Open-source packages and AI agents have made building a marketing mix model (MMM) easy, but they have not made it any easier to know whether the model is right. The barrier moved from building the model to supervising it: calibrating it against experiments, comparing models and naming who signs off on the number. Southeast Asia, with short data histories and many channels per marketing dollar, carries the most exposure.
Key takeaways
- An AI agent can now pull an open-source package, load spend and revenue data and produce a working marketing mix model with almost no statistical background.
- Robyn (Meta) asks the least of the user, Meridian (Google) needs data science coding and PyMC Marketing needs Bayesian expertise; agents erase the setup gap but not the interpretation gap.
- In fifteen large advertising experiments studied by Gordon, Zettelmeyer, Bhargava and Chapsky, observational methods were off by a factor of three or more in about half the studies.
- Models often break on data before modelling starts: AI-assistant referrals with no referrer, unpaid brand building and stale competitor and audience inputs.
- Supervision means calibrating against experiments, treating disagreement between models as a finding and naming the person who signs off on the number.
What changed in marketing mix modelling?
Building a marketing mix model has become something almost anyone can do. Knowing whether that model is telling the truth remains rare. AdExchanger recently covered the rise of open source marketing mix modelling, coining the label OS MMM along the way, and the excitement behind the piece was genuine. It just was not pointed at the right achievement.
Julian Runge at Northwestern's Medill School has made a version of this argument that deserves more attention than it gets. An AI agent can now be prompted to pull an open source modelling package, load a client's spend and revenue data, and produce a working marketing mix model with almost no statistical background required. Anyone who has paid a consultancy six figures and waited a quarter for an answer to a question as simple as which channel is actually working should pause on that for a moment.
Then comes the part of the argument that does not travel well in a summary slide. What became cheap is the act of producing a model. What still costs exactly what it always cost is being right about one.
Which part of marketing mix modelling actually got cheaper?
Runge and Koen Pauwels published a practical comparison of the major open source packages this year, and it is unusually clear-eyed about what each one demands. Robyn, built by Meta, combines ridge regression with an evolutionary search process. It asks the least of the user and can return a working model in under a day. Google's Meridian expects real data science coding ability. PyMC Marketing expects genuine Bayesian expertise and often takes a couple of weeks to run properly.
| Package | What it asks of the user | How it calibrates against experiments |
|---|---|---|
| Robyn (Meta) | The least; ridge regression with evolutionary search, a working model in under a day | Through multi-objective optimisation, which Runge and Pauwels flag as less understood and less tested |
| Meridian (Google) | Real data science coding ability | Experiment results as Bayesian priors; Meridian GeoX lets geo experiments anchor the model |
| PyMC Marketing | Genuine Bayesian expertise; often a couple of weeks to run properly | Experiment results as Bayesian priors |
An AI agent erases the setup gap between these three tools. It does nothing to erase the interpretation gap. If anything it widens it, because now all three approaches can produce a confident-looking chart by the end of the week, and every one of those charts looks equally trustworthy on a slide.
Why can't you trust a model's output on its own?
Picture the scenario. An agent has run overnight and a clean deck is sitting in the inbox by morning. One channel comes back reporting a return on ad spend of 4.1. Brand search comes back at 0.9. Someone in the room wants budget reallocated by the end of the week and asks the one question that actually matters. Is that number correct?
Nobody in the room can answer that from the output alone, because the output gives no indication of its own reliability.
The scale of what can go wrong here is not new. Researchers Gordon, Zettelmeyer, Bhargava and Chapsky ran fifteen large scale advertising experiments covering roughly half a billion user experiment observations and over a billion and a half impressions, then compared the results against the observational methods the industry typically relies on. In about half of the studies, the estimated lift was off by a factor of three or more. One campaign with a true lift of around two percent was estimated by observational methods at over three thousand percent. Another campaign with a genuine lift near seventy percent came back estimated at more than four times that figure.
That research is several years old now. Nothing about agentic AI changes the underlying problem it uncovered. A faster pipeline running on the same observational data will reproduce the same bias. It will simply do so sooner and with better formatting.
Where do marketing mix models break before the modelling starts?
Most poor marketing mix modelling output is not really a modelling failure. It is a data failure that the model then presents with unearned confidence.
Start with channels the model cannot see at all. If customers are asking an AI assistant like ChatGPT or Perplexity which product to buy and arriving at a site with no referrer attached, that traffic simply has no column in the dataset. The model either folds it into baseline demand or quietly credits it to whatever campaign happened to be running nearby. This is already showing up in the data. Touchigh, which works on generative engine optimisation for cross-border sellers entering the US market, has reported AI referral traffic growing by close to seven hundred percent across the most recent holiday season, with conversion rates from AI assistant referrals running thirty to forty percent above the site average. The precise figures belong to that one study. The structural point does not. A channel growing at anywhere near that pace, and completely absent from the dataset, is a specification error rather than a rounding error.
