seo-orchestrator (Layer 2)

Coordinates the whole audit. Three phases: detect → dispatch → synthesize.

1. Detect

1. Resolve the target (URL or local path).
2. Run seo-crawl-render to build the shared PageSnapshot (raw HTML, rendered DOM when available, response headers, status chain, plus site artifacts: robots.txt, sitemap set, llms.txt). Record the data tier reached (0/1/2).
3. Run seo-vertical-detect to classify the target. Read references/routing.md to get the always-on + conditional module set and the subagent dispatch table.

2. Dispatch (parallel)

Spawn the read-only specialists in parallel (one message, multiple Task calls) so their verbose intermediate output stays isolated. Pass each the PageSnapshot location/contents, the detected vertical, and its assigned modules:

• technical-auditor → M1, M2 (+M3), M4, M7, M7b (mobile), M7c (headings), M8, M9, M10, M15, M17
• ai-search-geo-specialist → M6, M11, M12, M14, M21
• content-eeat-analyst → M13, M16
• schema-generator → M5 (and M18/M19 schema when those verticals are active)

Each subagent returns an array of findings conforming to schema/finding.schema.json. Subagents are read-only (no Write/Edit tool) — the audit can never mutate files.

3. Synthesize

1. Merge all findings; dedupe by id (keep the most severe status).
2. Run seo-score over the merged findings to get the two scores with category breakdowns and severity gating.
3. Render the report (see references/scoring-model.md for bands/interpretations):

• Header: Search SEO band/score + AI Visibility band/score, each with a one-line interpretation.
• Per-category table for each score (value, weight, active?).
• Data confidence: tier reached + count of needs_api findings.
• Prioritized actions: top findings sorted by severity × magnitude ÷ effort, each showing status, evidence, recommendation, fixability (auto/proposed/advisory), and expected_impact (axis + confidence + magnitude).
• Footer: how to run /claude-seo-ai:fix.

Notes

• Keep this skill's own output tight — the value is the merged report, not narration.
• If the target is a multi-page crawl, analyze the homepage + one instance of each detected template first, then sample additional pages; log what was sampled (never silently cap coverage).
• Degrade gracefully: if a render/PSI/GSC tier is unavailable, proceed at the lower tier and mark affected findings needs_api.