Business problem

Marketing could not defend its budget split because last-click attribution credited channels that were merely closing demand.

Dataset

3 years of weekly spend, impressions and revenue across 6 channels and 42 geographies.

Methodology

• Modelled adstock and saturation curves per channel.
• Fitted a hierarchical Bayesian regression with priors from prior tests.
• Validated posterior ROI against geo-holdout results.

Key findings

• Paid search was over-credited by roughly 30%.
• Two upper-funnel channels showed genuine incremental lift.

Recommendations

• Move 18% of spend to the incremental channels at flat budget.
• Run a rolling geo-holdout calendar to keep priors fresh.