Brand Coherence Audit
JuJu78/brand-coherence-audit
Audite consistência de marca para mecanismos de busca com IA
Ver no GitHubSkill que audita a consistência discursiva de uma marca ou pessoa em até 15 fontes web (site oficial, LinkedIn, Wikipedia, podcasts, Crunchbase, etc.). Detecta inconsistências que prejudicam o reconhecimento por mecanismos de busca com IA (Perplexity, ChatGPT Search, Claude com web) e gera relatório estruturado com scores ponderados e plano de ação.
Instalação
npx convez-ai JuJu78/brand-coherence-audit1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Auditoria de posicionamento de marca antes de aparecer em buscas com IA
- Identificar e corrigir contradições entre site, redes sociais e menções na web
- Validar que entidades (pessoa/empresa) têm narrativa consistente para reconciliação por LLMs
Veredito

Ferramenta especializada e bem-pensada que resolve um problema real (consistência discursiva para IA search) que ferramentas tradicionais de brand monitoring ignoram. Requer configuração de múltiplas chaves de API e entendimento técnico, mas entrega valor claro para estratégia de visibilidade em busca generativa.
README
A tradução para o português deste README ainda está na fila.
brand-coherence-audit
Audit the discursive coherence of a brand or person across the web — built for the GEO (Generative Engine Optimization) era.
A Claude Code skill that scrapes a brand or person's footprint across up to 15 web sources (official site, LinkedIn, Wikipedia, press, podcasts, Crunchbase, etc.), extracts structured claims (positioning, offer, expertise, proofs), and surfaces the inconsistencies that hurt your chances of being cited by Perplexity, ChatGPT Search, Google AI Mode, Claude with web, and Brave Ask.
Why this skill exists
Modern AI search engines perform entity reconciliation before they cite a source. When your website says one thing, your LinkedIn another, and a podcast interview a third, the LLM lowers its confidence and falls back to safer, more consistent sources.
Existing brand monitoring tools measure sentiment and mentions. None of them measure discursive consistency — the single signal that determines whether AI engines confidently understand who you are.
This skill fills that gap.
What you get
A markdown report with:
- 4 sub-scores (Positioning 35%, Expertise 30%, Offer 20%, Proofs 15%) plus a weighted aggregate.
- A cross-source coverage matrix showing what each source claims on each axis.
- A prioritized list of inconsistencies, each with verbatim quotes from at least 2 sources and a recommended fix.
- A 5-step action plan that says exactly which source to update and how.
The report is rendered inline in the chat — nothing is written to disk by the skill itself.
Installation
Clone into your Claude Code skills directory:
git clone https://github.com/JuJu78/brand-coherence-audit ~/.claude/skills/brand-coherence-audit
Or for project-scoped use:
git clone https://github.com/JuJu78/brand-coherence-audit ./.claude/skills/brand-coherence-audit
Restart Claude Code (or reload skills) and the skill becomes invocable.
Required tools
The skill orchestrates external tools — it does not bundle any code. You need at least one tool from each layer.
API keys at a glance
Before running the skill, know exactly which keys you need. Set them as environment variables — never paste a key into a file or anywhere git-tracked.
| Key (env var) | Powers | Required? | Free tier | Where to get it |
|---|---|---|---|---|
BRAVE_API_KEY | Source auto-discovery (step 2) | Recommended. Only skippable if you always pass an explicit --sources list. | 2,000 queries/month (~100 audits) | brave.com/search/api |
APIFY_API_TOKEN | Social profiles: LinkedIn, X/Twitter, Instagram, YouTube, Crunchbase (step 3a) | Optional. Without it, social sources degrade to ~10-20% signal. | Pay-as-you-go, ~$0.01 per audit | console.apify.com/account/integrations |
| (none) — Firecrawl MCP | Premium scrape fallback (step 3) | Optional. Configured via its own MCP, not an env var here. | Free quota on firecrawl.dev | See Firecrawl MCP |
Bare minimum to run an audit: just BRAVE_API_KEY (for discovery) plus any one scraping tool from the layer below — even Claude Code's built-in WebFetch works. Everything else only improves coverage.
Discovery layer
- Brave Search API — recommended, used by default. Requires the
BRAVE_API_KEYenv var. Get a free key at brave.com/search/api (2,000 queries/month on the free tier — more than enough for ~100 audits).- Set persistently on Windows:
[Environment]::SetEnvironmentVariable("BRAVE_API_KEY", "<your-key>", "User")(PowerShell) - Set persistently on Linux/macOS: add
export BRAVE_API_KEY="<your-key>"to~/.bashrcor~/.zshrc - The skill calls the API via
curlfrom within the workflow — no MCP server needed.
- Set persistently on Windows:
Why Brave Search API and not Brave Ask: faster (~1-2s vs ~30s), returns 20 results vs 5-10, structured JSON output that's easier for the skill to parse, and the AI-synthesized answer of Brave Ask isn't needed for source discovery.
