attribution-ads-helper（归因广告助手）

Attribution Helper

Purpose

Core mission:

• Diagnose attribution discrepancies across channels.
• Compare attribution window assumptions and their budget impact.
• Build practical attribution decision framework for optimization.
• Produce actionable attribution-aligned allocation guidance.

When To Trigger

Use this skill when the user asks for:

• attribution model comparison
• conflicting ROAS/CAC by channel
• budget decisions under attribution uncertainty
• tracking and model interpretation support

High-signal keywords:

• attribution, tracking, model, predict
• roas, cpa, revenue, allocation, budget
• meta, googleads, tiktokads, youtubeads, dsp

Input Contract

Required:

• channel_metrics_by_window
• attribution_windows
• conversion_event_definitions
• decision_context

Optional:

• offline_conversion_data
• holdout_or_incrementality_data
• MMM_or_ltv_inputs
• confidence_threshold

Output Contract

1. Attribution Mismatch Map
2. Window Sensitivity Analysis
3. Decision-safe KPI View
4. Budget Reallocation Recommendation
5. Validation Experiment Plan

Workflow

1. Normalize event and conversion definitions.
2. Compare performance under each attribution window.
3. Quantify decision deltas from model differences.
4. Propose allocation with confidence labeling.
5. Output validation experiments for unresolved gaps.

Decision Rules

• If attribution views diverge materially, use blended guardrail plan.
• If one channel is highly view-through sensitive, reduce reliance on last-touch only.
• If incremental evidence exists, prioritize it over proxy metrics.
• If uncertainty remains high, allocate budget in capped test tranches.

Platform Notes

Primary scope:

• Meta (Facebook/Instagram), Google Ads, TikTok Ads, YouTube Ads, Amazon Ads, Shopify Ads, DSP/programmatic

Platform behavior guidance:

• Keep window comparisons explicit per channel.
• Separate platform-reported and unified-attribution decisions.

Constraints And Guardrails

• Never mix inconsistent conversion definitions in one conclusion.
• Flag time-lag effects for high-consideration products.
• Avoid binary conclusions when model variance is large.

Failure Handling And Escalation

• If event taxonomy is inconsistent, output normalization checklist first.
• If offline conversion pipeline is unavailable, mark blind spots and conservative policy.
• If budget decision is high-stakes, require experiment-backed confirmation.

Code Examples

Window Comparison Table

channel: Meta roas_1d_click: 1.9 roas_7d_click: 2.6 delta_pct: 36.8

Allocation Rule Under Uncertainty

if attribution_variance_pct > 25: budget_mode: guarded max_shift_pct: 10

Examples

Example 1: 1d vs 7d dispute

Input:

• Team split on attribution window

Output focus:

• sensitivity table
• decision-safe policy
• validation plan

Example 2: Channel reallocation decision

Input:

• Meta and Google show conflicting contribution

Output focus:

• mismatch diagnosis
• allocation options
• risk labels

Example 3: Incrementality integration

Input:

• Holdout test data available

Output focus:

• model reconciliation
• updated budget recommendation
• confidence update

Quality Checklist

• [ ] Required sections are complete and non-empty
• [ ] Trigger keywords include at least 3 registry terms
• [ ] Input and output contracts are operationally testable
• [ ] Workflow and decision rules are capability-specific
• [ ] Platform references are explicit and concrete
• [ ] At least 3 practical examples are included