How to read your Lighthouse Agentic Browsing score — what each number actually predicts

If you've run Chrome's Lighthouse Agentic Browsing audit and seen a number between 0 and 100, you've got the headline score. What that number actually predicts about your site — and what the per-audit detail underneath it reveals that the headline hides — is the part most site owners don't unpack.

This post walks the audit's scoring shape: what each score range means in concrete terms, which per-audit signals matter more than others, and how to read the report when you're deciding what to fix first or whether the install you paid for actually worked.

The companion post if you're not a developer is the 5-minute install audit — five browser-only checks instead of running Lighthouse. This post is for buyers who've already pulled up the Lighthouse panel and want to understand what they're looking at.

---

The headline score — what each range predicts

Lighthouse Agentic Browsing emits a single 0-100 number for the category. The translation table:

Score	What it predicts
90-100	Top decile of websites for agent-readiness. Agents reading the site find the kit files, parse them cleanly, can quote your content with structure, and can transact via declared actions. Edge cases only.
75-89	Solid pass. Most agents work; some specific audits may be failing (often schema-org-density partial or webmcp-annotations). Cost of incomplete audits is small in 2026, growing over the next year.
60-74	Partial. The kit files probably exist but at least one is malformed, or auto-discovery tags are missing, or Schema density is below threshold. Agents work for some queries and fail for others — the specific failures predict which queries.
40-59	Substantial gaps. Either the kit files are absent or significantly broken. Agents that find your site struggle to read it cleanly. Search-driven AI traffic loses meaningful share to better-prepared competitors.
0-39	No agent-readiness layer. Agents read your HTML directly with no structured layer. You're missing all of the citation, transaction, and discovery surfaces that 2026 agents look for.

The score is multi-modal in practice — most sites cluster around 0-15 (no install) or 70-85 (install present but partial). Sites in the 40-60 range almost always have a partial or broken install (a llms.txt that fails validation, an agents.json with typing errors, kit files present but no auto-discovery tags).

A 17 → 72 lift is the modal install outcome — the kit closes six of nine audits outright, leaving Schema density and WebMCP and (often) sitemap discoverability as the remaining audits. The kit's reported case-study data sits around that delta, with site-specific variance.

---

What the headline number hides

The category score is a weighted average across nine per-audit scores. The weights as of mid-2026:

Audit	Weight
schema-org-density	20%
agents-json-actions-typed	15%
llms-txt-present	10%
llms-txt-well-formed	10%
agents-json-present	10%
agent-runbook-present	10%
auto-discovery-links	10%
sitemap-discoverable	10%
webmcp-annotations	5%

Weights shift as the category stabilizes. The Lighthouse CLI deep-dive covers the JSON-output shape if you want the per-audit weights pulled programmatically.

What this means in practice:

• Schema density (20%) is the largest single audit. A site that passes all eight other audits but fails Schema density caps out around 80. A site that passes Schema density alone scores 20, which sounds bad but actually closes the audit that drives the most agent-citation traffic.
• agents-json-actions-typed (15%) is the second-largest. Most-missed audit on hand-written files — the deep-dive post walks the typing rules.
• The six middle audits (each 10%) are the file-presence + auto-discovery + sitemap checks. Together they're 60% of the category score, which is why the kit's "six of nine" framing matters — closing those six audits puts you in the 70-85 range even before Schema work.
• WebMCP (5%) is intentionally low-weight. Almost nobody passes it in 2026 and the audit is informational rather than load-bearing.

Reading the score: if your number is 70-85, look at which audits are red. If it's schema-org-density, you have content / theme work ahead (largest impact). If it's webmcp-annotations, ignore it — that's the audit nobody passes. If it's agents-json-actions-typed, your agents.json has typing issues and is the highest-ROI single fix.

---

How to read the audit details panel

When you expand a failing audit in the Lighthouse panel, you get one of three diagnostic shapes:

Pass/fail with a single status

agent-runbook-present: ❌ FAIL
Details: File not found at /agent-instructions.md

The most common shape. The diagnostic tells you exactly what's missing. Fix is mechanical — upload the file, re-run.

Pass/fail with per-item details

agents-json-actions-typed: ❌ FAIL
Issues found:
  - Action "search": parameter "q" has no declared type
  - Action "contact": parameters object is null
  - Action "subscribe": parameter "email" missing format declaration

The details.items array surfaces per-issue diagnostics. Each one points at a specific action or parameter; fix walks the actions-typed deep-dive for the typing rules.

Score with per-page sample data

schema-org-density: ⚠️ PARTIAL (score 0.45)
Pages sampled:
  - /                  Score: 0.7 (Organization, WebSite present; SearchAction missing)
  - /products/wallet   Score: 0.3 (Product present; offers.price empty)
  - /blog/agent-ready  Score: 0.5 (Article present; author empty)

The score-with-sample shape gives you a working list of specific URLs to fix and what's missing on each. The schema-org-density deep-dive covers the JSON-LD shape per page type.

The audit's per-page sampling is small — usually 3-5 inner pages plus the homepage. Sites with many page types should re-run the audit on representative inner pages individually (Lighthouse runs against any URL, not just homepages).

---

What the score doesn't measure

Three things the audit doesn't catch that still matter for agent-readiness:

1. Content quality and accuracy. A Product Schema with a wrong price scores the same as one with the right price. The audit checks structure, not correctness.
2. Update freshness. A llms.txt that hasn't been regenerated in a year still passes the audit if the listed URLs resolve. Agents reading stale files get stale signal — visible in conversation quality, not in audit score.
3. Action-completion rate. An agents.json that declares a purchase action passes the audit if the typing is right. Whether the action actually completes when an agent calls it depends on your backend, your auth flow, your cart logic — none of which Lighthouse tests.

These are the "audit-passing-but-still-broken" cases. The audit is necessary but not sufficient. The honest follow-up is operational monitoring: track agent user-agent failure rates in your access logs, watch for spikes that don't show up in human-traffic monitoring.

---

Practical scoring framework

Use the score as a triage signal:

• Below 30 — install hasn't happened. Skip the per-audit analysis; ship the kit files first. Re-run after.
• 30-65 — partial install or one major audit failing. Look at which of the four highest-weighted audits (schema-org-density, agents-json-actions-typed, the three file-presence audits) is red. Fix highest-weighted first.
• 65-85 — solid install with one or two audits remaining. Schema density is the audit most likely still red. The schema-org-density on product pages deep-dive covers the fix path; for non-product sites the per-audit fix reference walks the per-page-type Schema shapes.
• 85+ — done. Focus on operational quality (content accuracy, freshness, action-completion monitoring).

The score number is a single signal. The per-audit detail is where the install-quality and fix-priority information lives.

---

Related

• Lighthouse Agentic Browsing FAQ — every common question, answered → — direct-answer FAQ for the queries readers run after seeing a score they didn't expect; pairs with this score-reading guide
• Lighthouse Agentic Browsing — every audit, every fix → — the canonical per-audit fix reference for all nine audits
• Chrome added an Agentic Browsing audit to Lighthouse → — the category announcement and overview
• Inside Chrome 146 — what lighthouse --only-categories=agentic-browsing actually does → — the CLI deep-dive with the JSON output shape
• Cloudflare's Agent Readiness Score vs Chrome Lighthouse → — which scoring framework actually predicts agent transactions
• The 5-minute install audit for SMBs who can't read code → — the non-Lighthouse companion for non-technical owners
• How the BridgeToAgent kit maps to every audit → — which audits the kit closes vs which are CMS/theme-side
• Run the free readiness audit → — five-second check on your site