Google SEO Audit - multi-agent page audit

Audits a page across every Google-discoverability dimension and returns a scored report with prioritized, cited recommendations. Every recommendation must be grounded in the local corpus via the `google-search-ads-analytics-docs` skill (.claude/skills/google-search-ads-analytics-docs/search.py) and cite a Google source_url.

Fast path - the Workflow

For a thorough, adversarially-verified audit, run the bundled `seo-audit` workflow (Workflow tool), saved at .claude/workflows/seo-audit.js. It implements the four phases below and scales agents to the work - no fixed cap: Snapshot -> Verify (one agent per dimension) -> Refute (adversarial verifiers per finding, false positives dropped) -> Synthesize (AUDIT-<page>.md). Pass the target as args (URL or local path); optional { refuters: N } raises the adversarial panel size. If the Workflow tool isn't available, do the same steps manually with the Task tool, below.

Step 1 - Pick the target(s)

• A live page: pass the URL.
• This repo (default when the user says "my page/repo"): find local pages

with Glob "**/*.html" (ignore .venv/, node_modules/). The repo's page is app/index.html. Audit each found page.

Step 2 - Snapshot (deterministic, run once per target)

Use a per-target output file so multiple pages don't clobber each other:

.venv/bin/python3 .claude/skills/google-seo-audit/fetch_page.py <url-or-path> --out /tmp/snap-<page-slug>.json

Produces a JSON of: status/headers/HTTPS, title, meta description, meta robots, canonical, viewport (mobile), lang + hreflang, H1/heading outline, word count, images+missing-alt, JSON-LD structured data (types + a parse-OK flag - not schema validation; the structured-data agent must still verify required props), Open Graph, analytics/ads tags (GA4/GTM/Ads), and robots.txt (remote only).

Step 3 - Fan out the audit (one subagent per dimension)

Spawn the dimensions in parallel (Task tool, or the Workflow). Give each agent the snapshot JSON and tell it to consult `google-search-ads-analytics-docs` first - run .venv/bin/python3 .claude/skills/google-search-ads-analytics-docs/search.py "<topic>" (and --doc <doc_id> to resolve a citation), then read the matched files - and cite source_url on every finding. Dimensions:

1. Crawlability & Indexing - HTTP status, redirects, meta robots /

X-Robots-Tag, canonical, robots.txt, sitemap. KB: crawling-indexing/*.

1. On-page SEO & Content - title (len/uniqueness), meta description,

exactly one H1, heading hierarchy, content depth, helpful-content signals. KB: fundamentals/seo-starter-guide, creating-helpful-content, appearance/snippet, appearance/title-link.

1. Structured Data & Rich Results - JSON-LD presence, valid types, required

props, rich-result eligibility. KB: appearance/structured-data/*.

1. Page Experience & Core Web Vitals - HTTPS, HSTS, mobile viewport,

intrusive-interstitial risk, CWV considerations. KB: appearance/page-experience, appearance/core-web-vitals, crawling-indexing/mobile/*.

1. International (hreflang) - lang, hreflang alternates, locale handling

(only if the page is multilingual). KB: specialty/international/*.

1. Measurement & Tagging - GA4/gtag, GTM, Google Ads conversion, GA4

events/Measurement Protocol readiness. KB: google-analytics/*, google-ads-help/*.

Each finding: {check, verdict: pass|warn|fail, evidence, recommendation, severity: high|medium|low, citation_url}.

Step 4 - Synthesize the report

Aggregate into AUDIT-<page>.md:

• Score: each dimension 0-100; overall = average (call out fails).
• Executive summary (3-5 lines).
• Prioritized fixes: a table sorted by severity -> effort, each with the

concrete change and a source_url citation.

• Per-dimension findings (pass/warn/fail with evidence).
• Be honest: don't invent issues; mark "N/A" where a dimension doesn't apply

(e.g. hreflang on a single-language page).

Notes

• Requires the project venv (.venv/bin/python3) with Scrapling, and the

google-search-ads-analytics-docs skill present for grounding.

• Local files can't show HTTP-level signals (status/headers/HTTPS/robots.txt) -

flag those as "deployment-dependent, verify on the live URL".

Safety

• Authorization: only audit pages you own or are explicitly authorized to

test. fetch_page.py issues a normal browser-like GET to the target and its /robots.txt; do not point it at internal/private hosts you don't control (it has no SSRF allow-list - it fetches whatever URL you pass).

• No credentials, no writes to the target. The snapshot is read-only and is

written only to the local --out JSON (or stdout). Nothing is sent anywhere except the HTTP GET to the page being audited.