SEO Agent

One engine, one config per site. Every run: measure, diff against yesterday, fix the few things that matter through PRs, and write down what moved.

Requirements

• Claude Code with Bash. Python 3 (scripts are stdlib only, except gsc.py: pip install google-auth).
• Optional: Chrome/Chromium (for render_check.py), npm i -g lighthouse (for cwv.py), gh CLI (PRs and preview-host discovery).
• A site config at sites/<slug>.md. Copy sites/example.md and fill the placeholders.
• GSC is optional but turns guesses into measurements. Auth is yours: $GSC_SERVICE_ACCOUNT_JSON (service account added as a Restricted user on your property) or gcloud auth application-default login with the webmasters.readonly scope.

All commands below run from this skill folder (S=scripts). Output goes to ~/.claude/seo-reports/ (reports) and ~/.claude/seo-state/ (snapshots).

Modes

mode	trigger	does
daily (default)	/seo-agent <slug>	steps 1-6, lean and diff-driven
technical	/seo-agent <slug> technical, "deep audit", weekly, or after any routing/middleware/sitemap/layout change	full checklist in references/technical-seo.md
audit-only	the user says read-only	steps 1-4 and 6; every fix becomes a proposal with an owner; no branches, PRs or CMS writes

Steps

1. Load the config

Read sites/<slug>.md first. If it does not exist, stop and offer to create it from sites/example.md (ask for domain, repo, build command, GSC property). Never redo anything under known_wins.

2. Measure (free, no paid APIs)

Run in this order; each writes machine-readable output the report uses.

python3 $S/linkgraph.py <domain>                  # link graph + orphans, snapshot, diff vs last run
python3 $S/crawl.py --sitemap <domain> --sample 30   # per-page title/meta/canonical/H1/words/schema
python3 $S/validate_schema.py <money page URLs>   # JSON-LD parses, @types, canonical mismatch
python3 $S/psi.py <5-10 top pages>                # mobile LCP/CLS/INP + lab TBT (PSI_KEY optional)
bash -c "python3 $S/techaudit.py <domain> --out <dir> --max 3000 --thin 250 --depth-limit 4 \
  --preview-from-gh <owner/repo> $PROBES"         # when linkgraph shows new 4xx/orphans/lost schema

Use bash -c for anything with $PROBES (--probe /api/ --probe /admin ...): zsh does not word-split it and the probes silently vanish.

GSC (if configured): preflight under a 60 s cap, then snapshot every run.

perl -e 'alarm 60; exec @ARGV' python3 $S/gsc.py <property> preflight
python3 $S/gsc.py <property> snapshot                     # 28d pages+queries, movers vs last snapshot
python3 $S/gsc.py <property> searchanalytics --days 28 --dim query --top 200

Check robots.txt by hand if techaudit did not run: GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-SearchBot, PerplexityBot, Google-Extended, Applebot-Extended, CCBot must not be blocked unless that is a deliberate policy.

3. Research

• With GSC: queries at average position 11-30 with real impressions are the cheapest page-1 pushes. Start there.
• Without GSC: WebSearch the config's priorities terms, note who ranks and which SERPs are weak.
• Prefer long-tail problem language (what buyers type before they know the category) over head terms in avoid.

4. Prioritise

Build one table: issue | impact H/M/L | effort S/M/L | evidence (URL + observed value) | action.

Rank in this order:

1. Linkgraph regressions (inlink drops, lost schema, new orphans, title drift). Highest confidence, usually small.
2. Anything blocking indexing: noindex/canonical conflicts, 4xx/5xx in sitemap, robots blocks, preview hosts indexable.
3. GSC page-1 pushes: position 10-20 with impressions in the top quartile for the site.
4. Thin money pages (<600 words), missing/invalid schema, title 25-65 chars and description 70-165 chars violations.
5. CWV failures at p75: LCP >2.5 s, CLS >0.1, INP >200 ms (field data only; lab is diagnosis).

