Amazon Ads Audit

Use for account audits, account handovers, broad health reviews, or when the user wants the most important paid-media risks and opportunities across several domains.

Default mode: Suggest. Never mutate a live account.

Connector capability gate

When a conclusion depends on live MCP/API/connector data, load ../../references/connector-capability.md before interpreting missing or empty fields. If a machine-readable capability snapshot is available, identify the exact required capability IDs and run ../../scripts/evaluate_connector_capability_gate.py before metric interpretation.

• Pass — continue with the Skill's normal evidence, sufficiency, lineage and safety checks.
• Degraded — do not make a high-confidence dependent recommendation; keep the result to Directional, Hold, Alternate Source, Missing Data, or Manual Review.
• Blocked — do not treat the dependent observation as action-safe until the capability is resolved or an allowed alternate source is verified.
• missing_evidence_policy = never_zero — Partial, Unsupported, Unknown, or absent connector capability is never a numeric zero, unchanged state, or proof that the platform lacks the feature.

This gate is read-only and does not authorize live Amazon Ads mutation.

Progressive loading

Read this file first, then load only what is needed:

• full audit workflow and aggregation safety: references/account-audit-framework.md;
• metric definitions: ../../references/amazon-ads-metrics.md;
• cross-source/semantic/backfill comparability: ../../references/data-lineage.md;
• decision/execution boundaries: ../../references/decision-boundaries.md.

Do not load specialist Skills unless a material finding needs their decision logic.

Required context

Collect or mark missing:

• marketplace and profile/account scope;
• currency, timezone, exact date/comparison windows and attribution maturity;
• source freshness/completeness and truncation/coverage warnings;
• profile/account and campaign base metrics;
• keyword/target, search term and placement views when available;
• budget/bid/state/structure context;
• stated objective or target economics;
• retail/product context when conversion or scaling is material.

Workflow

1. Reliability gate — validate scope, dates, attribution, freshness and source comparability.
2. Canonical totals — choose one additive account grain and reconcile it to a complete campaign view when possible.
3. Metric integrity — sum compatible base metrics only when their aggregation_semantics is explicitly additive and source rows are disjoint; use ../../scripts/evaluate_metric_aggregation_gate.py when the semantic envelope is available. Never add overlapping entity grains, de-duplicated/non-additive metrics, or ratio KPIs; recompute ratios from compatible base components.
4. Account result — establish spend, attributed sales/orders, efficiency and business-objective fit.
5. Driver analysis — rank material contributors by absolute impact, then decompose through campaign/product/target/query/placement views.
6. Structure and routing — identify conflicting intents, Mixed-ASIN risk, unclear query routing, duplication, fragmentation or budget interdependence.
7. Waste/protection — identify evidence-backed waste without turning zero orders/high ACOS into automatic negatives or pauses.
8. Growth/headroom — identify qualified opportunities only after economics, retail readiness and marginal headroom checks.
9. Prioritize — blockers/risks first, then efficiency, growth and longer-term structure.
10. Route specialized findings to matching Skills instead of duplicating their rules.

Output contract

Return:

1. scope and data confidence;
2. executive verdict;
3. canonical account scorecard;
4. largest driver/contribution findings;
5. structure/routing and retail/economic findings;
6. prioritized issues/opportunities with evidence and confidence;
7. specialist action proposals when justified;
8. Hold / Need more data;
9. next review/measurement.

Safety boundaries

• Missing rows are not automatically zero.
• Profile/campaign/keyword/search-term/placement totals are not additive when they describe overlapping traffic.
• Account ratios must be recomputed from compatible aggregate components rather than averaged across rows.
• aggregation_semantics = unknown is not permission to sum. non_additive_deduplicated and ratio_or_derived metrics are never direct-sum safe; use a source-provided aggregate or recompute from compatible base metrics as appropriate.
• TACOS requires compatible total retail sales; profit claims require sufficient cost/economic inputs.
• Do not make high-confidence optimization decisions from unresolved marketplace/currency/date/attribution/source mismatches.
• Conversion-led issues must consider stock, Featured Offer / Buy Box, price, promotions, listing and variation-family changes.
• Default to Hold, Directional or specialist review when evidence is insufficient.
• Any real mutation requires explicit authorization and an external Connector / Executor.