I still hear the pitch: take your sales history, add your media spend, fit a trend line through the rest, and you have a Marketing Mix Model. It’s an appealing simplification, and it is also close to useless, because it ignores almost everything else that was actually moving the consumer. And what a model ignores doesn’t disappear. It gets handed to whatever is left – usually the media.
I once worked with a mobile retailer that ran a campaign built around a single headline offer: iPhone from £50 a month. Sales jumped. Drop that period into a model with no price or promotion variable, and every extra unit sold gets credited to whichever channels carried the message. The media plan looks spectacular. None of it is true. The offer sold the phones. The advertising told people about the offer.
Why the media gets the credit
A model can only credit what it can see. The offer and the campaign started in the same week and stopped in the same week. With no price term in the model, the media is the only thing that moved when sales did, so the model does the arithmetic and the media takes the lot. It isn’t a glitch: the more closely the two overlap, the bigger the inflation. Nor does it stay in the deck. Next year’s budget moves towards the channel, the offer isn’t there, and the sales don’t follow.
The rising tide
The same failure runs the other way too. A brand can grow while doing nothing especially well, simply because the category it sits in is growing, or a competitor has stumbled, or a macro trend is doing the work for everyone in the market at once. A rising tide carries all ships. Without a term in the model that separates the tide from the swimming, growth gets read as marketing effectiveness by default, and a brand that is coasting looks, on paper, exactly like a brand that is winning.
The warning signs
None of these proves a variable is missing, but each is a reason to ask:
- A return far better than the same channel achieves elsewhere, or in other categories.
- A channel whose effect spikes in exactly the weeks an offer was running.
- Base sales that stay flat while the category around them grows.
- Results that change sharply when someone adds a new variable.
One root cause
Both failures share the same root cause: a model built around the data that happened to be easy to collect, rather than the full set of things actually influencing the decision. Media is nearly always in the model, because it’s often the reason the project exists. Everything else is a fight to get there – and anything left out doesn’t have its effect ignored. That effect is handed to whatever variable moves with it, and that is often media, because campaigns are timed to coincide with offers and seasonal peaks.
It’s a common misconception that MMM stands for media mix modelling. It doesn’t. It’s Marketing Mix Modelling, and the distinction matters: the marketing mix is far broader than advertising. It takes in price, product, distribution and promotion, and a model should be built around everything that actually influences the decision, not just the part of it that was paid for in media.
The meeting before the data
The single most important meeting in any Marketing Mix Modelling exercise happens before a spreadsheet is opened. Its job is not to inventory the data you already hold. It is to work out everything that plausibly influences the consumer’s decision, whether or not you currently have a way to measure it, because the honest starting question is never “what data do we have?” It’s “what would change this person’s mind?”
In practice, the list tends to fall into four groups:
- The offer: price, promotions, discounts and finance deals.
- The shelf: distribution, availability, and whether the product was actually in stock.
- The market: category growth, seasonality, and what competitors were charging and spending.
- The wider world: the weather, the economy, consumer confidence and events in the news.
Not every factor matters for every business. But every one deserves a decision.
What you can’t measure
Some of the list won’t have data. That’s normal, and it’s manageable, provided it’s handled deliberately. A stand-in can work – search trends, for example, as a guide to interest in the category. A known event, such as a stock-out or a competitor’s launch, can be flagged in the model for the weeks it ran. And you can test how much the media results move when a factor is treated one way or another. What isn’t acceptable is leaving it out without saying so. Then its effect lands, quietly, on whatever’s left.
The list, not the model
One caution: everything on the list doesn’t mean everything in the model. Three years of weekly data can only support so many variables, and a model stuffed with them becomes unstable. Where price and media always move together, even a model that includes both will struggle to tell them apart. Sometimes the fix is in the plan rather than the model: run some media without an offer, so there’s variation to learn from. The point of the list is that every omission becomes a decision, not an accident.
The question to ask
Before you trust the next set of returns, ask your analyst: what else was influencing the consumer’s decision – and where is it in the model? A good answer names the factors, says how each was handled, and admits what’s missing.
A model missing half the picture will still hand you a number with three decimal places. The precision is real. The story it’s telling you almost certainly isn’t.
If your model has never had that meeting, drop a line to [email protected].