Marketing mix modelling
Marketing mix modelling (MMM) is statistical analysis of your own sales and spend history that shows what each channel actually contributes, including the channels clicks can’t measure. Here’s how it works, when it’s worth doing, and how we build one you own.
-
2021
Southern Phone — cross-channel media partner since 2021.
-
$13m+
in ad spend managed across our client base.
-
98%
client retention.
What you get
The scope, in writing
-
A calibrated marketing mix model
Covering every channel you spend in, online and offline.
-
Channel-level contribution
How much revenue each channel actually drives, split from your baseline sales.
-
Response curves
Showing where each channel saturates, so you can see the point where the next dollar stops paying.
-
Budget-reallocation scenarios
What happens to revenue if you move spend between channels, before you move it.
-
Seasonality, promotion and pricing effects
Separated out, so media isn’t credited for a discount’s work.
-
A plain-English readout your board can act on
Not a statistics lecture.
-
The model itself, handed over
Code, data pipeline and documentation, with no ongoing licence owed to us or anyone else.
-
A refresh plan
Quarterly re-runs that keep the model honest as your mix changes.
How we do it
6 steps
Jordan Parrello, our Advertising Director, leads every model. We start by auditing what you have: sales or revenue at weekly grain, spend by channel, promotions, pricing changes. Two years of history is the working minimum. If your data can’t support a reliable model yet, we tell you that here and suggest what to fix first.
We build the modelling dataset from your GA4 property, ad platform exports, CRM and finance data, plus external factors (seasonality, public holidays, competitor moments where known). This is most of the work, and it’s where cheap models go wrong.
We fit a Bayesian mix model in an open-source framework (Google’s Meridian or Meta’s Robyn), with adstock (delayed ad effects) and saturation curves per channel. Open source matters: the method is inspectable and the model is yours.
We test the model against held-out weeks it hasn’t seen, and against ground truth where it exists — past geo tests, known outages, platform lift studies. A model that can’t predict the recent past doesn’t get to predict your budget.
You get the contribution and saturation picture, then scenario modelling: the recommended budget split, and the expected revenue change from making it. We present this to whoever holds the budget, in their language.
Quarterly re-runs with new data, calibrated against incrementality tests as you run them. This is where most of the value sits, once the model starts steering real budget decisions.
Case study
Southern Phone
Southern Phone
Southern Phone has trusted us with its media since 2021: paid search, social and more, planned and bought as one budget. That’s the situation mix modelling was built for — several channels competing for the same dollars, with reallocation decisions to make every quarter, and a client who wants to know what’s actually working rather than what last-click says.
Read the Southern Phone storyQuestions, answered straight
FAQ
Attribution is the method your analytics uses to decide which touchpoint gets credit for a conversion — last click, first click, or a data-driven split. It works from tracked user journeys, which makes it fast and granular but blind to anything it can’t track: TV, out-of-home, podcasts, and increasingly, cookie-blocked digital journeys.
Attribution follows individual users; MMM works top-down from aggregate data — weekly sales against weekly spend, per channel, over years. It needs no cookies or user tracking, so it covers offline channels and survives privacy changes. The trade-off is granularity: MMM answers “how should we split the budget?”, attribution answers “which ad drove that click?”. Mature advertisers run both and reconcile them.
The working minimum is around two years of weekly sales data plus channel-level spend, with meaningful investment across at least three or four channels. If your spend is concentrated in one or two channels, an MMM will tell you little you don’t already know; a geo holdout test is cheaper and sharper. We’ll say so in the audit, before you’ve committed to a build.
A first build is a scoped project fee, priced after the data audit. The cost driver is data readiness, not model complexity. Because we build on open-source frameworks, there’s no software licence stacked on top, and quarterly refreshes cost a fraction of the initial build. Compare that with MMM platforms, which charge an ongoing subscription for a model you never own.
Typically eight to twelve weeks from kickoff to board-ready readout, and most of that is data assembly and validation, not the model fit itself. If your sales and spend data is already clean and centralised, it’s faster. Refreshes are far quicker because the pipeline already exists.
You don’t need a platform. The frameworks most platforms are built on, Meridian and Robyn, are open source and free. A platform buys you a dashboard and a subscription; an agency build buys you a model tuned to your business that you own outright, plus the people who built it in the room when budget decisions get made. If you outgrow us, the model leaves with you.