Agentic browsing: what Google recommends for AI agents
Chrome is testing a new Lighthouse category dedicated to agent-driven browsing. It no longer measures whether an engine cites your brand, but whether an agent can actually use your site to act on someone’s behalf.
Until now, the question behind GEO was: is my site picked up, cited or recommended in an AI-generated answer? With agents now able to browse, click and fill in forms on behalf of a user, a second question appears: once my site is recommended, can an agent actually use it?
This is exactly what Google is starting to measure. The Chrome Lighthouse team introduced an experimental category, agentic browsing, which evaluates how well a site can be understood and operated by a software agent rather than by a human in front of a screen.
Key takeaways
- Agentic browsing is an experimental Lighthouse category, separate from the classic Performance, Accessibility or SEO scores.
- The result is not a score out of 100: it is a ratio of passed audits across a set of deterministic checks, closer to a readiness status than a marketing score.
- Three pillars are evaluated: WebMCP integration, machine-centered accessibility, and visual stability paired with the presence of an llms.txt file.
- This evaluation does not measure citation in an answer, but whether an agent can actually take action on the site once it gets there.
What Lighthouse actually evaluates
Google describes this category as measuring "how well your site is constructed for machine interaction through a set of deterministic audits." Three families of audits make up this score.
- WebMCP integration: Lighthouse monitors, via Chrome's DevTools Protocol, whether the site registers tools an agent can use, whether declared in HTML (declarative) or in JavaScript (imperative), and validates their schema.
- Agent-centric accessibility: the page's accessibility tree is checked on three specific points, programmatic names and labels for every interactive element, the integrity of ARIA roles and parent-child relationships, and whether content is actually visible in that tree.
- Stability and discoverability: Cumulative Layout Shift (CLS) is measured, since an agent that clicks relies on element positions, and Lighthouse also checks for an llms.txt file at the domain root, a machine-readable summary.
Why results can vary from one run to the next
Google notes that this score can fluctuate between runs, for three technical reasons. Dynamic registration of WebMCP tools depends on exactly when the JavaScript executes, which can cause a tool to be missed by the audit. DOM complexity affects how the accessibility tree is built. Finally, elements injected after the fact, ads or third-party content, can move interactive elements between the moment an agent spots them and the moment it tries to interact with them.
Google is upfront about this variability: there is no definitive ranking or reference threshold yet. This is a work in progress, not a fixed standard.
What Google recommends to technical teams
The published recommendations stay deliberately concrete and actionable.
- Adopt WebMCP to explicitly expose the site's logic and forms to AI agents, instead of leaving an agent to guess how to trigger an action.
- Use proper semantic HTML and ARIA markup, which amounts to building a coherent "machine view" of each page, beyond its visual rendering.
- Reduce layout shifts so an agent that has spotted an element can still reach it at the moment it tries to click.
Being recommended and being usable: two questions converging
Agentic browsing asks a different question than grounding and source selection: it is no longer "why was this site cited?" but "can an agent act on it once it gets there?" A site can be perfectly recommended by an engine and still make an agent fail when it tries to fill in a contact form or compare an offer, simply because its structure is not readable by a machine.
For a brand tracking its AI visibility, this adds a layer to monitor beyond citations and recommendation: the site's technical readiness for action, not just for reading. Both topics stay connected through the same fundamentals already useful for GEO, structured content, explicit labels, a stable page, visible freshness, which is why this work naturally extends what is already underway rather than replacing it.
Read next
Agentic browsing
The short definition of the term and its link to a site’s technical readiness for AI agents.
llms.txt: how to build it and its impact on AI visibility
One of the three audits checked by Lighthouse: what this file contains and how to write it.
How AI engines select and cite their sources
The grounding mechanism, complementary to agentic browsing: being chosen as a source before being used as a destination.
Methodology
Understand how AIglebot observes a brand’s presence in the answers of the tracked AI engines.
Generative Engine Optimization
Connect this technical readiness to the broader GEO approach to brand visibility.
Frequently asked questions
What is agentic browsing at Google?
It's an experimental Lighthouse category, Chrome's audit tool, that evaluates whether a site is built to be understood and operated by a software agent, through audits covering WebMCP, accessibility and visual stability, rather than a simple search ranking.
Does the agentic browsing score replace the classic SEO score?
No. It's a separate, experimental and still evolving category that replaces neither the Performance score nor the SEO score in Lighthouse. It adds to them to cover a different use case, an agent acting on the site rather than a user browsing it.
What is the link between agentic browsing and GEO?
GEO aims to get a brand cited or recommended in an AI-generated answer. Agentic browsing goes one step further: once the brand is recommended, it checks whether the agent can actually use the site to complete an action.
AIglebotBetaAIglebot helps companies track their visibility, reputation and recommendation in ChatGPT, Gemini, Perplexity and Claude.