Placement Optimization

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.

Campaign objective gate

Before converting placement efficiency into an increase/decrease modifier proposal, load ../../references/campaign-objective.md and preserve campaign_objective, primary_metric, and guardrail_metrics whenever campaign mission changes the interpretation or acceptable trade-off.

• Prefer explicit user/account strategy, structured campaign metadata, or a documented operating plan as objective evidence.
• Do not infer campaign objective from ACOS, TACOS, campaign name, spend level, placement mix, bidding strategy, organic/rank movement, or a short performance window.
• Growth still requires marginal placement headroom plus retail/economic readiness; high historical placement ROAS is not proof that a larger modifier is incremental.
• Profit is constrained by contribution economics and marginal efficiency, not by a universal placement ACOS rule.
• Discovery may accept controlled learning cost, but exploratory traffic is not automatically a placement-scaling candidate.
• Defense must expose strategic rationale, cannibalization risk, and opportunity cost before paying more for protected traffic.
• Experiment preserves treatment/control integrity and declared primary/guardrail metrics; do not contaminate a placement experiment with opportunistic modifier changes.
• Control favors stable, interpretable traffic ownership and should not be amplified from one efficient placement window alone.
• Unknown stays unresolved. If modifier direction, tolerance, or sizing depends on mission, cap the result at Hold, Directional, Shadow, Experiment, or Manual Review; do not emit an action-safe exact modifier solely from ambiguous objective evidence.

This gate is additive to connector capability, realized-exposure, control-interaction, confounder, action-sizing and platform-capability checks. It does not authorize live writes.

默认模式

Suggest。输出 placement modifier 候选，不直接修改 campaign。

输入

需要尽量包含：

• placement name；
• impressions、clicks、spend、orders、sales；
• CPC、CVR、ACOS/ROAS；
• current base/target bid；
• current placement modifier (Top of Search, Rest of Search, Product Pages, and separately Amazon Business placement when source-supported and eligible)；
• campaign bidding strategy；
• current audience bid adjustment state when the campaign/surface supports it；for registered Sponsored Products bid/budget causal comparisons, complete audience lists still require unique nonblank audience IDs and valid finite nonnegative per-member bid boosts (observed zero is allowed); invalid/duplicate members mean Unknown, not an unchanged audience control；
• current Sponsored Products video bid adjustment state when applicable；
• active schedule/event bid rules or other material modifiers；
• material control-change timestamps；
• 日期窗口和目标。

当 recommendation 依赖 base bid、placement modifier、dynamic/rule-based bidding、audience bid adjustment、schedule/event rules 等控制的交互时，按需加载：

references/realized-bid-exposure.md

当 recommendation 依赖 Amazon 当前 placement adjustment 范围、dynamic bidding 的精确平台行为、audience/video/other bid boost capability、rule eligibility，或 console/API 是否支持某项控制时，再加载：

../../references/platform-capability-lineage.md

平台规则发生版本/来源冲突时，先完成 capability lineage reconciliation；不要把旧 guide、launch announcement、不同 ad product 或不同 control surface 的数值直接当成当前 action-safe rule。

Amazon Business placement control boundary

Amazon Ads documents a distinct Sponsored Products Amazon Business placements bid adjustment. When supported by the active campaign/marketplace and connector, collect its configured value and observation/effective time separately from the three ordinary placement modifiers. Do not treat missing/unsupported B2B evidence as 0 or as proof that the control is ineligible. When explicit B2B values are supplied to the read-only control comparator, invalid values or one-sided missing evidence must yield Unknown; overlapping observed changes are confounders. Legacy snapshots that omit the field remain compatible but cannot independently establish B2B-complete control coverage. Across-retailers Sponsored Products campaigns have different placement eligibility, so verify dated capability scope before an action-safe conclusion. Amazon Business exclusive-campaign site restriction is a separate routing control, not the same modifier.

Source: Amazon Ads, Adjust Sponsored Products bids (2026-09-28), https://advertising.amazon.com/help/GYYZVM7LGSRYGWV5 . No universal stacking formula or modifier recommendation is inferred.

For registered Sponsored Products bid/budget causal reviews, an explicitly supplied normalized targeting_or_routing.state.retailer_scope must be AMAZON_STORE or ACROSS_RETAILERS on both snapshots. An across-retailers scope contradicts an Amazon Business placement modifier (including 0) or Business-exclusive site routing; return Unknown rather than inferring an eligible, absent or unchanged B2B control. An actual verified retailer-scope change is a confounder. Both-absent legacy scope remains compatible without B2B eligibility proof; verify marketplace/campaign/connector capability independently. These are local read-side labels, not Amazon-native API enums.