Scraping layer (the skill uses a fallback chain)
- Primary (free, recommended): any MCP exposing a
scrape_url_contenttool. The skill is built around the pattern "HTTP simple → Crawl4AI fallback" — fast on static and lightly-rendered pages, free, and good enough for ~80% of sources. - Premium fallback: Firecrawl MCP — only invoked when the primary returns insufficient content. Preserves your free Firecrawl quota.
- Last resort: Claude Code's built-in
WebFetch.
If you only have one of those, the skill still works — it just degrades gracefully.
Social-platform layer (recommended for personal-brand audits)
- Apify MCP — for LinkedIn profiles, X/Twitter accounts, Instagram bios, YouTube channel pages, and Crunchbase organizations. Without it, the skill falls back to unauthenticated scraping which only captures ~10-20% of the brand-relevant signal on these platforms (LinkedIn shows only the headline behind a sign-in wall, X hides bio sub-fields, etc.).
- Configure once at user scope: add an
apifyentry in~/.claude/.mcp.jsonpointing tohttps://mcp.apify.com/?tools=actors,docswithAuthorization: Bearer ${APIFY_API_TOKEN}. SetAPIFY_API_TOKENas a system env var (Windows:setx APIFY_API_TOKEN "<key>"; Linux/macOS: addexportto your shell profile). - Cost: ~$0.01 per audit (LinkedIn + X + Instagram all included). Get an API token at console.apify.com/account/integrations.
- Trigger: pass
--social-profiles(or list LinkedIn/X URLs in--sources) to activate. Auto-discovery also routes any detected social URLs through Apify automatically.
Usage
/brand-coherence-audit "<subject>" [--type=person|brand] [--depth=quick|standard|deep]
Examples
/brand-coherence-audit "Stripe" --type=brand --depth=standard
/brand-coherence-audit "Yann LeCun" --type=person --depth=deep
/brand-coherence-audit "Anthropic"
# (auto-detects type and uses standard depth)
Optional flags
--sources=url1,url2,...— bypass auto-discovery and audit exactly these URLs.--exclude=url1,...— exclude specific URLs/domains from auto-discovery.--language=fr|en— force language. Auto-detected from the official site by default.
Depth presets
| Mode | Sources | Approx. duration | Includes |
|---|---|---|---|
quick | 5 | ~3 min | Top 5 highest-priority sources only |
standard (default) | 8-10 | ~7 min | Adds press articles and YouTube descriptions |
deep | up to 15 | ~15 min | Adds podcast/video transcripts and Crunchbase / Producthunt |
Output structure
# Audit de cohérence discursive — [Subject]
Score global : XX/100 — [Verdict]
## Détail par axe (table)
## Synthèse exécutive
## Sources analysées (table)
## Matrice de couverture (one table per axis)
## Incohérences détectées (priorisées)
- 🔴 Critique
- 🟠 Modérée
- 🟡 Mineure
## Plan d'action priorisé (5 actions max)
## Limites de l'audit
What this skill is NOT
- Not a sentiment analyzer. It does not measure how people feel about the brand.
- Not a mention tracker. It does not count or monitor mentions over time.
- Not a SEO tool. It does not analyze keywords, backlinks, page speed, or any classical ranking factor.
- Not a Wikipedia-style fact-check. It only flags inconsistencies between sources the brand controls or influences. It does not evaluate truthfulness against an external ground truth.
For competitor-on-topic content gap audits, see entity-territory-audit (sister skill).
Scoring philosophy
Every finding must be verifiable: at least 2 sources quoted verbatim, both contradicting each other on a structured claim. No interpretive findings. No "tonal mismatch" verdicts. If you cannot quote both sides, the finding is dropped.
The 4 axes are weighted by their impact on LLM citation likelihood:
- Positioning (35%) — the LLM's first reconciliation question is "who is this entity?". Inconsistency here is the most damaging.
- Expertise (30%) — directly determines on which topics you get cited.
- Offer (20%) — matters for commercial-intent queries.
- Proofs (15%) — credibility booster but tolerates some natural variation (e.g. headcount evolves).
Full scoring rubric: see references/scoring-model.md.
Roadmap
- v1 (current) — Single-shot audit, markdown report, FR + EN.
- v2 (planned) —
--diffmode comparing two runs over time (track improvements after fixes). - v3 (maybe) — JSON output mode for CI-style coherence regression testing.
Contributing
Issues and PRs welcome. The skill is intentionally MCP-orchestrated, not Python-bundled, to stay portable across Claude Code installations. Keep that constraint in mind for contributions.
License
MIT.
About the author
Built by Julien Gourdon, SEO consultant specialized in AI search and Generative Engine Optimization (GEO). I help brands and independent experts become consistently understood, and confidently cited, by AI search engines such as ChatGPT Search, Perplexity, Google AI Mode and Brave Ask.
- 🌐 Website: julien-gourdon.fr
- 📖 The full method behind this skill (in French): Audit de cohérence de marque pour les moteurs IA : méthode complète
- 🐙 More open-source skills: github.com/JuJu78
If this skill is useful to you, a ⭐ on the repo helps others find it.