Then there is the spend line that never appears at all. Founder publishing, category authority built slowly over time, the long unpaid habit of simply showing up consistently in a category's conversation. None of it carries a media cost, none of it produces a visible burst, and its effect often shows up a full quarter or two later, which is precisely the shape a marketing mix model handles worst. It gets swallowed into trend and disappears from the analysis entirely.
And then there is the least glamorous failure of all. Input quality. A model fed a thin, outdated picture of competitors and audience behaviour will confidently allocate budget against a market that no longer exists. There is a documented case involving Mothercare Singapore that cut the time spent on audience research by eighty percent, freeing up more than twenty-five hours a week, while tracking over four thousand digital ads annually within its category. The genuinely interesting figure there is not the time saved. It is that the market picture feeding every downstream model became current instead of stale by a quarter.
Where does the modelling expertise go now?
It does not disappear. It moves from building the model to supervising it. Three things are worth fixing before the next model run.
Calibrate against an experiment rather than against instinct. Meridian and PyMC Marketing both allow experiment results to be fed in as Bayesian priors. Robyn calibrates through its multi-objective optimisation process, which Runge and Pauwels specifically flag as far less understood and far less rigorously tested. Google introduced Meridian GeoX in May 2026 specifically so that geo experiments can anchor a model directly rather than sitting alongside it as a separate exercise.
Treat disagreement between models as a finding in itself. Running two different packages on the same dataset and watching them disagree by an order of magnitude on a single channel teaches something real about how weakly identified that channel is within the data. Most teams run exactly one model and never discover this.
Name the person who signs off on the number. Before any figure reaches a live budget decision, someone needs to own the risk of being wrong about it. This is the single control most teams quietly skip.
That layer of supervision has to live somewhere, and for a smaller or independent agency it often lives nowhere at all, simply because there is no dedicated analytics bench to absorb the work. That gap is part of what makes tools built to give smaller shops the intelligence and automation layer of a larger network so relevant right now. Time saved in reporting only creates value if it is reinvested into judgement rather than absorbed into more reporting.
Why does this land hardest in Southeast Asia?
Southeast Asia stands to gain the most from free modelling tools and carries the most exposure when those tools are used without proper supervision. Data histories tend to be shorter. The number of channels per marketing dollar tends to be higher, spanning TikTok Shop, Shopee, Lazada, Meta, a Grab placement, and a WhatsApp community, often all within a single campaign. Fewer markets are large enough to support a genuinely clean holdout test. A regional brand can easily end up with more channel columns in its dataset than it has clean weeks of history behind it, which is exactly the condition under which a marketing mix model will happily return a precise, stable, and wrong answer.
Free access to these tools is unambiguously good news. What arrives alongside that access is a much shorter distance between a bad assumption and a budget that has already been reallocated on the strength of it.
So before the next model runs, the question worth asking inside any organisation is simple. Who is accountable if the answer turns out to be wrong?
Where does SOMIN fit into this?
The gap this article describes is not really a modelling gap. It is an intelligence gap, the space between a chart that looks confident and a chart that has actually been checked against how people in a market behave. SOMIN was built around closing that exact space. SoMonitor's Brand Tracker and Perspective Studies exist to keep the audience picture underneath any model current rather than a quarter stale, while SODA and the GWI partnership give teams a way to test what a model is claiming against real consumer signal before that number ever reaches a budget decision. The tools for building a marketing mix model have become far more accessible. The judgement required to trust one has not, and that is the part SOMIN is built to support.
FAQ
Can an AI agent build a marketing mix model?
Yes. An agent can pull an open-source package, load a client's spend and revenue data and produce a working model with almost no statistical background required. What it cannot do is tell you whether the model is right, because the output gives no indication of its own reliability.
What is the difference between Robyn, Meridian and PyMC Marketing?
Robyn, built by Meta, combines ridge regression with evolutionary search and asks the least of the user. Google's Meridian expects real data science coding ability, and PyMC Marketing expects Bayesian expertise and often takes a couple of weeks to run properly. Meridian and PyMC Marketing accept experiment results as Bayesian priors; Robyn calibrates through multi-objective optimisation.
How wrong can observational advertising measurement be?
In fifteen large advertising experiments analysed by Gordon, Zettelmeyer, Bhargava and Chapsky, observational estimates of lift were off by a factor of three or more in about half the studies. One campaign with a true lift of about two percent was estimated at over three thousand percent.
How do you check whether a marketing mix model is right?
Calibrate it against experiments, for example geo experiments through Meridian GeoX, rather than instinct. Run two different packages on the same data and treat large disagreements as a sign the channel is weakly identified, and name the person who signs off before any number reaches a budget decision.
Why is marketing mix modelling riskier in Southeast Asia?
Data histories are shorter, campaigns spread across many channels such as TikTok Shop, Shopee, Lazada, Meta, Grab and WhatsApp, and fewer markets are big enough for clean holdout tests. A brand can have more channel columns than clean weeks of data, which is when a model returns a precise, stable and wrong answer.
