Performance Drop Diagnosis

Find what changed, when it changed, where the loss came from, and which causes are supported by evidence. ACOS/ROAS movement is an outcome, not a root cause.

Default mode: Read-only or 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. Load only what is needed:

• Detailed causal workflow and Mixed-ASIN safety: references/causal-drop-diagnosis.md
• Platform-managed surface/product/creative realization: ../../references/realized-ad-identity.md
• Cross-source, freshness, attribution-variant, modeled/direct conversion, allocation-coverage, or metric-version reconciliation: ../../references/data-lineage.md
• Marketing Stream v1/v2 dataset-generation, total/delta and version-reconciliation semantics (when stream data drives a suspected performance change): ../../references/marketing-stream-dataset-coverage.md
• Metric definitions: ../../references/amazon-ads-metrics.md
• Shared decision boundaries: ../../references/decision-boundaries.md
• General optimization logic: ../../references/optimization-framework.md

Do not load data-lineage.md for a single stable measurement path whose source, completeness, attribution variant and metric definitions are already known to be comparable. Load realized-ad-identity.md only when the ad product can change delivery surface, product mix, creative/message realization, or another shopper-visible experience independently of ordinary manual campaign edits.

Required context

Collect or explicitly mark missing:

• marketplace, profile/account scope, currency, timezone, date window, attribution maturity;
• attribution model/window/variant when conversion reporting can expose multiple methodologies such as standard/default vs all views;
• modeled-versus-direct conversion semantics and allocation coverage when the ad product/report can combine modeled and directly measured conversions; at a lower reporting grain, preserve any Unallocated/sentinel dimension rows rather than treating them as zero or dropping them;
• cross-ad-product activity when diagnosing attributed conversion movement: eligible Sponsored Products, Sponsored Brands, Sponsored Display and DSP activity can compete for conversion credit, so a product/campaign attribution drop can occur without an equivalent shopper-demand drop;
• primary KPI and business objective;
• daily or weekly account/campaign performance;
• source/metric metadata when windows use different reports, refresh paths, attribution variants, semantic versions, modeled/direct measurement composition, allocation coverage, or unknown measurement definitions;
• advertised ASIN, purchased ASIN, parent/variation-family mapping where relevant;
• campaign/ad group/keyword/target/search-term/placement data;
• budget, bid, state, negative-targeting and structure-change history;
• platform-managed delivery-surface / ad-experience changes when the marketplace may auto-enroll existing campaigns into a new format or surface;
• realized product/creative/message mix when the selected ad format can dynamically choose among eligible products or shopper-facing realizations;
• retail context when available: stock, Featured Offer/Buy Box, price, coupon/deal, listing suppression, reviews, delivery promise, variation-family changes;
• timestamp/freshness for retail snapshots;
• total retail sales when making TACOS or organic-momentum claims.

Workflow

1. Reliability gate — reconcile date windows, attribution model/variant, attribution lag, completeness, modeled/direct conversion semantics, allocation coverage and measurement comparability; a source, attribution-methodology, allocation-methodology or metric-definition cutover near the break is a competing cause. For Marketing Stream, distinguish v1 deltas from v2 total-value records: verify v2 full-record identity and highest streamBatch.version selection before accepting an apparent summed spike/drop.
2. Find the break point — identify the first sustained KPI change and classify it as abrupt, gradual, intermittent, or isolated.
3. Size the loss — compare matched windows; normalize unequal windows only after the data are comparable. Preserve Unallocated conversion rows in campaign/order-level totals when the source defines them as attributable but not assignable to the requested dimension.
4. Decompose the bridge — impressions -> clicks -> CPC/spend -> orders/CVR -> sales/AOV -> ACOS/ROAS. Keep campaign-level conversion movement separate from lower-grain allocation movement.
5. Rank contributors — prioritize ASINs, campaigns, targets, search terms and placements by lost business contribution, not noisy percentages. Do not rank an allocated lower-grain slice as complete when material conversions remain Unallocated.
6. Check attribution competition — when attributed purchases/sales fall while traffic is comparatively stable, check whether activity changed in other eligible Amazon ad products before treating the measured conversion drop as a shopper-conversion failure. Keep the observed credit shift separate from the hypothesis about why it shifted.
7. Check retail/market confounders — inventory, Buy Box, price, promotion, listing, reviews, delivery, demand, competitor and parent/variation-family changes.
8. Audit controllable and platform-managed changes — bids, budgets, placements, states, negatives, product-ad mapping, launches, pauses, automation or bulk edits, plus auto-enrolled delivery surfaces/ad experiences and dynamic product/creative realization that can change traffic or conversion mix without a manual campaign edit.
9. Run Mixed-ASIN safety — distinguish clean routes from halo-heavy or mixed-ASIN routes before target-level actions.
10. Assign causality — Confirmed, Likely, Directional, Rejected, or Missing Data.
11. Propose recovery — only actions that pass the actionability gate, with validation and rollback criteria.

