XNews

Web Quality Audit Skill for AI Agents

Web Quality Audit Skill for AI Agents

Addy Osmani released a web quality audit skill for AI agents, covering five dimensions with a rigorous evidence-based methodology. - Five-step workflow: determine audit goals (representative URLs, states, user journeys, public/logged-in, mobile/desktop scope); collect real-time baseline (trace + Lighthouse audit, no broad code scan); locate source code using runtime evidence; grade by user impact and confidence, provide or execute fixes; retest verification (same conditions, report what was verified and what still needs manual confirmation). - Four core design principles: evidence over scores (aggregate scores are diagnostic summaries, real evidence is metric values like LCP, INP, CLS); measure first, then read code (runtime baseline before source review, clearly separate measured conclusions from code-inferred assumptions); evidence type hierarchy (CrUX field data, first-party RUM, DevTools performance trace, Lighthouse lab scores, static code checks—do not conflate); reproducibility discipline (record full measurement conditions, run key conclusions at least 3 times, take median, report range, never compare single local trace to CrUX p75). - Five audit dimensions: Performance (CWV thresholds LCP<2.5s, INP<200ms, CLS<0.1; resource optimization, loading strategies); Accessibility (WCAG four principles, contrast AA 4.5:1, keyboard reachable, focus visible, form label associations); SEO (crawlability, page elements, technical items, use CWV field data as evidence, no ranking promises); Best Practices (security HTTPS/HSTS/CSP, no vulnerable dependencies, no source maps in production, no deprecated APIs, console zero errors, no intrusive interstitials); Agent Browsing (forward-looking category measuring AI agent understanding via semantic HTML and accessibility tree exposure, WebMCP validation, optional llms.txt, explicitly separated from SEO). - Tool routing: preferred Chrome DevTools MCP (performance_start_trace then performance_analyze_insight for performance; lighthouse_audit for other four categories—intentionally excludes performance). Fallback: Lighthouse CLI → PageSpeed Insights → CrUX Vis → static checks. Token-saving principle: text snapshot before screenshot, large reports saved to disk with only actionable items extracted, drill down on few key insights. - Output grading: Critical (security, complete failure → fix immediately), High (CWV failures, major accessibility barriers → fix before launch), Medium (performance opportunities, SEO improvements → fix within iteration), Low (minor optimizations → fix opportunistically). Report structure: evidence table (signal / scope & conditions / result / source), then issues by severity (each with impact, evidence type, specific fix), finally priority order and verification status.

Key insights

Evidence-Based Auditing Prioritizes Real Metrics Over Aggregate Scores

The skill explicitly states that Lighthouse scores are diagnostic summaries whose weights change over versions and cannot serve as proof of quality. True evidence comes from specific metric values (LCP, INP, CLS) and field data. This shifts the focus from chasing scores to improving actual user experience.

Strict Measurement Epistemology Distinguishes Field Data from Lab Data

The skill defines a hierarchy of evidence types—CrUX field data, first-party RUM, DevTools trace, Lighthouse lab, static analysis—and forbids conflating them. It also mandates reproducibility: at least 3 runs, median reporting, and never comparing a single local trace to CrUX p75. This rigor aims to eliminate false confidence in audit results.

Agent Browsing Introduced as a Forward-Looking Audit Dimension

A new category, Agent Browsing, assesses how well AI agents can understand and interact with a page through semantic HTML and accessibility tree exposure. It is explicitly separated from SEO, with no claim that agent-friendliness affects search ranking or AI visibility. This reflects the growing importance of machine-readable web experiences.

Tool Routing Ensures Audits Proceed Even Without Optional Tools

The skill defines a preferred tool (Chrome DevTools MCP) and a clear fallback chain down to static checks. It also includes a token-saving principle to minimize cost. This pragmatic design ensures audits are not blocked by missing tools and remain efficient.

Source posts