Marketing Mix Modeling

Your pixel says Meta drove 400 sales last month. Meta’s own dashboard claims 600. The bank statement lands somewhere between the two, and neither platform will tell you which channel actually caused the revenue.

Both Google and Meta now hand you a free tool that promises to settle this from the top down, using aggregate spend and sales rather than individual clicks. The two tools have stopped agreeing on almost everything, including whether they have a future.

Marketing mix modeling reads the account from the top. It correlates aggregate weekly spend, price, seasonality and outside factors against revenue, then attributes a share of sales to each channel.

Clicks never enter the model. That is why MMM keeps returning numbers through cookie loss and iOS opt-outs, the same conditions under which click attribution degrades.

The two free tools now point in opposite directions

Google’s Meridian runs on Bayesian causal inference, using a No-U-Turn Sampler to fit the model, per its open-source repository. Meta’s Robyn uses ridge regression instead, and its own repository still labels it “experimental”.

The methods diverged first. The commitment diverged after.

Google released Meridian in 2024 and has kept investing, to the point of tying internal sales KPIs to Meridian adoption, per July 2026 reporting. Meta open-sourced Robyn in 2020 and has since gone quiet on it, with two sources in that same reporting describing the engineering team as “dismantled.”

The sources do not fully agree here, and that gap is worth keeping. The reporting says Robyn is being wound down, while the package remains on Meta’s repository with no deprecation notice, which is the state a brand relying on it would actually encounter.

Whose media is doing the measuring

A tool that scores channels is built by a company that sells one of those channels. That is the conflict AdExchanger raised when it asked whether Meridian and Robyn are a gift to measurement or a trojan horse.

The specific charge against Meridian is that it over-indexes to Google’s own properties when it splits credit. Measurement vendors have also built proprietary MMM products on top of Meridian’s source code, per the same reporting, which means a brand can be buying “independent” measurement that runs on Google’s engine.

Google and Meta’s framing is the opposite: these are neutral, open-source frameworks handed to the industry for free. Both readings are defensible from the public record, and an entry that picks one is hiding the more useful fact that credible people disagree.

What a D2C brand needs before MMM says anything

MMM needs years, not weeks. Reliable output requires roughly one to three years of weekly aggregate data with real variation in spend, per an analysis of MMM’s structural limits — a history most sub-$50M D2C brands have not accumulated across every channel.

Two further constraints rarely appear in the vendor pitch. The feedback loop is slow, because a model reading weekly patterns cannot tell you how Monday’s new creative is doing until weeks of data accrue. And when channels are scaled up and down together, their effects correlate so tightly that no volume of extra data cleanly separates them.

MUD\WTR is the shape of the brand this applies to, splitting spend across Meta and Google in 122 countries. Its channels move in the same promotional rhythm, which is exactly the correlation MMM struggles to untangle.

This is why brands running MMM pair it with incrementality testing. The model proposes where the money worked; a holdout test checks whether turning a channel off actually changes the number.