Auditing Website Quality

Overview

Audit a live website across four built-in dimensions — mobile-friendliness, accessibility, PWA-readiness, SEO/performance — plus an optional fifth, GEO/AEO (AI-search citability), when a skill for it is installed. Each dimension is an independent, read-only parallel agent, so adding one is just adding an agent.

Core principle — two layers, and don't confuse them:

1. Static layer (cheap, scriptable). Anything in the raw HTML + HTTP headers: manifest link, meta tags, Open Graph, JSON-LD, robots/sitemap, heading structure, transport headers. scripts/static-audit.sh extracts all of it in one fetch. No browser required.
2. Lab/manual layer (the verdict). Core Web Vitals field data, tap-target size, colour contrast, keyboard/screen-reader operability, offline support. These need Lighthouse, PageSpeed Insights, axe, DevTools, and real testing.

A green static pass is necessary but never sufficient. Automated a11y tools catch only ~30–40% of issues; the authoritative Core Web Vitals verdict is field data, not a lab score. Report a clean static run as "static checks pass," never as "the site is accessible / fast / installable."

When to Use

• Pre-launch or pre-review audit of a site or web app
• "Why isn't my site installable / indexed / responsive / shareable?"
• Producing a prioritized fix list for a frontend
• Verifying a fix actually landed

When NOT to use: pure GEO/AEO (getting cited by AI assistants) — that's the optimizing-for-ai-search skill, which complements this one. Use both for a full web-presence audit.

The audit dimensions

Each dimension is an independent audit unit with its own reference file. The first four are built in; GEO/AEO is an optional fifth (see Pluggable dimensions).

Workflow — parallel dimension audit

Dimensions are independent (no shared state), so audit them concurrently — one read-only subagent per dimension, the same pattern as reviewing-code's lenses.

1. One shared fetch. Run ./scripts/static-audit.sh <url> once. It is a single HTTP fetch covering every dimension; capture its output and hand the same copy to each agent — don't make five agents re-fetch the page.
2. Dispatch one read-only subagent per in-scope dimension, in parallel (add the GEO agent only if that skill is installed). Give each the contract below.
3. Aggregate by severity into one report — Critical / Important / Suggestions. De-dupe overlaps (structured data, SSR, robots show up in both SEO and GEO — keep one finding, note both lenses).
4. Hand back the consolidated lab/manual checklist. Agents can't run a browser (see below); present the merged Lighthouse / PSI / axe / keyboard / screen-reader / DevTools steps for the human (or main agent) to run, then fold results in.

Per-dimension agent contract

Each dimension agent is read-only and receives exactly:

• The target URL(s) and the shared static-audit.sh output.
• One instruction: "Load references/<dimension>.md. Audit <url> for THIS dimension only. Do any deeper dimension-specific static analysis (extra pages, closer DOM/markup inspection). Return findings as: severity · the concrete threshold · location (file/selector/tag) · the fix. Separately list the lab/manual checks you could NOT perform."
• Constraints: don't edit anything; don't stray into other dimensions; return structured data, not prose.

Honest limit — what an agent cannot do. A subagent has no browser, so it does the static + analysis + triage + fix-list layer only. It cannot run Lighthouse, PageSpeed Insights, axe, a screen reader, or DevTools Application — the tools that produce the authoritative verdicts (Core Web Vitals field data, contrast, keyboard/SR operability, installability/offline). Those run outside the agent and get folded in. An all-green agent pass = "static checks pass," never "the site is fast / accessible / installable."

Pluggable dimensions (GEO/AEO and beyond)

A new dimension is a new parallel agent — nothing else changes. GEO/AEO (being reached, read, and cited by AI answer engines) is the natural fifth:

• If a GEO/AEO skill is installed (e.g. optimizing-for-ai-search), dispatch it as another parallel agent with the same contract, pointed at its own audit (it ships its own geo-audit.sh + references), and fold its findings into the same severity buckets.
• If not installed, note GEO as out of scope rather than half-doing it.

GEO and this skill's SEO dimension overlap on structured data, SSR/rendering, and robots/sitemap but optimize different goals (SERP rank + Core Web Vitals vs passage-level citability). When both run, de-dupe shared findings and keep both lenses' framing.

Quick Reference — pass thresholds

Tools (and what static can't see)

Google retired the standalone Mobile-Friendly Test, its API, and the Mobile Usability report (2023-12-01) → use Lighthouse mobile + PSI instead.

The two PWA "bars" (common confusion)

Since Chrome 108 (Android) / 112 (desktop), a service worker is no longer required to be installable via the menu — only a valid manifest + HTTPS. A service worker (with a fetch handler) is still required for offline and to fire the automatic beforeinstallprompt. So distinguish:

• Installable via menu = HTTPS + valid manifest (icons 192+512, display≠browser).
• Auto-prompted + works offline = the above plus a registered service worker.

iOS honors only some manifest fields (name, short_name, start_url, scope, display: standalone, theme_color, icons ≥15.4), ignores background_color/ shortcuts, has no install prompt (Share → Add to Home Screen), and supports Web Push only for installed PWAs on iOS ≥16.4. Details in the PWA reference.

Common Mistakes

Worked Example — resume.schwannden.com (benchmark)

A Next.js static site on GitHub Pages. The shared static-audit.sh pass (step 1) fed the dimension agents; their aggregated findings:

• SEO: strong. Title (53 chars), meta description, lang="en", single <h1>, robots.txt + sitemap.xml (200), HTTP→HTTPS 301, gzip, rich JSON-LD (Person, ProfilePage, Occupation, Organization, Article, WebSite, SoftwareApplication).
• FAIL — no og:image/twitter:image despite twitter:card=summary_large_image → social link previews render with no image. High-value, one-tag fix.
• WARN — no rel=canonical and no HSTS header (GitHub Pages limitation).
• FAIL — zero PWA readiness: no manifest, no theme-color, no apple-touch-icon, no service worker → not installable on any platform.
• Static can't judge: mobile tap targets/CWV, contrast, keyboard/screen-reader → flagged for Lighthouse mobile + PSI field data + axe.

Prioritized fix list this produced: (1) add og:image + twitter:image; (2) decide if PWA-install is a goal — if yes, add manifest + icons + theme-color + apple-touch-icon (+ service worker for offline); (3) add canonical; (4) run Lighthouse mobile + PSI to confirm CWV and tap targets.