Lighthouse Agentic Browsing: Google's New Score for AI-Agent Readiness
Google shipped a new Lighthouse category on May 5, 2026 that scores how well your site can be read and operated by AI agents. It uses a pass-ratio — not a 0–100 score — and most enterprise sites fail it. Here is what it checks, how it is scored, and exactly how to pass.
What Is Lighthouse Agentic Browsing?
Lighthouse Agentic Browsing is a category in Google's Lighthouse tool that scores how well a website can be read and operated by AI agents. Google published the scoring methodology on May 5, 2026, and shipped it in Lighthouse 13.3.0[1]. It answers a question traditional SEO never asked: when an AI agent — ChatGPT browsing, Perplexity, Project Mariner, Gemini's agentic mode — lands on your page, can it actually understand and act on it?
For most enterprise sites today, the answer is no. The category is experimental and based on proposed standards, so the rubric will evolve[1]. That early status is precisely why getting ahead of it is an advantage.
Why Did Google Add an Agentic Browsing Score?
Google added the category because AI agents now interact with the web at scale, and they do not browse the way people do. A human tolerates a slow hero animation or a layout that jumps. An agent — often working from the accessibility tree or a screenshot-and-coordinate map — needs structure, stable positions, and machine-readable labels. When those are missing, the agent fails silently.
Three forces converged in 2026: agentic browsing went mainstream (Project Mariner, ChatGPT browsing, Perplexity, Gemini agentic mode all parse pages programmatically); WebMCP emerged as a proposed standard for exposing a site's tools to agents; and Google needed a benchmark developers could measure against — so it put one inside the tool every technical team already runs.
The result reframes web quality around a second audience. Building for humans is no longer sufficient; the site must also be legible to the machines acting for them.
How Is the Agentic Browsing Score Calculated?
The agentic browsing score does not use the traditional 0–100 weighted model. It uses a fractional pass-ratio — the number of applicable readiness checks your page passes out of the number that apply[1]. A site is reported as a fraction such as 1/3 or 3/3 in the category header, alongside pass, fail, and warning statuses on individual audits.
This is a meaningful design choice. A fraction is concrete and unambiguous: “you pass one of three agent-readiness checks” tells a stakeholder exactly where they stand, with no weighting to debate. Audits that do not apply — for example, WebMCP audits on a site that has not adopted WebMCP — are marked Not Applicable and excluded from the ratio.
Because the audits depend on dynamic conditions — tool-registration timing, accessibility-tree construction, and layout shifts from ads or injected content — results can fluctuate between runs even though each audit is deterministic[1].
What Does the Agentic Browsing Audit Check?
The category groups its checks into three areas. The single most important takeaway: the accessibility tree is not an afterthought for agents — it is how they see.
| Area | What it checks | Why an agent needs it |
|---|---|---|
| WebMCP integration | Registered tools (declarative HTML + imperative JavaScript), form coverage, and schema validity | Lets agents call your site's tools and submit forms directly |
| Agent accessibility tree | Names and labels on interactive elements, role and parent-child tree integrity, visibility of interactive content | The accessibility tree is the agent's primary data model — its map of your page |
| Stability & discoverability | Cumulative Layout Shift (CLS) and the presence of an llms.txt file | Stable coordinates let agents click reliably; llms.txt gives a machine-readable summary |
What Is a Good Agentic Browsing Score?
Passing every applicable check. For a site without WebMCP, that means all applicable stability and accessibility audits.
Google's “good” Cumulative Layout Shift threshold; above 0.25 is “poor.” Agents browsing by coordinates can't click an interface that shifts.[3]
A machine-readable summary file agents can read first. Indexable's own site passes all applicable checks (3/3).
Indexable's Technical SEO Agent Audits Agentic Browsing Readiness
One of the ten agents in Indexable's system is purpose-built for the technical layer of agent-ready websites — it audits every template against the agentic browsing checks and hands engineering a line-level fix sheet.
Why Do Most Enterprise Sites Fail the Agentic Browsing Audit?
