Failed Experiment: When Google’s Lightweight MMM Turned My Ads into Zombies

Sep 03, 2025

Prologue: The Experiment Nobody Warned Me About

Let’s rewind. I’m a data scientist who has spent more years than I’d like to admit tinkering with Marketing Mix Modelling (MMM). I’ve worked with big media agencies, publishers, and brands, and I’ve seen the good, the bad, and the “how the hell did that pass as statistics” of marketing analytics.

So when Google released Lightweight MMM — an open-source Python package promising a faster, easier, Bayesian way to measure media effectiveness — I was hooked. Free, powerful, and with Google’s shiny logo attached? Sign me up.

But like most free experiments in life (free tattoos, free food samples from a sketchy street vendor, free trial SaaS tools that need three days of debugging before running), what looks good upfront often leaves scars.

What started as excitement quickly spiralled into one of the strangest episodes of my MMM career. And it all revolved around one thing: zombie ads.

While Google Lightweight MMM gained popularity as a free open-source framework for marketing mix modeling, it has recently been replaced by Google Meridian. This upgraded Google open-source MMM tool offers enhanced features, better scalability, and improved accuracy for marketers and data scientists exploring media effectiveness — Google’s Lightweight MMM has been one of the most widely adopted open-source tools for marketing mix modeling over the past few years. However, it has now been succeeded by Google Meridian, a more advanced open-source solution designed to improve speed, flexibility, and reliability for MMM practitioners.

Falling in Love with Lightweight MMM

If Meta’s Robyn was a messy love affair (remember the AWS bills and evolutionary computing nightmare?), Google’s Lightweight MMM felt different. It was cleaner.

Installation was a breeze:

pip install lightweight_mmm

No need to spin up half of AWS just to get a model to converge. No evolutionary computing black hole. Just Python, some JAX magic, and Bayesian priors. Honestly, the first time I ran it, I thought:

“This might be it. This might actually save me from the endless cycle of agency black-box MMM models and overpriced consultants.”

The package had everything a data scientist dreams of:

Bayesian regression models baked in
Adstock transformations for media lag effects
Diminishing returns functions to model saturation
Forecasting tools to predict scenarios
And, of course, the warm glow of Google’s name

It looked like an MMM playground where data people like me could build elegant, transparent, trustworthy models.

But like any playground, it came with rusty swings and broken slides.

With Google LightweightMMM, the marketing mix modeling process becomes easier to manage thanks to its transparent workflow, ensuring a consistent and systematic methodology for building MMM models — Google Lightweight MMM provides a clear step-by-step workflow that guides users through the entire modeling process, making MMM a more structured and systematic approach.

Enter the Zombies (a.k.a. Adstock Gone Wrong)

The real trouble started when I fed in my first media dataset. A mix of digital, TV, and other media spends over a couple of years. I applied the default geometric adstock transformation. Everything looked fine.

Until the results came back.

One media channel — let’s call it “Radio” to protect the innocent — showed ROI numbers that defied capitalism, physics, and human intelligence.

According to Lightweight MMM, every dollar I put into radio advertising was delivering so much incremental revenue that I should have mortgaged my house and bought out every radio station in the country.

But here’s the kicker: even after I stopped spending on radio in the dataset, the model insisted radio was still delivering revenue.

Weeks later. Months later. Like a zombie crawling out of the grave, still dragging incremental sales behind it.

This wasn’t advertising anymore. This was necromancy.

Not long ago, a potential client contacted us after noticing a strange pattern in their MMM outputs: the model kept attributing sales impact to ads for several months after spend had ended. Their previous provider had developed the MMM using Google LightweightMMM, but when it came time to plan their new media strategy, the abnormal results raised doubts. They approached us for a fresh perspective and an impartial audit, looking for clarity and trustworthy guidance — Recently, we heard from a prospect who raised a serious concern: their MMM results were showing ad impact lasting for months even after the media spend had completely stopped. The model, built by a service provider last year using Google Lightweight MMM, was now being used for media planning — but the results appeared unrealistic. They reached out to us seeking a second opinion and an independent review to validate the model and ensure reliable decision-making.

The Beauty and the Horror of Bayesian Stats

Now, to be clear: Lightweight MMM wasn’t “wrong” in a statistical sense. The model was doing exactly what it was told: applying adstock decay, fitting priors, and finding the posterior distributions that maximized likelihood.

The problem? The defaults were too generous.

The adstock parameter wasn’t constrained tightly enough, which meant the decay rate let media effects linger… forever
Diminishing returns weren’t kicking in early enough, so spending looked like it had infinite upside
Priors were set in a way that gave too much credibility to weak signals

In Bayesian language, it was elegant. In commercial language, it was insane.

This is where marketers and data scientists live in different universes. A data scientist sees pretty posterior plots. A marketer sees a recommendation to double their spend on a channel that hasn’t run an ad in three months.

And if you deliver that to a CMO? You’re not getting invited back.

