The challenge
A global gaming launch where "reach" was already a given. Budgets locked, channels booked. The real question was which dollars actually moved processors off shelves. Attribution in consumer hardware is a mess: purchases happen at Best Buy, on Amazon, or at a retailer we can't see. So we stopped trying to prove it per click and built a modeled answer instead.
What we built
We layered a predictive-bidding model over the programmatic stack that scored each impression for purchase propensity using first-party enthusiast audiences, game-ownership signal, and device profile. We then overlaid a weekly MMM against actual sell-through to calibrate it, so the model was learning against real revenue, not proxy conversions.
- Audience model: 14 first-party segments × 3 intent tiers × creative rotation
- Pacing: weekly budget redistribution based on modeled ROAS, not last-click
- Creator matching: picked influencers by audience-fit score instead of follower count. We paid less and converted more.
- Retail attribution: bridged Amazon Attribution + Best Buy co-op data into a single weekly revenue readout
What happened
By week four, the predictive model was outperforming platform-native bidding by 41% on modeled ROAS. By week eight, we'd shifted more than half the budget off "safe" channels onto higher-velocity creator + CTV combinations. Final blended ROAS landed at 6.1x, nearly double the historical benchmark for a hardware launch of this scale.
What stuck
The bigger outcome: the modeling framework became the template for the next three Intel launches. Same audience taxonomy. Same pacing logic. Same retail-attribution bridge. That's compounding.