Output contract

Return:

1. data reliability, attribution/measurement comparability, modeled/direct and allocation coverage when relevant, and actionability;
2. executive verdict and exact windows;
3. KPI bridge and primary driver;
4. ranked ASIN/campaign/target contribution, explicitly labeling incomplete lower-grain allocation;
5. attribution-competition, retail and control-change findings, including family-level and platform-managed realization effects when relevant;
6. Mixed-ASIN safety labels;
7. facts vs hypotheses vs missing data;
8. prioritized recovery proposals with confidence;
9. monitoring plan appropriate to attribution/sample maturity.

Safety boundaries

• Do not infer causality from correlation alone.
• Same source/table/metric name does not prove comparability when attribution model/variant, modeled/direct composition, allocation coverage, completeness, filters, grain or metric semantic version changed.
• Standard/default conversion metrics and all views conversion metrics are not interchangeable merely because they share labels such as Purchases, Sales or ROAS.
• Modeled conversions combined into ordinary conversion columns are not equivalent to directly measured-only conversions unless the source semantics establish comparability.
• An Unallocated or sentinel dimension row is not zero and is not disposable noise when the reporting contract says it carries conversions attributable to the parent campaign/order but not allocable to the requested lower grain. Preserve it in valid parent-level totals; never redistribute it across targets/search terms/placements without supported allocation evidence.
• If material conversion volume is Unallocated, lower-grain conversion rate, ROAS, waste, winner/loser and contribution conclusions are incomplete. Downgrade target-level causal claims and traffic-suppression actions to Directional, Hold, or Missing Data until allocation coverage is reconciled.
• An attribution-methodology cutover near a performance break is a competing measurement cause until reconciled; do not generalize a documented view-attribution change to unaffected campaign/inventory scopes.
• A decline in attributed purchases/sales for one eligible ad product does not by itself prove shopper conversion propensity fell: cross-ad-product attribution competition can reassign credit after activity changes elsewhere in the account.
• Do not recommend aggressive bid cuts, pauses or traffic suppression solely from a product-level attributed-conversion drop until material cross-product attribution competition has been checked or explicitly marked unavailable.
• No manual control change does not prove that delivery conditions stayed constant when Amazon can auto-enroll existing campaigns into a platform-managed delivery surface or dynamically realize different products/creative experiences.
• Configured eligible products are not the same thing as the realized product mix when a format supports platform-managed selection.
• Missing realized product/creative/surface fields from the active connector are not evidence that realization was unchanged or zero.
• Do not recommend bid cuts, pauses or negatives solely because a row has poor ACOS.
• Do not call a target waste without sufficient evidence and compatible attribution maturity.
• If purchased-ASIN attribution or Mixed-ASIN scope is unclear, downgrade target-level actions to Directional or Blocked.
• A change made after the decline began is a possible fix, not a root cause.
• A stale retail snapshot cannot prove Buy Box, inventory, price or listing health during a later decline window.
• A child-ASIN conversion decline is not isolated ad inefficiency when parent/variation-family or sibling retail changes plausibly explain substitution.
• When missing/stale retail state, family-level retail shock, source-lineage drift, attribution-variant drift, modeled/allocation drift, metric-version drift, cross-ad-product attribution competition, or a platform-managed realization change can materially explain the break, downgrade aggressive traffic suppression until reconciled.
• Every recommendation must state evidence, confidence, expected effect, validation window and rollback trigger before it can ever reach an external executor.