SMBs: when marketing mix modeling is worth it, and when it is not

Marketing mix modeling is coming back strongly, driven by open-source tools and by privacy constraints. It answers a real question of budget allocation. It also demands a history and a spending variation that most Quebec SMBs do not have, and that has to be said before committing to anything.

On this page

Why the method is coming back now

This article is for owners and marketing leads who have been offered a marketing mix modeling exercise, or who wonder whether their company is large enough to draw a decision from it.

The method has existed for decades. It was set aside when individual tracking seemed to explain everything, and it is coming back for two reasons.

The first is privacy. Google states that this approach relies on aggregated data and uses neither cookies nor user-level information. In a context where a share of visitors refuse to be tracked, a method that never needed that tracking regains value.

The second is accessibility. Google has released an open-source framework, Meridian, with its methodology and its code, which removes the licence-cost argument. A caution, though: free does not mean simple. The tool is open-source, the skill to use it is not, and that is where the real spending sits.

What you need before considering it

Three conditions, and all three must be met. A single one missing is enough to make the exercise decorative.

The second condition is worth pausing on, because it is counterintuitive. A disciplined company, one that spends steadily and without jolts, is harder to model than an irregular one. If you are seriously considering this exercise, deliberately introducing variation over the coming quarters is an investment in future measurement.

When it is not worth it

This is the part you will rarely be told, because it stands in the way of selling the engagement.

On a media budget of a few tens of thousands of dollars split across two channels, the allocation question does not justify a model. The answer lies in a simple experiment: cut one channel for eight weeks and measure the effect on your sales.

On a long sales cycle with few sales a year, the number of observations is too small. Fifty sales a year give fifty data points, and no serious model draws a conclusion from fifty points with ten variables.

For a company whose only significant channel is paid search on its own brand name, the question is not how to allocate a budget. It is whether that channel is incremental, which is a matter for a test and not a model.

A provider who offers marketing mix modeling to a company in one of these three cases is selling an impressive deliverable that will change no decision.

Calibration, without which nothing holds

A model estimates effects from correlations over time. Nothing guarantees that those correlations are causal.

That is why calibration through an experiment counts as much as the model itself. Google notes that its framework builds in calibration through incrementality experiments as one of its central innovations. In other words, good practice is to actually measure the effect of a channel through a test, then impose that result on the model as an anchor point.

An uncalibrated model can attribute a large effect to a channel that has none, simply because its spending followed the same curve as your seasonal sales. The test method is described in our article on incrementality tests.

The question to ask any provider is therefore direct: which experiments is your model calibrated on? If the answer is none, the deliverable is a well-presented hypothesis.

What we do instead

For most Quebec SMBs, three pieces of work produce more, for less, and in the right order.

First, connect real sales to the leads that produced them, to know the cost per sale by source. Many companies look for a model when they do not yet know this basic figure.

Next, one geographic experiment a year on the largest spending line. This gives a measured causal effect on what matters most, for the cost of a temporary deprivation.

Finally, introduce deliberate variation into the budgets, recording the periods. Three years of this discipline create exactly the history that will make modeling possible later, if your size ever justifies it.

How to choose a provider for this work

Three questions are enough to separate a serious offer from a decorative deliverable.

The first concerns calibration: which experiments will the model be anchored on? A provider who answers none is selling a well-presented correlation.

The second concerns uncertainty: will the deliverable show a range around each estimated contribution? A model that presents only points with no margin hides exactly the information management needs in order to decide.

The third concerns ownership: who owns the code, the prepared series and the documentation at the end of the engagement? Since the framework is open-source, nothing justifies your leaving without the model you paid for. It is a contract point, not a favour.

The advertising market shifts that move those trade-offs are quantified in the shift in digital advertising spend.

This point sits inside the plan described in the marketing plan and its budget.

Frequently asked questions about marketing mix modeling