Bayesian regression plays a major role in marketing mix modeling (MMM), but it comes with drawbacks. The accuracy of the results heavily depends on how coefficient priors are defined. If posterior results suggest media channel contributions that don’t pass marketers’ “spend share vs. impact share” checks, the model needs to be re-tuned. This iterative process can be lengthy, as adjusting the prior for one coefficient often alters the impact distribution across other variables. Such sensitivity makes model calibration challenging and highlights one of the key limitations of Bayesian regression for MMM. Unless constrained Bayesian regression with MCMC is applied, marketers and analysts face repeated cycles of adjustment before achieving commercially acceptable outputs — In most cases, Bayesian regression for marketing mix modeling (MMM) is highly sensitive to the priors assigned to coefficients. Modelers often simulate coefficient distributions in advance and then update them with data to generate posterior estimates. However, when the media-attributed impact doesn’t align with marketers’ expectations, the model must be re-run. Marketers usually validate results through a “spend share vs. impact share” sense check, ensuring that media spend aligns closely with attributed impact. This creates a back-and-forth tuning cycle, since even small changes in coefficient priors can shift the impact of other variables. As a result, Bayesian regression in MMM often struggles with stability, unless a constrained Bayesian regression approach with MCMC sampling is adopted.

We’ve developed an updated Bayesian regression approach that uses constraints to regulate coefficient sampling, reducing the heavy reliance on prior definitions and assumptions. Instead of depending solely on priors, we incorporate the expected impact share for each variable directly into the likelihood function, alongside the priors, data, and constraints. If you’re interested, you can learn more in our latest blog post: Upgrading Google Meridian: A Constrained Bayesian Approach for Robust Marketing Mix Modelling.

The Zombie Apocalypse in the Boardroom

I’ll never forget the first time I presented one of these Lightweight MMM results to a marketing team.

I walked them through the model structure. I showed them the tidy Bayesian intervals. I presented the channel-level ROI.

The client stared at the slide. Then someone said:

“Wait. Are you telling me radio is still selling our product, even though we haven’t run radio since last year?”

Silence.

I scrambled. “Well, technically, what the model suggests is that there are lingering effects from brand equity…”

They cut me off.

“So basically… you’re saying our dead radio ads are zombies?”

I laughed nervously. They didn’t.

That was the moment I realized: what looks like statistical beauty to us is commercial nonsense to them.

Why This Matters (and Why It’s Not Google’s Fault)

Here’s the truth: I don’t blame Google. Lightweight MMM is a gift to the data science community. It’s transparent, free, and pushes MMM out of the agency black box.

But it’s not built for marketers.

It’s built for data scientists who:

Understand how to set priors
Know when and where to constrain adstock parameters
Can interpret credible intervals instead of taking point estimates as gospel
Have the patience to debug why ROI looks like it was written by a drunk economist

Marketers? They want clean answers. They want ROI that matches reality. They don’t care if your posterior has a 95% credible interval. They care if they can trust the numbers enough to take them to their CFO without getting laughed out of the room.

And trust me, no CFO believes in zombie ads.

Marketing mix models must reflect that ad impact diminishes over time. It’s not enough to set adstock parameters correctly; the results also need to align with reasonable business logic. Otherwise, a model might imply that one ad run could generate sales for years, which is unrealistic and impractical for real-world marketing planning — Ads cannot have infinite or continuous impact over time. Properly selecting adstock transformations and tuning their parameters is crucial, but it’s equally important to ensure the results make practical sense. Without this, a model could suggest that a single ad campaign continues driving sales indefinitely — a scenario that clearly defies realistic marketing practices.

Lessons from the Graveyard

Here’s what I learned (the hard way) from my Lightweight MMM experiment:

Adstock needs guardrails
Never let adstock decay rates run free. If you don’t constrain them, you’ll end up with immortal ads.
Priors are everything
Lightweight MMM makes Bayesian easy, but the defaults are too weak for commercial use. Tailor your priors or prepare for weirdness.
Marketers need auditability, not just plots
When results don’t match reality, you need a framework to debug and explain, not just statistical outputs.
Open source ≠ ready for the boardroom
Data scientists can love Lightweight MMM. But hand it raw to a marketer, and you’ll burn trust faster than a dodgy influencer campaign.

Epilogue: From Zombies to Trustworthy MMM

So where does that leave us?

Lightweight MMM is a powerful experiment, but like my Meta Robyn adventure, it proved one thing: open-source MMM is too raw for most marketers. It’s brilliant for tinkering, for learning, for research. But for actual commercial decision-making? It’s a minefield.

That’s why I built More Than Data.

No zombies. Our MMM models have guardrails so ads don’t live forever.
No black boxes. We make the stats transparent but the outputs business-ready.
No betrayal in the boardroom. What we deliver is audit-ready, CFO-proof, and marketer-friendly.

Because at the end of the day, marketers don’t care about pretty Bayesian plots. They care about trust.

And trust, my friends, is something zombie ads can never give you.

With Minute MMM from More Than Data, each media variable undergoes precise adstocking and diminishing return adjustments. Taking BVOD as an example, the model shows a clear relationship between media spend and its contribution. Importantly, the impact doesn’t persist indefinitely once the spend stops. The carry-over effect exists but decays rapidly, vanishing entirely within 2–3 weeks, making the outputs practical for real-world marketing planning — More Than Data’s Minute MMM carefully applies adstock and diminishing return transformations for each media variable. For example, with BVOD (Broadcasting Video on Demand), the model accurately reflects spend versus contribution. Unlike flawed models, there’s no endless “spend-off but impact still continues” scenario. The carry-over effect is visible but decays naturally, completely disappearing after 2–3 weeks, ensuring realistic and actionable results.

Final Thoughts (and a Sequel Teaser)

If you enjoyed this story, you’ll love my other failed experiment with Meta’s Robyn, where I discovered the joy of evolutionary computing eating AWS bills alive. Spoiler: it didn’t end well either.

Open-source MMM is like free street food: sometimes delicious, sometimes food poisoning. Either way, you’ll have a story.

The question is: do you want a story, or do you want results you can actually take to your CFO?

At More Than Data, we choose results.