Bayesian MMM for Mobile App Developer Acquisition

Appodeal

Marketing & Media
PyMC-Marketing

The Challenge

Appodeal needed to figure out where their marketing dollars were actually working. They were acquiring app developers onto their platform through multiple channels, but the path from first touch to sign-up was tangled — developers might see a paid ad, read an organic article, attend a webinar, and then convert weeks later. Standard attribution gave credit to whichever touchpoint happened to be last, which told them almost nothing about what was genuinely driving growth.

The real question wasn't just "which channels contribute?" It was "how confident can we be in those estimates?" Because when you're reallocating a media budget, acting on a shaky point estimate is worse than having no estimate at all.

Our Approach

We built a media mix model that disentangles channel contributions from noisy, overlapping marketing data. The model accounts for the fact that advertising effects don't appear instantly — spend today influences sign-ups over the following days and weeks, with the decay pattern varying by channel. It also captures diminishing returns, so the model knows that doubling spend on a channel doesn't double its impact.

The structure shares information across channels where it makes sense, which stabilizes estimates even when individual channel data is noisy. We validated the model rigorously against held-out data and its own internal consistency, and every channel estimate comes with a credible range — not just a best guess, but an honest accounting of uncertainty.

Results

Appodeal walked away with a clear decomposition of what each marketing channel was contributing to developer acquisition.

More importantly, every return-on-ad-spend estimate came with uncertainty bounds, which changed how the team thought about budget allocation. Instead of chasing the channel with the highest point estimate, they could weigh expected return against confidence — investing more heavily where the evidence was strong and running targeted experiments where it wasn't.

PyMC Labs Team

  • Niall Oulton
  • Thomas Wiecki
  • Bill Engels
  • Sef M

More Work in Marketing & Media

Live Nation

Built a marketing effectiveness model for concert ticket sales that handles the unique rhythm of tour promotion — from announcement buzz through on-sale surges — across more than 125 artists.

MMMHSGPHierarchical ModelsReverse AdstockZero-Inflated Models

Fox Broadcasting Company

Helped Fox Broadcasting's data science team build marketing models for TV advertising revenue and measure the true impact of NFL promotional campaigns.

MMMCounterfactual AnalysisBass Diffusion ModelCausal Inference

Gain Theory

Worked with Gain Theory across multiple fronts — strengthening their Bayesian modeling practice, delivering marketing models for their clients, and building a viewership forecasting model for streaming content.

MMMHierarchical Bayesian MMMBass DiffusionHSGPLift Test Calibration

Let's Chat, We Respond Fast

Tell us about your problem. We typically respond within 24 hours.

Schedule a Consultation