How our audit works
All 55 checks and why an AI never touches the score.
Read →Agent view — served from /lighthouse-agentic-browsing.md.
Guide · updated September 2026
What is the Lighthouse Agentic Browsing audit?
Agentic Browsing is an experimental Lighthouse category, shipped in version 13.3 in May 2026, that checks whether a page is usable by AI agents. It was the first time a major web-quality tool scored sites for agent readiness. It reports pass or fail on a small number of checks rather than a 0–100 score, and Google labels it experimental.
What does it actually check?
A short list, focused on whether an agent can parse and operate the page: the presence and structure of agent-facing files, whether forms carry machine-readable annotations, whether a capability manifest exists, and page stability during load.
| Area | What it looks for |
|---|---|
| Agent files | llms.txt and agents.json presence, and whether declared actions are typed |
| Forms | WebMCP annotation coverage and schema validity |
| Capabilities | An agent runbook describing what the site can do |
| Stability | Cumulative Layout Shift and accessibility tree quality |
Notably absent: content negotiation, JavaScript rendering dependency, crawler policy, structured-data accuracy, and anything about whether an agent can complete a real task end to end.
Why does it give pass/fail instead of a score?
Because the standards it measures are not settled. A 0–100 score implies agreed weightings, and there is no consensus on how much llms.txt matters relative to form annotations — partly because Google’s own advocates have publicly questioned llms.txt.
Pass/fail is the honest output for a field this young. It also means Lighthouse cannot tell you whether you improved, which is the gap our versioned scoring exists to fill.
Is agentic browsing a ranking factor?
No, and Google has not suggested it is. It is a diagnostic in a developer tool, marked experimental, and experimental Lighthouse categories have been removed before.
The useful reading is directional rather than tactical. Google building the audit at all indicates it expects agent traffic to matter. That is worth acting on. Treating four experimental checks as a ranking checklist is not.
How does it compare to a full agent readiness audit?
| Lighthouse Agentic Browsing | Code4X | |
|---|---|---|
| Checks | 4 | 55 |
| Output | Pass / fail | 0–100, versioned |
| Tracks change over time | No | Yes, same rubric |
| Generates fixes | No | Yes |
| Cost | Free | From $4,500 |
| Status | Experimental | Production |
Run Lighthouse first. It is free, it is in your browser already, and if it passes everything you have cleared a genuine bar. It simply covers a narrow slice, and passing it tells you nothing about the fifty-one checks it does not perform.
Where this is contested
The obvious risk to a business like ours is that Google ships a proper 0–100 agentic score and commoditises the scoring layer entirely. We think that is reasonably likely, and we would rather say so than pretend otherwise.
It would not remove the need for remediation, longitudinal comparison, or representation monitoring — Lighthouse has never fixed anything it measured. But anyone whose product is only a score should be planning for it, and you should ask any vendor how they answer that question.
Questions
Frequently asked
What version of Lighthouse has the agentic browsing audit?
Lighthouse 13.3, shipped in May 2026. The category is labelled experimental and reports pass or fail rather than a 0-100 score.
Is Lighthouse agentic browsing a Google ranking factor?
No. Google has not indicated it affects ranking. It is a diagnostic in a developer tool, marked experimental, and experimental Lighthouse categories have been removed in the past.
What does the agentic browsing audit not check?
Content negotiation, JavaScript rendering dependency, AI crawler policy in robots.txt, structured-data accuracy, and whether an agent can complete a real task end to end.
Should I use Lighthouse or a paid agent readiness audit?
Run Lighthouse first, since it is free and already in your browser. It covers four checks. A full audit covers 55, tracks change on a fixed rubric, and generates the remediation code.
See how AI reads your site
Eight checks against your domain in about fifteen seconds. No email.