分析思路

1. 看效率，不只看销售额

高销售额版位可能只是获得更多流量。比较：

• spend share vs sales share；
• CPC；
• CVR；
• ACOS/ROAS；
• 样本量；
• query/target/product/audience mix 是否可比。

2. 区分配置控制与 realized exposure

Base/target bid、placement adjustment、campaign bidding strategy、audience bid adjustment、video bid adjustment、schedule/event rule 都是配置控制；placement report 中的实际流量、CPC、CVR 和效率是 realized outcome。

不要从配置值自行发明一个精确 auction-level effective bid。控制组合与 Amazon 的 auction-time logic 共同影响实际 exposure，应该通过控制时间线 + placement delivery/CPC 来诊断。

当“Amazon 会如何自动提高/降低 bid、某 placement 最大 adjustment、某 audience control 是否可用、某 control 是否在当前 surface/marketplace 可用”等事实影响推理时，先按 ../../references/platform-capability-lineage.md 标记 capability scope 和 conflict status。Conflicted / Unknown 的 exact platform rule 不能支撑精确 effective-bid 推断或 modifier 建议。

3. 看增量潜力

高效 Top of Search 不代表应该立即大幅提高 modifier。需要判断：

• 当前 impression/click 规模；
• base bid 是否本身偏低/偏高；
• modifier 提升可能带来的边际 CPC；
• dynamic/rule-based bidding、audience bid adjustment、video bid adjustment 或 bid rules 是否已经改变 exposure；
• 目标是否偏增长还是利润；
• 是否存在可解释的 marginal headroom，而不只是历史平均 ROAS/ACOS 好看。

4. 识别放大与归因风险

Base bid、dynamic/rule-based bidding、placement modifier、audience bid adjustment、video bid adjustment、schedule/event rules 可能共同影响实际竞价。不要把 modifier 当作独立旋钮。

如果多个 material controls 在同一 attribution/measurement window 内变化，placement 结果只能 Directional / Confounded，除非有额外设计能分离影响。尤其不要把 audience bid adjustment 引起的 traffic/audience mix 变化归因为 placement modifier 的独立效果。

常见状态

High-efficiency / underexposed

可考虑提高 modifier、调整 base bid 或设计实验，但需判断哪一个控制最可解释，并确认没有重叠的 pending control change。

High-efficiency / already dominant

继续扩量可能边际效率下降，先观察，不追涨。

Low-efficiency / high-spend

检查 query mix、CVR、retail state 和 coupled controls 后，再判断降低 modifier、base bid 或其他控制。

Low-volume

样本不足，不做强结论。

Coupled / confounded

多个竞价控制或规则同时变化，或当前 control state 不完整。优先 Hold / Shadow / Experiment，不要输出假精度 modifier。

变更原则

• 一次只做可归因的单一/最小控制变化，除非实验明确设计组合 treatment；
• 若同时需要改 base bid 和 placement，优先分阶段；
• 已有 base bid / placement / bidding-strategy / audience bid adjustment / video bid adjustment / bid-rule 变化 pending evaluation 时，不叠加同一路由上的另一个 material change；
• 数值幅度需要账户策略、校准响应或明确实验约束；必要时加载 ../../references/action-sizing.md；
• 平台允许范围/自动 multiplier 与 advertiser-specific sizing 分开；前者按 ../../references/platform-capability-lineage.md 验证，后者按 action-sizing 决定；
• 设验证窗口，避免短期反向调整；
• 大促、event rule active window、零售状态变化需要单独解释。

输出

每个 placement：

• current modifier；
• current base/target bid；
• bidding strategy / audience bid adjustment / video bid adjustment / active bid-rule state；
• control state as-of / material change timeline；
• platform capability scope / conflict status（当平台规则影响结论时）；
• spend share / sales share；
• CPC/CVR/ACOS；
• baseline comparison；
• interaction risk；
• causal attribution: Supported / Directional / Confounded / Unknown；
• recommendation: increase/decrease/hold/investigate/experiment；
• recommended control；
• proposed modifier（仅证据和 sizing basis 充分时）；
• confidence；
• evidence；
• guardrails；
• validation window；
• rollback condition。

禁止

• 不因为 Top of Search CVR/ROAS 高就默认提高 modifier；
• 不忽略 dynamic/rule-based bidding、audience bid adjustment、video bid adjustment、schedule/event rules 或其他 material bidding controls；
• 不把配置值拼成未经平台行为证实的精确 effective-bid 公式；
• 不把历史 placement 平均效率当成增加 modifier 后的保证边际效率；
• 不在 placement 样本很小的时候做大幅动作；
• 不同时大改多个竞价层造成无法归因；
• 不把未完成 capability lineage reconciliation 的旧/冲突 Amazon 文档数值写成当前确定的平台上限或自动调整公式；
• 不在 material control state 缺失/过期时输出 action-safe 精确 modifier。
• Sponsored Products 三个必需 placement modifier 的 Unknown / Unsupported / Partial、布尔值、负值或非有限数不是已观测稳定控制；即使前后相同，也不得据此声称单一 base bid 或 budget 变化可归因。明确观测到的数值 0 不等于缺失；上限与适用市场另查平台能力证据。

Sponsored Products site-routing evidence in causal reviews

For registered Sponsored Products bid/budget comparisons, do not infer unchanged delivery routing from a complete control inventory alone. targeting_or_routing.state.site_restriction must be a source-supported recognized repository-normalized label (ALL_ELIGIBLE or AMAZON_BUSINESS); repeated Unknown/Unsupported/Partial, absent, malformed or unmapped site modes return Unknown, not zero or stable routing. A verified switch to Amazon Business-exclusive routing is a competing control, distinct from an Amazon Business placement bid modifier. Consult ../../references/control-state-comparability.md and dated platform/connector capability evidence before any B2B interpretation. These labels do not certify campaign eligibility or actual delivery, and this Skill remains Suggest-only.