The question.
Last-click attribution can reward channels that capture existing demand while missing marketing that creates it.
The research distinguishes observational model assumptions from experimentally identified effects. Budget scenarios include uncertainty and practical constraints.
The proposed approach.
- Align spend, outcomes, geography, and non-media controls
- Model lag and saturation effects with documented priors
- Calibrate suitable model components with experiment evidence
- Explore constrained budget scenarios without presenting point estimates as guarantees
Target architecture.
Spend and outcomesControls and experimentsBayesian mix modelDiagnosticsBudget scenarios
The difficult parts.
- Confounding and correlated channels
- Limited variation in historical spending
- Extrapolating beyond observed budgets
How to evaluate it.
- Posterior predictive checks and sensitivity to priors
- Agreement with held-out experiments where available
- Stability of scenario recommendations under uncertainty
Planned deliverables.
- A marketing data and experiment specification
- An assumption-aware mix-model study
- A scenario planning dashboard design
Design principles.
Attribution and incrementality answer different questions.
An optimization output is only as credible as its causal assumptions.