02-audit

02 · Audit — Claude Visibility Baseline

This command measures what Claude tells real users when they ask the brand's target queries — without telling Claude who the client is.

We deliberately scope the baseline to Claude only. Multi-LLM coverage (ChatGPT / Perplexity / Gemini / AI Overviews) sounds appealing but requires per-engine API keys and per-engine output normalization, which turns the plugin into a heavy ops project. Single-engine + reproducible beats multi-engine + flaky.

If you later need cross-engine coverage, the vendored seo-geo-optimizer (199-bio) skill has multi-engine analysis paths (see its scripts/platform_optimizer.py).

---

Inputs

• clients/<slug>/brand_context.json — required. Reads target_queries, layer_1_business_identity.company.name, competitors.

Output

• clients/<slug>/visibility_baseline.json (round 1) — validates against schemas/visibility_baseline.schema.json.
• clients/<slug>/baselines/round-N.json (round 2+) — preserved snapshots for 07-reaudit diff.
• clients/<slug>/baseline-report.md — human-readable summary.

---

Procedure

Step 1 — Prepare query list

jq '.target_queries | map({query, query_id: (.query | gsub(" "; "-") | ascii_downcase), priority, intent})' \
  clients/<slug>/brand_context.json > /tmp/queries.json

Strip P2 if budget-tight. Default: run all P0 + P1 + P2.

Step 2 — Define the unbiased query runner

For each query, spawn a fresh sub-agent context that does NOT see the brand_context. The sub-agent answers the query like a normal user — it uses WebSearch + WebFetch as Claude does by default, no system prompt nudging it toward our client.

Procedure for each query (loop):

Sub-agent task prompt (template):
  "You are answering a user's question. The user asked: <query>.
   Search the web (WebSearch + WebFetch as needed), then write a
   substantive answer (300-600 words) the way you would if asked
   conversationally. List specific brands/companies/products by
   name when relevant. Include the URLs you actually cited."