Build vs buy marketing mix modeling: what each path involves, and how to decide
Building a marketing mix model means using an open-source library such as Robyn, Meridian or PyMC-Marketing with your own data team, which gives full control and needs data scientists and engineering time. Buying means paying for a platform that supplies the data pipeline, modeling, refreshes and support. Build suits teams with spare modeling skills. Buy suits most others.
The first model is the smaller part of either decision. Most of the cost and most of the risk come afterwards: keeping data connectors working, refreshing the model, calibrating it with experiments, and keeping it trusted when the person who built it moves on. This page sets out what each path includes, compares them on six dimensions, lists the cost components to budget for over three years, and describes when each path is the better choice.
What building includes
Building starts with a library. Meta's Robyn, Google's Meridian and PyMC-Marketing are free to download and cover the modeling itself. They ship no user interface, no data connectors and no support contract.
The rest has to be built and run by the team. Data pipelines pull spend from every ad platform and outcomes from sales systems, on a schedule, and are fixed when a platform changes its export format. Modeling means choosing variables, delayed-effect and saturation settings and priors, then validating the result. Refreshes repeat the fit as new weeks arrive, and check that results have not drifted for the wrong reasons. Reporting turns the output into something a planning meeting can use. Experiments have to be designed and run separately, because the libraries accept calibration inputs but do not design or run tests.
A build needs at least one experienced modeler and engineering support, and it depends on those people staying.
Validation is part of the build as well: checking the model against periods it was not trained on, against experiment results, and against what the business knows to be true. Without it, the first model is a draft.
What buying includes
Buying means paying a vendor for a platform or a managed service. The vendor supplies connectors to the main ad platforms and sales systems, the modeling engine, scheduled refreshes, dashboards and planning tools, and support.
Offerings vary widely. Some are self-serve software, some are services where the vendor's analysts run the model, and some combine the two. Some include experiment design and calibration, and some leave that to the client.
Buying reduces the headcount needed and shortens the time to a first result. It adds a dependency on the vendor, and the method is only as visible as the vendor chooses to make it.
Before signing, check four things in the contract or the demo: what data and outputs you can export, whether you can see and change the model's assumptions, how experiments and calibration are handled, and what happens to your data and model history if you leave.
The differences, dimension by dimension
Time to a first model. Build: months, most of it data engineering. Buy: weeks, where connectors already exist.
Control. Build: full control over the specification and data. Buy: control limited to what the platform exposes.
People. Build: modeling and engineering skills in-house. Buy: someone to own the relationship and use the results.
Ongoing effort. Build: the team maintains pipelines, refreshes and reporting. Buy: the vendor does, within the contract.
Continuity. Build: knowledge sits with a few people. Buy: it sits with the vendor and the platform.
Transparency. Build: as transparent as the team makes it. Buy: depends on the vendor, so it is worth checking before signing.
The cost components to budget for
Costs vary too much by business to give reliable figures, so the useful step is to list them before comparing quotes.
For a build: salaries for modelers and data engineers, cloud computing for fitting and refreshes, time to build and maintain connectors, time to design and run experiments, and the cost of rebuilding knowledge when someone leaves.
For a buy: the subscription or service fee, internal time to supply data and review results, any experiment costs the vendor does not cover, and the cost of switching if the vendor stops fitting the business.
Compare the two over three years. The first-year cost of a build is mostly setup. The later years are mostly maintenance, and that is the part most comparisons underestimate.
Two costs are easy to miss on both paths. The first is the internal time to turn results into decisions: preparing plans, answering finance and explaining changes between refreshes. The second is the cost of acting on a wrong number, which is why validation and calibration belong in the budget.
Why a model goes stale quietly
A model that is not maintained keeps producing numbers. A platform changes its export format, a channel is added or renamed, a promotion calendar shifts, and the model continues to fit on data that no longer means what it did.
