Page Analysis.

A vision-capable LLM (Claude / GPT-4o / Gemini 2.x) audits a full-page screenshot together with the rendered HTML, captured console messages, and a network-traffic summary. Returns a numerically scored list of issues across eight categories — the baseline every other test type builds on.

What it does analyzePage()

Playwright loads the URL in headless Chromium (stealth-patched: real Chrome 126 UA, navigator.webdriver patched, no jank fingerprints). After DOMContentLoaded plus a 10s networkidle ceiling, the runner captures a full-page JPEG, the post-render HTML, the console stream (capped at 2,000 chars per message), and a network summary (URLs, methods, statuses — never headers or bodies). All of that goes to the configured LLM with a structured prompt. The model returns strict JSON: a 0–100 score, an array of issues, and a one-line summary. Each issue has a category, severity, confidence (1–10), priority (1–10), affected selector, and a suggested fix.

What it finds

Concrete examples from a real run:

• Visual: "Hero CTA button overlaps the navigation bar at viewport widths between 1100–1280px" — overflow, truncation, contrast clashes, misaligned columns
• Usability: "Primary signup button labelled 'Continue' is below the fold on a 1366px viewport; visitors don't see it without scrolling" — hidden CTAs, dark patterns, dead-end flows
• Performance: "Hero image is 4.2 MB unoptimized PNG; LCP delayed by ~1.8s on a 3G connection" — render-blocking scripts, oversized payloads, slow third-party calls
• Reliability: "Uncaught TypeError in analytics.js:142 on every pageview — telemetry is silently failing" — console errors, failed network calls, broken images
• Security: "Page served over HTTPS but loads http://cdn.example.com/widget.js — browsers will block this as mixed content" — mixed content, exposed tokens, missing CSP/HSTS
• Accessibility: "Login form's password field has no associated <label> — screen readers will announce it as 'edit text, blank'" — quick a11y wins (full audit is the WCAG test)
• Content: "Footer still says '© 2023 Acme Corp'" — stale copy, placeholder text, broken external links, missing legal pages
• SEO & meta: "Page has two <h1> tags and no <meta name=description> — Google will likely truncate the snippet" — canonical issues, missing Open Graph, duplicate headings

Coverage

• Static observation — doesn't click through multi-step flows (that's Test Flows and Exploratory)
• No keyboard-navigation traversal — WCAG covers that with a dedicated in-browser audit
• No real-device performance profiling — payload sizes and timing only, not CPU/memory traces
• No mobile-specific bug surfacing here — the responsive phase runs separately at 768×1024 and 390×844
• Won't catch issues that only appear post-login or behind interactive state (use the auth option to authenticate, or define a Custom Test)

Sample finding

// One entry from report.analysis.issues[]
{
  "category":    "reliability",
  "severity":    "high",
  "title":       "Uncaught TypeError in analytics.js:142",
  "description": "Pageviews trigger 'Cannot read properties of undefined (reading \\'session\\')'.\nTelemetry is silently failing on every visit — A/B test data + funnel attribution\nis incomplete from this date forward.",
  "selector":    null,
  "confidence":  9,
  "priority":    8,
  "interestingness": 7,
  "evidence":    ["console.log: TypeError @ analytics.js:142"],
  "prompt_to_fix_this_issue": "Open analytics.js around line 142.\nThe code reads `window.user.session.id` — `window.user` is undefined for\nsigned-out visitors.  Add a null-check, or initialise window.user early."
}

See also