In 2015, Mastercard paid $600 million for Applied Predictive Technologies. Retailers did not pay APT for software. They paid for a design discipline. Cluster the estate, draw stratified samples from each cluster, run every test cell independently, pool the reads, and price the effect down to the individual store.
I learned that method from APT years ago. It is still the cleanest way I know to settle a marketing argument. For most of its life it belonged to companies that could pay for a platform, a statistician and a consultant to explain the output.
That is over.
The tools went free
Meta open-sourced Robyn, a full marketing mix modeling package. Its own README says MMM “used to be a resource-intensive technique that was only affordable for ‘big players.’” Google followed with Meridian, its open-source MMM, opened to everyone in early 2025.
In September, Google made Meridian GeoX generally available. It is an open-source library for geo experiments, the same matched-market logic APT built a business on, and it feeds test results straight into the model.
The core method behind a $600 million acquisition now ships free.
The code stopped being the barrier
Free libraries never closed the gap. The scarce part was the bench. Someone who could run a power analysis before launch, confirm the clusters balance on the pre-period, fit a random-effects model across cells, and tell a real location-level effect from noise.
AI compresses that bench. It writes the power calculation, the balance diagnostics and the hierarchical model, and states every assumption it made along the way. You still have to know which question is worth the markets. That part does not automate.
I hit this last month. I wanted a conjoint study on whether a CFO would approve a marketing budget. The off-the-shelf platform cost more than the study was worth. So I am building it myself in Python with Claude.
Your data stays home
The fair objection is privacy. Nobody wants customer records pasted into a chatbot.
You do not have to. Have the AI build and test the code on synthetic data shaped exactly like yours. Same columns, same formats, fake rows. Then run the finished code on the real data, on your own machine. The model never sees a customer.
That is the rule I hold for every client.
What the design buys
A pooled read you can defend. Each cell is measured against its own matched control. A random-effects meta-analysis pools the cells, weights each by its precision, and reports how much the effect varies between them. One noisy market cannot swing the verdict.
The effect where it lives. A flat pooled lift is rarely flat everywhere. Estimate lift per location, shrink each estimate toward its cluster so small markets cannot manufacture winners, then regress lift on location attributes. The test that read zero in total comes back as a targeting rule. These markets pay back, those do not, and here is what separates them.
How channels lean on each other. Run the cells as a factorial. Cut paid social in one arm, cut search in another, cut both in a third. The interaction term shows whether search was riding on social. That is the answer that holds up across the table from a CFO.
Priors for the model. Feed the experiment estimates into the MMM as calibrated priors. The model stops guessing at channels the tests already settled and points to the channel worth testing next.
It also ends the argument. Every marketing leader carries a plan they believe in. So does the CEO. When those plans disagree, the old tiebreaker was seniority. A test is a better tiebreaker. Run the plan into the real world and it either proves out or fails. Either answer is worth having.
The annual consultant MMM never solved this. It looked backward, cost serious money and broke the moment the market shifted. The new stack flips the order. Run experiments first. Let the experiments calibrate the model. Keep testing.
What the plan locks before launch
The decision threshold. The lift that would change the budget, stated as return in dollars.
Power. The minimum detectable effect given your market count and pre-period variance. If the design cannot see the threshold, it does not run.
Strata and cells. Clusters built on pre-period sales, trend and seasonality, then stratified random assignment inside each one.
Pre-period fit. Test and control tracked each other before launch, or the read is void.
The read rule. Pooled effect, per-location effects with shrinkage, and interaction terms, all written down before anyone sees a result.
The bar moved
Big companies had measurement. Everyone else ran on dashboards. That split came down to money and talent, and both constraints are gone.
The tools are free. The code is a prompt away. The data never leaves your building. Marketing science is now the standard for any team that has to defend a budget.
Bring the plan you already believe in. Prove it or fail it.
I publish one free Claude skill every month for exactly this work. The marketing mix model and customer segmentation skills are already in the repo. [REPO LINK]
Until next month, keep your CRM honest,
Bill
Marketing Systems Guild. Build it. Prove it. Scale it.

