The numbers the AI is allowed to trust

Precisely-defined product signals with sourcing and known limitations — every drift rule and confidence input in the AI Workflow traces back here.

Owner Product, with Data/Analytics · all five metrics are platform-native — credibility depends on matching what the marketer already sees.

Core metrics as product signals

ROASRevenue ÷ spend. North-star input; e-commerce drift target.

CPASpend ÷ conversions. Drift target for lead-gen.

CPLSpend ÷ leads. Secondary target, multi-stage funnels.

CPCSpend ÷ clicks. Leading indicator ahead of CPA/ROAS shifts.

CTRClicks ÷ impressions. Primary creative-fatigue signal.

Attribution — v1 decision

Platform-native, not a custom multi-touch model. Meta and Google's own conversion numbers are ingested as-is. Disagreeing 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 — disclosed in-product, not hidden.

North-star metric

Incremental ROAS improvement attributable to accepted AI suggestions. Measured against a trailing baseline per campaign, excluding campaigns with no accepted suggestions.

Guardrail metrics

LLM cost % of managed spend Suggestion acceptance rate Escalation rate