Defer low-impact/high-effort work and anything needing a business call; list it under "needs a human".

5. Ship fixes (review gate)

Code (only if repo_path is set):

• Branch fix/seo-<topic> off default_branch. One concern per PR.
• Run the config's build; never push a failing build.
• gh pr create with: what changed, the evidence row it fixes, build result, what to watch in GSC.
• Never push to the default branch, never merge, never touch design/CSS/components unless asked.

CMS (dispatch on cms.driver; token from the env var named in token_env):

• Patch scalar fields only: title, meta description, excerpt. Do not rewrite rich-text bodies unless asked.
• Hard gate before every patch: python3 $S/lint_brand.py "<new text>" <each brand_voice.forbidden item>. Exit 1 means rewrite and lint again.
• driver: none or token missing: write the change as a proposal in the report instead.

Optional: after shipping changed URLs, python3 $S/indexnow.py <domain> <key> <urls> (Bing and others; key file must be served at /<key>.txt).

6. Report

Write ~/.claude/seo-reports/<slug>-<YYYY-MM-DD>.md (format below). Every 7th run also write <slug>-week-<YYYY-Www>.md: what shipped vs what moved in GSC.

Lean rule: if the linkgraph diff is empty, GSC shows no material movement and no priority changed, write a 3-line "no material changes" report and stop. Do not manufacture work.

Decision rules

• Numbers in reports: same source, same window length, before/after pairs via --start/--end around the change date. Never compare a UI export to an API pull or a 7-day window to a 28-day one.
• Lab vs field CWV never share one claim.
• FAQ rich results only show for authoritative government/health sites; FAQ markup elsewhere is fine for parsers but promises no SERP feature.
• Keyword overlap is not cannibalisation until GSC query-by-page shows two URLs swapping for the same query.
• A preview host (*.vercel.app, *.netlify.app) is yours only after you check its title; look-alike names belong to strangers.

Failure handling

problem	do this
GSC preflight fails or hangs	record the exact error, use the newest ~/.claude/seo-state/<host>/gsc-*.json labelled "last-known, window X", continue; mark GSC rows "not measured today"
no GSC configured	WebSearch site: checks; label every number "inferred, not measured"
linkgraph refuses to save (>5% of URLs failed)	network blip: rerun once; if it repeats, report the failing URLs; use --force only if the failures are real 4xx
sitemap missing/invalid	that is the top critical finding; crawl from / via techaudit.py instead
Lighthouse/Chrome missing	use psi.py (Google-hosted) and say lab data came from PSI
cwv.py slow (~90 s/URL)	run it with nohup ... & and poll its log; long silent foreground commands can time out
build fails on a fix branch	fix it or drop the PR; never push red

Report format

# SEO daily: example.com, 2026-03-04
Crawl: 412 URLs, 398 indexable, 0 5xx. Linkgraph: 2 regressions (below).

| issue | impact | effort | evidence | action |
|---|---|---|---|---|
| /pricing lost 6 inlinks | H | S | footer nav change, 9 -> 3 inlinks | PR #41 restores footer link |
| /blog/x missing Article schema | M | S | validate_schema: 0 JSON-LD | PR #42 |

Shipped: PR #41, PR #42 (build green). CMS: 1 title patch, lint_brand PASS.
GSC 28d vs last: /guide/y pos 14.2 -> 11.8, impressions 1,920 -> 2,410 (same window length).
Needs a human: merge PRs; decide whether /old-landing should 301 to /pricing.
Carry forward: page-1 push for "<query>"; thin /features page (310 words).

Deep technical audits use the severity/evidence/fix/effort/owner contract and report skeleton in references/technical-seo.md.

Running it daily

In-session: /loop 24h /seo-agent <slug> (only fires while Claude Code is open). Unattended: scripts/daily.sh <slug> from cron on a server; see references/deploy.md.