Most enterprise sites fail because they were built for human shoppers over the last decade, not for machines. The failure pattern is consistent across large sites:
- JavaScript-heavy single-page apps that render content client-side, so the agent sees a near-empty page.
- High Cumulative Layout Shift from ads, hero animations, and injected content — often many times past the 0.10 threshold. See CLS and AI agents.
- A malformed accessibility tree — unlabeled buttons, broken roles, interactive elements the agent cannot identify. See the accessibility tree.
None of these hurt a human visitor enough to notice. All of them are disqualifying for an agent. That gap — beautiful to people, invisible to machines — is the exact problem the category surfaces.
How Do You Improve Your Lighthouse Agentic Browsing Score?
Improving the score is mostly disciplined fundamentals, applied for a new audience:
- Reduce Cumulative Layout Shift below 0.10. Reserve space for images and ads, avoid injecting content above existing content, and load fonts without reflow.
- Fix the accessibility tree. Give every interactive element a name and a correct ARIA role; keep parent-child relationships intact. This doubles as human accessibility improvement.
- Prioritize semantic, server-rendered HTML. Deliver content in the initial HTML, not only after client-side hydration, so agents read it without executing your app.
- Add an llms.txt file at your root with a machine-readable summary of your site.
- Adopt WebMCP to expose your forms and tools to agents declaratively — the forward-looking step as the standard matures.
You can run the audit yourself with the Lighthouse CLI. See the step-by-step guide:
How Does Agent-Readiness Relate to AI Search Visibility?
Agent-readiness is the technical foundation beneath AI search visibility (GEO/AEO). Getting cited by ChatGPT or Perplexity depends on content being structured for retrieval — but increasingly, those systems also send agents to read and verify your page in real time. If the agent cannot parse the page, the strongest content in the world never gets surfaced.
Think of it as a stack: agent-readability is the floor, structured content is the middle, and citation is the result — the distinction we draw in SEO vs AEO vs agent-readiness. A high agentic browsing score does not guarantee citations, but a low one caps how often you can be cited, because the layer underneath is broken.
Fix the architecture and the rest of your AI-search program compounds; skip it, and nothing above it fully works. For the standard that powers the WebMCP audits, see What is WebMCP; for the broader thesis, see Agentic Websites: Why the CMS Era Is Over.
Frequently Asked Questions
Is Lighthouse Agentic Browsing the same as the SEO or Accessibility score?
No. It is a separate, experimental category that scores machine-readiness for AI agents, using a fractional pass-ratio rather than a 0–100 score. There is overlap with Accessibility — the agent accessibility-tree audits build on accessibility best practices — but the audience and scoring differ.
Does a low agentic browsing score hurt my Google rankings?
Not directly today; the category is experimental. But it measures the same structural qualities — semantic HTML, stability, machine-readability — that increasingly determine whether AI search surfaces cite you, which is where a growing share of discovery happens.
Do I need WebMCP to score well?
No. WebMCP audits are marked Not Applicable if you have not adopted it, so they do not lower your ratio. You can pass the applicable checks with strong CLS, a clean accessibility tree, and an llms.txt file. WebMCP is the forward-looking upgrade.
How often should I run the audit?
Run it after any significant template or layout change, and on a recurring schedule for key templates. Because results can fluctuate with dynamic content, compare trends rather than single runs.
Most Enterprise Sites Pass 1 of 3. Yours Could Pass All Three.
Indexable's ten-agent system audits agentic browsing readiness across every template, fixes the gaps, and monitors your score as the standard evolves.
Primary Sources Cited
- Google / Chrome Developers. Lighthouse Agentic Browsing scoring documentation. Published May 5, 2026; shipped in Lighthouse 13.3.0. developer.chrome.com/docs/lighthouse/agentic-browsing/scoring
- W3C Web Machine Learning Community Group. WebMCP (navigator.modelContext) draft. webmcp.link
- web.dev. Cumulative Layout Shift (CLS) thresholds: good ≤ 0.10, poor > 0.25. web.dev/articles/cls