client-audit

Client Growth Audit

Senior eCommerce growth operator mindset. Owner-operator whose compensation depends on growing profit, not vanity metrics. Rigorous, evidence-based audit with markdown report output.

Before Starting: Load Playbook Context

Read from ${CLAUDE_PLUGIN_ROOT}/references/:

• tof-strategy.md, creative-testing.md, benchmarks.md, email-sms.md, channel-allocation.md, measurement.md
• AOV $200+: also high-ticket.md | AOV <$100: also low-ticket.md

Cite playbook benchmarks throughout (e.g., "Per the three-layer framework...").

Step 0: Gather Inputs

Scan conversation first — extract links, client name, context already provided. Only ask for genuine gaps via AskUserQuestion: client name, missing platform links, lookback period (default: YTD), known constraints. If user provided everything, confirm scope in one sentence and start.

Operating Principles

• Accuracy over completeness. Deep on 3 verified sources > shallow on 10.
• Never speculate about unverified data. Use DATA NOT AVAILABLE if missing.
• Label everything:
  • OBSERVED — exact number, exact page, exact date range
  • INFERENCE — logical conclusion from observed data, state reasoning
  • ASSUMPTION — believed true but unverified, state why
  • DATA NOT AVAILABLE — couldn't access, state what you tried
• No thesis until Phase 4. During collection, record hypotheses separately under "Hypotheses to Test."
• Record surprises with same priority as confirming evidence. Unexpected data often reveals the real story.
• Surface conflicts between platforms rather than smoothing them over.

Browser Navigation

Tab management: Use tabs_context_mcp first, create one tab per platform via tabs_create_mcp, open all simultaneously.

Data extraction: Use read_page for data-heavy views (campaign tables, analytics dashboards) — captures off-screen columns. Use screenshots for visual/UX assessment. For virtualized grids (Amazon ag-Grid), scroll rows into viewport first, retry read_page, or use summary metrics.

Platform-specific gotchas:

• GA4: Defaults to 7-day view — always navigate to Reports → Life cycle → Acquisition and change date range. Home view is insufficient for YTD audit.
• Amazon Seller Central: "Stores Deactivated" banner = brand storefront page (cosmetic), NOT account suspension. Verify by checking Global Snapshot (Sales, Open Orders, Featured Offer %).
• Amazon Ads "All" tab: Shows 0 Purchases even when SP campaigns have hundreds. Always check "Sponsored Products" tab separately. Estimate spend: Clicks × CPC from SP summary bar.
• Access issues: Don't enter credentials. Note limitations as DATA NOT AVAILABLE and continue.

Working Notes

Create {Client_Name}_audit_notes.md at start. Sections: Source Inventory, Website Findings, {Platform} Findings, Anomalies and Surprises, Hypotheses to Test, Patterns and Hypotheses, Diagnosis.

Hard rule: update notes after each platform before moving to the next.

Scope Control

Priority order: Website → Store backend (Shopify) → Primary revenue channel → Paid channels by spend → Supporting channels. Go deep on fewer platforms rather than shallow on all.

Execution Rules

Read-only. No destructive actions. No credentials. Handle partial access gracefully. Fewer high-confidence findings > many weak claims.

The Audit Process

Phase 1: Access and Inventory

Open all platform links. Confirm access. Build source inventory in notes: URLs, accessible reports, blocked items, date ranges, immediate red flags. Save before Phase 2.

Phase 2: Evidence Collection

Record what you see. Do not diagnose. Interpretive thoughts go under "Hypotheses to Test."

Minimum data points per platform (where visible): revenue/sales trends, spend/cost, volume metrics (sessions/clicks/impressions/orders), rate metrics (CVR/CTR/ACOS/ROAS), AOV, traffic sources/channel mix, campaign structure and status, budget allocation, product/SKU performance, customer composition (new vs returning, geo), funnel data, UX observations (website), copy issues, platform warnings/alerts.

Anomaly detection — actively watch for:

• Bot traffic signals: high sessions from data center cities (Council Bluffs IA, Ashburn VA, The Dalles OR) with zero conversions
• Cross-platform contradictions: ad platform purchase counts ≠ store backend
• Website copy inconsistencies: guarantee terms differing between sections, price discrepancies, broken domains
• Trust claims that conflict with observable evidence
• Amazon "All" tab showing 0 Purchases (check SP tab — see gotchas above)

Save notes after each platform before proceeding.

Phase 3: Evidence Summary

Organize findings across platforms: strongest/weakest performers, spend vs revenue concentration, trends, anomalies, missing data. What surprised you? Write under "Patterns and Hypotheses." Keep labeled as hypotheses.

Phase 4: Diagnosis

Determine root causes from Phase 3 patterns:

1. What are the most significant patterns? (Let evidence decide the count)
2. Most likely explanation for each? (Cite specific evidence, use OBSERVED/INFERENCE labels)
3. What must be true for each explanation to hold?
4. What can't be scaled until fixed?
5. Where do sources agree or disagree?

Every claim must cite supporting evidence.

Phase 5: Opportunity Mapping

For each opportunity: what could be done, supporting evidence, likely impact, confidence (high/med/low), dependencies.

Phase 6: Prioritization

Rank by: impact, confidence, speed, dependency risk.

• Tier 1 (7 days): High impact, high confidence, low dependency
• Tier 2 (30 days): Important, needs more setup
• Tier 3 (60-90 days): Strategic, depends on Tier 1/2

Phase 7: Generate Audit Report

Output a markdown file saved as {Client_Name}_Growth_Audit.md.

7-section structure:

1. Situation — 3-5 sentences: what the business is, what was audited, date range, 1-2 sentence verdict (what's working, what's broken, what's at stake)
2. Data Summary — One markdown table per platform. Columns: Metric | Value | Source | Verdict (Strong/OK/Weak/Critical). Include OBSERVED labels and show calculation formulas.
3. Diagnosis — Cross-platform connections. WHY the numbers look this way. Use INFERENCE labels. Address data source conflicts. Connected analysis, not a list.
4. Channel Verdicts — One subsection per channel. Verdict: Scale / Fix then Scale / Hold / Cut. Flexible length — 2-3 sentences if healthy, as long as needed if broken. Include specific campaign names and metrics.
5. Action Plan — Numbered, prioritized. Each gets [7d], [30d], or [90d]. Every action references specific data. No generic advice.
6. Open Questions — Data gaps, unverified assumptions, cross-platform conflicts, things to investigate.
7. Appendix: Raw Metrics — Compact tables by platform. No analysis, just numbers for implementation reference.

Every number cites its source. Write as an operator. Apply human voice protocol.

Phase 8: Report Verification

Cross-check every key metric against working notes. Verify names are accurate (not hallucinated). Confirm labels are correct. Check Open Questions covers all DATA NOT AVAILABLE items.

Spot-check 5-8 high-stakes data points against live platform tabs. Build comparison table (Report Says vs Live Shows vs Match?). Fix discrepancies. Show spot-check table when delivering.

Anti-Hallucination Checklist

Before finalizing:

• Every metric directly observed on a platform (not estimated)
• No "industry average" benchmarks unless user requested benchmarking
• No generalizing from a single campaign/product/day without stating limitation
• No causal claims from correlational evidence
• Cross-platform conflicts surfaced, not smoothed
• Diagnosis emerged from patterns, not pre-formed theories
• Working notes updated after every platform

Requirements

• Claude in Chrome extension — Required for platform data collection. Settings → Desktop app → Claude in Chrome.
• Computer use — Settings → Desktop app → Computer use.
• Platform access — Must be logged into all platforms in Chrome before starting.