Metrics & Attribution Spec
This doc defines the performance marketing metrics Ad Copilot depends on — not as marketing vocabulary, but as precisely-defined product signals with sourcing, calculation, latency, and known limitations. Every drift-detection rule and confidence input in the AI Workflow Design traces back here.
1. Why this doc exists
The AI Workflow Design references “target CPA,” “signal strength,” and “data volume.” None of that is trustworthy unless the underlying metrics are defined precisely — performance marketers stop trusting the product the first time its numbers don’t reconcile with Ads Manager.
2. Core metrics as product signals
| Metric | Definition | Source | Used for |
|---|---|---|---|
| ROAS | Attributed revenue ÷ ad spend | Platform-native | Primary drift target for e-commerce; north-star input |
| CPA | Ad spend ÷ attributed conversions | Platform-native | Primary drift target for lead-gen |
| CPL | Ad spend ÷ attributed leads | Platform-native (mapped from account’s lead event) | Secondary target for multi-stage funnels |
| CPC | Ad spend ÷ clicks | Platform-native | Leading indicator ahead of CPA/ROAS shifts |
| CTR | Clicks ÷ impressions | Platform-native | Primary creative-fatigue signal |
All five metrics are sourced from the ad platforms’ own reporting — consistent with the PRD non-goal of building a custom attribution engine. Credibility depends on matching what the marketer already sees natively.
3. Attribution approach
v1 decision: platform-native attribution, not a custom multi-touch model.
- Meta and Google each report conversions with their own windows/models; Ad Copilot ingests these as-is.
- No attempt to reconcile or override with an independent model.
- Why: disagreement with the platform dashboard is an instant credibility problem and roughly doubles engineering scope for a 0-to-1 build.
Known limitation: cross-platform journeys can double-count. v1 does not de-duplicate across Meta and Google; this is disclosed in-product via a “platform-reported, not deduplicated” tooltip.
4. Conversion tracking (product-level)
- Relies on the advertiser’s existing setup (Meta Pixel/CAPI, Google conversion tracking) — no new tracking requirement.
- Data freshness: pulled hourly (NFR-1); platforms may still attribute retroactively — recent 24–48h weighted less heavily when calculating trend stability.
Data-quality risks surfaced, not hidden
- iOS / ATT — modeled vs observed conversions flagged where the platform provides it.
- View-through vs click-through — reported separately; marketer chooses whether view-through counts toward target.
- Provisional/sampled API data — recommendations on provisional windows are annotated as such in reasoning.
5. How metrics feed the AI system
- Drift detection — live ROAS or CPA vs target over a rolling window generates recommendations.
- Signal strength — magnitude + stability of the deviation.
- Data volume — conversion counts / lookback length drive the <100 conversions / <7 days conservative rule.
- Creative fatigue — CTR trend decline on a stable audience triggers scheduled refresh (US-4).
6. North-star metric (product-level)
Incremental ROAS improvement attributable to accepted AI suggestions. Compare campaign performance after acceptance against a trailing baseline for that same campaign (not a control group in v1). Excludes campaigns with no accepted suggestions to avoid diluting the signal.
7. Guardrail metrics
- LLM cost as % of managed ad spend — hard ceiling; rising trend is a product-health alarm.
- Suggestion acceptance rate — trust leading indicator; reviewed alongside incremental ROAS to catch “trivially easy suggestions.”
- Escalation rate — should decrease as trust tiers mature; sudden drops investigated as threshold miscalibration.
8. Explicit non-goals
- No custom attribution model.
- No incrementality/lift testing methodology in v1.
- No cross-platform conversion deduplication (disclosed in-product).
More in this series