GEO Audit Toolkit — Claude Code skill

Philosophy: audit-only, evidence-based, no sales framing. Findings are reported in severity × confidence × per-LLM impact, never in pricing tiers or revenue projections.

Single command

/geo audit <url> [--positioning "<query>"] [--vs comp1.com,comp2.com] [--output-lang fr|en]

Examples:

# Audit with positioning + 2 competitors, output in French
/geo audit https://example.com --positioning "agence powerpoint" --vs mprez.fr,laboiteaslides.fr --output-lang fr

# Audit without RRF measurement (on-page only)
/geo audit https://example.com

Internally calls python3 -m scripts.run_audit <url> .... The orchestrator is sequential (not parallel-subagent) and produces a single Markdown audit at site-client/<slug>/audits/<date>-audit.md.

Per-LLM retrieval map (rev. 2026-04-26)

Bing Copilot, Meta AI, Grok are out of scope (user decision, 2026-04-26). Mapping validated empirically: 80–90% Google SERP similarity vs ~30% Bing when comparing ChatGPT citations to scraper-baseline. See docs/LLM_RETRIEVAL_ARCHITECTURE.md.

Audit pipeline (5 phases, sequential)

Audit output structure

1. État des lieux par LLM        ← verdict table + 5 detail blocks
2. Plan d'action                  ← unified, per-LLM impact tags + experimental sub-bucket
3. Ce que l'audit ne mesure pas

Annexes techniques:
  A. RRF mesuré (détail)
  B. Politique des crawlers IA
  C. Données structurées (schema)
  D. Content health (BLUF, etc.)
  E. Findings — vue agrégée

Glossaire technique

Every audit lands in site-client/<slug>/audits/<YYYY-MM-DD>-audit.md. Path resolved via scripts/client_paths.py.

Active sub-skills (6)

The fork dropped 9 upstream skills that did not fit the audit-only pivot (geo-prospect, geo-proposal, geo-brand-mentions, geo-citability, geo-llmstxt, geo-platform-optimizer, geo-technical, geo-content, geo-report, geo-report-pdf, geo-compare). See tasks/lessons.md.

Local-only tooling

local/ is gitignored. Contains personal scripts that don't ship in the public toolkit:

• local/docx_export.py — convert audit Markdown to a styled .docx (Google Docs ready). Usage:

  python3 local/docx_export.py site-client/<slug>/audits/<date>-audit.md

Quality gates

• Audit-only: no pricing tiers, no MRR projections, no "client-ready" framing. Findings carry effort × score-gain × confidence.
• Evidence triage: every finding tagged with confidence (high → experimental). Single-study correlations and emerging patterns are flagged, not promoted to critical.
• Free, no-setup: pipeline runs without paid APIs or signup. SERP scraping uses public interfaces (Google via Startpage proxy, Brave HTML).
• Output language follows audited site's locale (override with --output-lang fr|en).
• Per-LLM retrieval map re-verified every 6 months — partnership deals change quickly. See tasks/lessons.md (2026-04-26 entry).

Reference docs

• docs/LLM_RETRIEVAL_ARCHITECTURE.md — per-LLM backend mapping
• docs/CRAWLERS_TAXONOMY.md — 29 AI bots catalogued
• docs/STRUCTURED_DATA_BEST_PRACTICES.md — Schema.org / EEP methodology
• docs/CONTENT_HEALTH_PATTERNS.md — Mount AI anti-patterns + BLUF
• docs/BRAVE_GRAIL_SIGNALS.md — Brave-rooted LLM evidence (Test #1 result)
• docs/sources/ — 8 archived methodology articles with critical-thinking triage