explicit dimensions on images and embeds. Add width and height attributes to prevent layout shifts
  • Create /llms.txt. Write a valid Markdown file with an H1, a blockquote summary, and links to your key pages
  • Keep CLS below 0.1. Test with Lighthouse and fix any late-loading elements that shift content

Recommended (emerging standards)

  • Add rich Schema.org JSON-LD on every page — Service, FAQPage, Article, Organization, LocalBusiness
  • Create /agents.json declaring what actions are available on your site
  • Write /agent-instructions.md with behavioral guidance for AI agents visiting your site
  • Ensure your sitemap.xml is referenced in robots.txt and includes <lastmod> dates
  • Allow AI crawlers in robots.txtGPTBot, ClaudeBot, PerplexityBot, Google-Agent

Content strategy (for AEO + agentic readiness)

  • Structure content with clear Q&A patterns that AI can extract — FAQ sections with proper markup
  • Place proof near claims — statistics, certifications, case studies adjacent to the claims they support
  • Write concise first-paragraph answers to the main query each page targets
  • Build topical authority with interlinked service pages, case studies and blog content

Why This Matters for Your Business

The agentic browsing audit is not a ranking factor today — it is purely diagnostic. But it signals the direction Google is heading. As AI agents become the primary way users interact with the web — booking services, comparing vendors, filling forms — the sites that agents can reliably use will capture the traffic.

Sites that fail agentic browsing readiness will increasingly become invisible to this new class of AI-mediated visitors. Just as mobile-unfriendly sites lost traffic after Google's mobile-first indexing, agent-unfriendly sites will lose traffic as agentic browsing becomes mainstream.

Whether you are building AI chatbots for your customers, implementing RAG applications for your team, or looking at enterprise AI integration — the agentic web is the environment your AI products will operate in. Making your own website ready for it is the logical first step.


Summary

Google's new Agentic Browsing category in PageSpeed Insights checks three things: your accessibility tree is clean, your layout is stable, and you have an llms.txt file. These are the minimum requirements for AI agents to navigate your site. The broader agentic web stack — agents.json, WebMCP, agent-instructions.md — is still emerging, but the direction is clear.

The websites that will thrive in the agentic era are the ones that are simultaneously findable (SEO), citable (AEO/AIO), and usable (agentic browsing). Start with the three default audits, build from there, and keep your structured data, content depth and accessibility fundamentals strong.

The AI agents are already visiting your site. The question is whether they can do anything useful when they get there.