Marketing Mix Modeling and Incrementality Testing for Growing Brands

Attribution answers the wrong question

Every ad platform will tell you it drove your sales. Add up the conversions each one claims and you'll often get more than you actually sold. Click-based attribution counts who touched a customer last, or along the way, but it can't tell you what would have happened without the ad. A retargeting campaign shown to people already on their way to checkout looks brilliant in a dashboard and may add almost nothing.

The question that matters for budget decisions is incrementality: how many extra sales did this spend cause? Two methods answer it, and they work best together: marketing mix modeling (MMM) and controlled experiments.

Marketing mix modeling in brief

MMM is a statistical model of your sales over time. It uses aggregated weekly or daily data, so it needs no user-level tracking and isn't affected by cookie loss or consent rates. In simplified form:

sales_t = baseline + trend + seasonality
        + Σ  β_channel × saturation(adstock(spend_channel, t))
        + controls (price, promotions, holidays) + error

Two transformations make it realistic:

• Adstock captures carry-over. An ad seen this week still influences purchases next week. A decay rate per channel describes how quickly the effect fades; TV and video usually fade more slowly than search.
• Saturation captures diminishing returns. The first thousand spent on a channel works harder than the fiftieth thousand. A curve such as a Hill function models this.

The fitted model tells you each channel's contribution, its return on spend, and where on its saturation curve you currently sit, which is exactly what you need to reallocate budget.

Tools have made MMM accessible

MMM used to mean an expensive consultancy project once a year. Open-source libraries such as Google's Meridian, Meta's Robyn and PyMC-Marketing now let an analyst build one in-house. Bayesian approaches are popular because they let you encode prior knowledge (for example, "search returns are probably positive") and report uncertainty honestly instead of a single overconfident number.

What you need: around two years of weekly data if possible, spend by channel, outcome (sales or qualified leads), and the main non-marketing drivers like price changes, promotions, stock-outs and holidays. Missing an important control is the fastest way to get a misleading model, because the model will credit marketing for something else.

Experiments: the ground truth

Models infer; experiments measure. Three practical designs:

Geo holdout tests

Split regions into matched groups. Turn a channel off (or up) in the test regions for several weeks and compare against control regions. Synthetic-control methods build a weighted blend of control regions that tracked the test regions closely before the experiment, making the comparison sharper.

Platform conversion lift studies

Many ad platforms can randomly withhold ads from a share of your target audience and compare conversion rates. These are convenient but run inside the platform's walls, so validate the setup carefully.

On/off and budget-step tests

For smaller brands, pausing a channel for a fixed period, or stepping its budget up and down on a schedule, gives a rough but useful read when regional splits aren't possible.

Plan each test before you run it: what decision it informs, how long it needs to run to detect a meaningful effect, and what result would change your mind.

Calibrating the model with experiments

The modern approach links the two. Experiment results become priors or constraints in the MMM, so the model's channel estimates agree with what tests actually measured. The model then extends those findings to channels and periods you couldn't test. Each quarter, run a new test on the channel where the model is most uncertain or the spend is largest. Over time your model gets more trustworthy and your test budget is spent where it teaches the most.

A simple example

Adasify Retail spends across search, social, video and email. Its platform dashboards suggest retargeting has the best return. The MMM disagrees, showing retargeting near saturation and video under-invested. A four-week geo test confirms that pausing retargeting in matched regions barely moves sales. Budget moves from retargeting into video, and the next model refresh, now calibrated with the test, guides the following quarter's plan.

Where AI is changing this work

AI speeds up the slow parts: cleaning and joining spend data from many platforms, spotting anomalies such as a missing week of spend, and running thousands of budget scenarios to find an allocation that maximises expected return under constraints. Language models also make the output usable, translating posterior distributions into clear statements like "moving 10% of social into video is likely to raise sales, with a wide range of uncertainty". The judgement about what to test next stays with people.

Key takeaways

• Attribution shows who touched a sale; incrementality shows what caused it.
• MMM uses aggregated data, so it's robust to cookie and consent loss.
• Adstock and saturation are essential to realistic channel estimates.
• Geo tests and lift studies provide ground truth; use them to calibrate the model.
• Rerun the cycle every quarter: model, test where uncertain, recalibrate, reallocate.