The results still look reasonable, which is why this is the most common way an in-house build loses value. Refreshing, checking stability between refits and reconciling inputs with the source systems are the tasks that prevent it.
Simple checks catch most drift: compare each refresh's channel estimates with the previous one, investigate any large move that has no business reason, and reconcile the model's total spend and revenue with the finance figures every cycle.
Calibration on either path
A marketing mix model learns from history, so on its own it can mistake a channel that moved with demand for one that drove it. Experiments, such as geo tests, supply that separation, and their results can be fed back into the model.
On a build path, the team designs and runs the experiments and connects the results to the model. On a buy path, it depends on the vendor: some include experiments, some accept results from outside, and some do neither. It is worth asking about specifically.
When building is the better choice
Building is the better choice when the business already has a strong data science team with time to spare, when the data or the business model is unusual enough that standard platforms fit poorly, when full control over the method is a requirement, and when the team can commit to maintaining the model for years. It is also a good way to learn the method before deciding what to buy.
When buying is the better choice
Buying is the better choice when the team is small, when results are needed within weeks, when the business cannot depend on one or two people staying, and when the results have to arrive on a planning calendar the business already runs to. It is also the better choice when experiments are needed and the team has no capacity to design them.
And it suits businesses where measurement is one of many priorities for a small analytics team, so that the team's time goes on using results rather than on keeping the pipeline running.
What neither path guarantees
Neither path guarantees a model the business will use. A model has to answer the questions planners ask, in their terms, at the time they plan.
Neither path removes the need for history: enough of it behind each channel, and enough movement in each channel's spend, before the estimates are stable.
And neither path is permanent. Teams move from building to buying when a key person leaves, and from buying to building when the in-house team grows. Both moves are easier when the data and the experiment record are kept in a form that does not depend on either model.
How to decide
Build if you have modelers who will still be there in two years, time to maintain pipelines and refreshes, and a need for full control.
Buy if you need results quickly, if the model has to survive staff changes, or if experiments and calibration are part of what you need.
Either way, list the cost components above for three years, and ask how calibration with experiments will work before committing.
Cassandra is a marketing mix modeling platform on the buy path, with experiment calibration included. Its production engine is Bayesian. Data refreshes daily through the connectors, and each re-fit is a new model linked to the one before it, so earlier results stay intact.
Geo experiments are designed and run inside the same platform, using the same method family as GeoLift, and their results become constraints on the model. Reads sit at campaign level, in the planning layer.
A price is published, so the build-or-buy comparison can use a real figure on the buy side.
Moving between the two paths
From build to buy, the main asset is the cleaned history of spend and outcomes. It shortens onboarding, and the old model's results give a reference point for the new one. Run both on the same period before switching off the old model.
From buy to build, ask the vendor for exports of the data, the outputs and the settings used. How much is available depends on the vendor, which is a reason to ask about exports before signing.
In both directions, keep the record of experiments. Test results remain valid evidence whichever model uses them.
Plan the switch around the planning calendar. Moving models in the weeks before a budget cycle leaves no time to compare the old and new results, and the first question at the next planning meeting will be why the numbers changed.
Questions
Which is better for me, to build or buy software?
Build if you have spare modeling capacity, need full control and can maintain the model for years. Buy if you need results quickly, want the model to survive staff changes, or need experiments included.
Are open-source MMM libraries free?
The code is free. Robyn, Meridian and PyMC-Marketing ship no user interface, connectors or support, so the cost moves to the people who build and run the model.
How long does it take to build a marketing mix model in-house?
Usually several months for the first model, most of it spent on data pipelines and validation, followed by ongoing work on refreshes and maintenance.
What should I ask an MMM vendor before buying?
How the model is specified and whether you can see it, whether priors can be changed, what can be exported, how experiments are handled, and what happens to your data and outputs if you leave.
Can I start by building and buy later?
Yes. A first in-house model is a good way to learn the method and clean the data, and that cleaned history makes a later platform faster to set up.