Meridian
A production Bayesian marketing-mix model built on Google's open-source Meridian, fit on 137 weeks of real point-of-sale revenue across 8 brands, 12 media channels and 16 Chilean regions. Eight independent per-brand models, all converged, retrained weekly on GPU.
Attribution stopped working. Last-click credits whoever showed up at the end, view-through credits whoever showed up at all, and neither answers the only question a board actually asks: if we move a million pesos from out-of-home to Meta, what happens to revenue? Worse, most measurement quietly grades its own homework — fitting ad-platform conversion data against ad spend and calling the correlation a result.
A separate Bayesian MMM per brand, fit on real net point-of-sale revenue pulled from the warehouse — not platform-attributed conversions — weekly across 16 Chilean regions, with media spend re-weighted to each brand's actual store footprint. Every run reports R-hat convergence alongside its results, so a model that didn't converge says so rather than quietly reporting a number.
The Number That Changed Our Mind
Our first build was fitting ad-platform attributed revenue against ad spend — circular by construction. On that data, one brand's email ROI came out at 61x. Refit on real point-of-sale revenue, the same channel returned 2.9x. We rebuilt the pipeline rather than publish the 61. If a marketing model tells you a channel returns sixty times, the model is measuring itself.
One Model Per Brand, Not One For The Group
Eight brands, eight independent models, each with its own ROI priors and its own optimizer. A sneaker banner and a sporting-goods banner do not share a response curve, and averaging them produces a number that describes neither. All eight now converge at R-hat 1.00–1.04.
Fit On Revenue, Not On Pixels
The dependent variable is real net point-of-sale revenue by week, region and brand, straight from SAP through the warehouse — retail and ecommerce. National media spend is re-weighted across each brand's true regional footprint, so ROI denominators stay honest instead of silently dropping the spend that falls outside it.
Convergence Is Reported, Not Assumed
Every run writes R-hat, the share of badly-converged parameters and a converged flag alongside its results. That is how we caught a GPU precision setting silently diverging every scheduled training for months. The failure being visible is the difference between a model you can run a business on and a dashboard.
Managed marketing-mix engagements list at roughly US$25,000 per model. At eight brands that is about US$200,000 for a single run, before weekly refreshes — and multi-geo enterprise engagements are quoted at two to three times that. Three-year total cost of ownership for a bought solution is commonly put at US$1M–$2.7M. What a vendor will not do is tell you their last run was wrong.