Why marketing mix modeling is still hard to get right

Advertising combine modeling (MMM) is changing into extra accessible, however getting began stays a problem.

After a number of conversations about MMM adoption, I observed the identical query saved developing: “We consider within the idea of MMM, however we don’t know find out how to get began.”

The reply is that viable open-source platforms have dramatically lowered the barrier to entry. They haven’t lowered the extent of experience required to provide reliable, actionable outcomes.

Open-source MMM has modified the start line

MMM adoption is accelerating. Almost half (46.9%) of U.S. marketers will make investments extra in MMM over the following 12 months, they usually ranked MMM as probably the most dependable measurement methodology (27.6%).

The open-source revolution in MMM is actual. Three production-grade libraries now cowl the total methodological spectrum:

• Robyn (Meta, R): Automated hyperparameter search through Nevergrad, Pareto frontier mannequin choice, and built-in decomposition and response curve plots — probably the most approachable entry level. It’s the one I take advantage of most as a result of it’s extremely customizable.
• Meridian (Google, Python/TensorFlow): Bayesian inference with geo-level priors and principled uncertainty quantification — extra rigorous, with a steeper studying curve.
• PyMC-Advertising (PyMC Labs, Python): Essentially the most versatile possibility, providing a full probabilistic mannequin that’s closest to academic-grade Bayesian MMM — but it surely additionally requires probably the most statistical fluency.

This technology of instruments has eradicated the $150,000 to $500,000 consulting gate that was once the one path into MMM. Any workforce with R or Python experience and comparatively clear historic information can now run a mannequin in-house.

The important thing caveat value making specific in any dialog with these exploring MMM is that this: “Free instrument” doesn’t imply “free mannequin.” The software program is free. The area experience required to configure it appropriately — a vastly necessary a part of the method — isn’t.

A crowded vendor panorama with an fascinating energy dynamic

The SaaS layer constructed on prime of open-source MMM has proliferated shortly. It’s value distinguishing a couple of tiers.

Knowledge-layer-first distributors

Platforms like Rockerbox and Northbeam began as attribution and information assortment platforms, then added MMM. Their edge is information pipelines and pace, not modeling depth or customization.

Measurement-first distributors

Platforms like Measured, Analytic Companions, Ekimetrics, and Nielsen Gracenote supply extra rigorous modeling at a better value level, with enterprise-grade capabilities.

Google Meridian and GA360

One level is value calling out. Google’s open-sourcing of Meridian was a beneficiant contribution to the sector and, on the identical time, a strategic one. When a walled backyard funds and packages the measurement methodology used to guage its personal channels, it’s value sustaining wholesome skepticism about mannequin priors and default assumptions, even with clear code.

The sensible query when evaluating distributors is: who owns your information layer, and does that create conflicts within the modeling layer?

Problem 1: Knowledge entry is the silent MMM killer

That is probably the most underappreciated implementation blocker, and it not often will get the eye it deserves. A well-specified MMM wants:

• Two to 3 years of weekly information as a baseline — sufficient to seize at the least two full seasonality cycles and a significant vary of spend variation.
• Constant channel-level spend granularity — not simply “digital,” however search, social, show, and video damaged out individually.
• Offline channels (TV, OOH, radio, occasions, unsolicited mail — which generally dwell in numerous methods) are owned by completely different groups, and infrequently use incompatible time granularities.
• Exterior covariates — macro indicators, competitor exercise, pricing information, and product launch calendars.
• For B2B particularly, longer gross sales cycles and decrease conversion volumes make the information necessities much more demanding. You typically want extra historical past.

In observe, what blocks most MMM tasks is the six-week information archaeology train that comes earlier than mannequin constructing. Finance owns income. The model workforce owns TV. The company owns digital spend. The spreadsheet somebody inbuilt 2021 is the one document of commerce promotions.

The mannequin is just nearly as good as the information archaeology that precedes it, and no one tells you that within the vendor demo.

Problem 2: You continue to have to roll up your sleeves

AI assistants have meaningfully lowered the syntax barrier. They will scaffold a Robyn run, generate a Meridian config, or assist debug a PyMC mannequin. What they will’t but do is navigate the judgment calls that make an MMM reliable:

• Select the place to take a seat on a Pareto frontier of a whole bunch of mannequin options (NRMSE vs. DECOMP.RSSD tradeoffs).
• Know when Nevergrad’s optimizer has meaningfully converged versus landed in an area minimal.
• Configure adstock transformation parameters (Weibull form/scale, geometric decay) to match practical channel dynamics.
• Diagnose why a mannequin assigns an implausible contribution to a channel, and whether or not to handle it with a previous, an information correction, or a variable exclusion.

In different phrases, vibe coding your strategy to an MMM will produce a mannequin that seems to work however is flawed in methods you gained’t catch. The scripting isn’t the onerous half. The area experience required to validate the output consists of working channel-specific incrementality experiments to calibrate your MMM.

Problem 3: The human experience layer isn’t elective

Even when the tooling matures to the purpose the place AI can run a reliable default MMM, the irreplaceable human contribution is encoding enterprise context — issues no mannequin can infer from the information alone:

• Adstock and carryover context: Your TV purchase has a four-week carryover. Your paid search has a three-day carryover. Your branded consciousness marketing campaign has a decay that spans months. This info isn’t discovered within the information. It’s within the minds of the channel specialists.
• Saturation curve form: Realizing a channel is probably going approaching diminishing returns earlier than the mannequin tells you so, and questioning the outcomes when the mannequin suggests in any other case.
• Guardrails and anomaly dealing with: Components like COVID troughs, product launches, pricing shifts, and macro disruptions should be modeled explicitly or flagged as structural breaks. AI doesn’t know your shopper had a pricing disaster in Q3 2022.
• Interpretation sanity checks: A modeled TV contribution of 40% for a model spending $2 million on TV might “really feel flawed” and warrant investigation. That instinct is earned, not computed.
• Organizational translation: Essentially the most technically right mannequin is nugatory in the event you can’t clarify why it recommends shifting 15% of the search price range to CTV in phrases a CMO and CFO will act on.

Lay the groundwork earlier than you construct a mannequin

The very best place to start is knowing what information it’s essential gas the mannequin and who wants to assist contextualize and translate that information into efficient advertising choices. Neither is simple or quick, however each are important if you wish to get significant insights out of your mannequin, no matter whether or not you select an open-source or subscription-based platform.

A sensible first step is to download Robyn’s demo script and experiment with the pattern information earlier than making use of it to your individual.