Marketing mix modeling in the AI era: Why expertise still matters

How marketing mix modeling has changed

Traditional MMM often meant commissioning a specialist team to collect data, build a model, interpret the results, and present recommendations. That could be a lengthy, expensive project. WARC’s coverage of MMM’s democratization describes its history as a high-cost tool associated with large advertising budgets.

Today, open-source packages such as Meta’s Robyn and Google’s Meridian make established modeling methods widely available. Automation handles parts of the work that previously demanded more manual effort. AI assistants can help investigate errors and guide parts of the modeling workflow.

These developments reinforce each other, but they do different jobs. Open source provides access to the code. Automation speeds up repeatable tasks. AI can help people navigate and execute more of the process. None of them removes the need to understand the business being modeled.

In a recent episode of AdExchanger’s Inside the Stack, M-Squared founder and CEO Madan Bharadwaj described the change in his own experience: work that once took months can now yield a first draft in days, with further work needed to reach a credible model.

The opportunity goes beyond receiving the same report sooner. Faster modeling makes it practical to explore more questions, compare different ways of representing a business, and investigate findings before acting on them.

Can marketing mix modeling really be free?

The software can be. Open-source tools such as Robyn and Meridian carry no software license fee for their open-source code.

You still need to prepare data, decide how the model should represent the business, check its findings, and turn them into a plan. Computing, maintenance, and the people doing that work also require resources.

If your team has the skills and capacity, building with open source can be a sensible choice. The question is whether someone has responsibility for the complete process, including what happens when the results look implausible or the business changes.

Forrester’s analysis of open-source MMM makes a related point: access to an algorithm doesn’t automatically make it understandable to marketing leaders. Interpretation, education, and support remain important to adoption.

A low software cost should make good measurement easier to access. It shouldn’t lower the standard for deciding what to trust.

The choices that shape your MMM recommendations

A model doesn’t arrive with a complete understanding of your customers, economics, or growth strategy. People make choices throughout its development that affect the findings.

As Madan put it in the podcast, when discussing which insights deserve confidence:

“That requires judgment, which is where the human beings come in.”

Here are some of the decisions where that judgment matters.

What business outcome are you trying to improve?

A model built to explain total revenue can lead to different recommendations from one focused on new customers or contribution margin.

Suppose a campaign brings in strong sales by promoting a heavily discounted product. A revenue-focused recommendation may look attractive while the finance team sees a weak contribution after product and fulfillment costs.

The outcome you choose needs to reflect the decision you’re making. Otherwise, the model can optimize toward a result your business doesn’t actually want.

What else could explain the sales?

Marketing operates alongside pricing, promotions, distribution, inventory, seasonality, and changes in consumer demand.

If you launch a campaign as a product enters more stores, sales may rise for both reasons. Leave out the distribution change and the model may assign too much of the increase to advertising.

An expert helps decide which influences belong in the analysis, whether the data can distinguish them, and where the answer remains uncertain. More variables won’t automatically produce a better model. The task is to represent the business well enough to support the decision.

What level of detail will help you act?

An answer for the whole brand may conceal differences between products or markets. At the same time, breaking the business into smaller pieces can leave too little data to estimate their effects reliably.

That creates a real judgment call. You need enough detail to make a useful investment choice, supported by enough information to trust the analysis.

How far can you take the recommendation?

A channel’s estimated average return doesn’t tell you how much more it can absorb profitably. Returns can change as spending rises, and a model has limited evidence about spending levels the business hasn’t tried.

Before scaling a recommendation, ask what the model knows about the proposed change, what assumptions it relies on, and whether a test would help resolve the uncertainty.

Why a beauty brand questioned its AI-assisted MMM

In the podcast, Madan described working with Jay Sung at Pacifica Beauty. Sung’s view was that individual products competed in different markets, so the decisions he needed to make required more than a single brand-level answer.

He had already used AI to run an open-source MMM. According to Madan, that initial model recommended broad cuts to Amazon advertising. Sung questioned the recommendation and brought in M-Squared to investigate.

The team developed models around product-level business structures and found meaningful differences in estimated media impact. The resulting investment recommendations varied by product.

The important lesson is the process. The initial output raised a question. Business knowledge challenged it. Further modeling investigated a structure that better matched the decisions Sung needed to make.

Madan described his starting point candidly:

“Of course, I don’t know the answer to his problem, but I know the process.”

An expert doesn’t need to know the result in advance. They need a rigorous way to work toward a recommendation the business can evaluate and act on.

Faster models make expert judgment more valuable

AI creates an opportunity to ask more of measurement. You can investigate alternatives that would once have taken too long to explore. You can revisit assumptions and examine differences that a broad model might miss.

It also becomes easier to generate an answer before understanding its limits.

When a model informs a multimillion-dollar marketing budget, a wrong recommendation can put millions in investment at risk. That doesn’t mean every imperfect model produces a major loss. It means the level of scrutiny should reflect the scale of the decision.

Before acting, you need to understand why a result appeared, how sensitive it is to reasonable modeling choices, and what other evidence supports it. A chart that looks convincing or a model that predicts sales well hasn’t automatically explained what caused those sales.

This is where incrementality testing can help challenge important assumptions. It can also be useful when two plausible model interpretations would lead to very different budget decisions.

EMARKETER’s 2026 MMM report identifies the gap between insight and action as an ongoing challenge. Generating more analysis only helps when the organization can interpret it and use it.

Where M-Squared adds value

At M-Squared, we apply marketing mix modeling within a measurement program built around the decisions your business needs to make.

That starts with learning the business. What do you believe drives growth? Where are you investing against that belief? What would you need to see to increase, reduce, or redirect the investment?

Our modeling process combines automated data preparation and model runs with established modeling practices and the specifics of your business. We examine different ways to account for your products, promotions, and financial outcomes, then compare the findings before recommending a change in spending. Expert judgment shapes both how the analysis is built and how its results are used.

We use our Marketing Accounting Framework to organize measurement around business outcomes and economics that marketing and finance can evaluate together.

Our Triangulation approach brings MMM findings together with experiments and other relevant evidence. When estimates disagree, understanding the conditions behind each result helps us decide what to recommend and what to investigate next.

Then comes the work of operationalizing the findings. Our Multiplier Engine helps apply measurement evidence to ongoing performance reporting, so teams can use estimated incremental impact alongside the platform metrics they work with every day. Calibration can raise or lower the credit a channel receives, helping teams evaluate where to increase, reduce, or redirect spending between studies. Those estimates need review as spending, seasonality, and business conditions change.

The Causal Insights Program connects that work around a learning agenda and planning calendar. The aim is to make each study useful to the next investment decision and build on what the business has already learned.