Credit for channels clicks miss
TV, audio, retail media and brand spend are measured alongside search and social.
- Home
- Marketing mix modeling
Marketing mix modeling
See what moving budget between channels would do.
We build custom marketing mix models on Google's Meridian or Meta's Robyn, fed by your own data pipeline, to measure what each channel contributes, including the ones clicks can't see. The model retrains as new data lands.
What you get
Budget decisions backed by a model you can question.
Budget moves you can defend
Reallocation rests on modeled revenue, not on which platform claims the most credit.
Unaffected by tracking loss
The model works on aggregate data, so cookie loss and iOS changes don't erode it.
Runs on your infrastructure
The model and its pipeline stay with you, not in a vendor's black box.
The problem
Attribution tells you less than it used to
Last-click flatters search, multi-touch models fail as tracking degrades, and platform-reported conversions each claim credit for the same sale. Meanwhile TV, audio, out-of-home and brand spend barely leave a click trail at all. MMM answers the question attribution can't: if we move a dollar between channels, what actually happens to revenue?
Before and after
What changes.
Click attribution
- Last-click flatters search
- Every platform claims credit for the same sale
- TV, audio, out-of-home and brand spend barely leave a click trail
Marketing mix model
- Planning starts from one shared model
- Spend pulled back before a channel saturates
- Budget shifts tested on paper before they're made
- Estimates that stay current without a new project
What an MMM engagement includes
What’s included.
- Data pipeline assembling spend, impressions, revenue, pricing, promo, and seasonality history
- Custom model build on Meridian, Robyn, or a bespoke Bayesian specification, whichever fits your data
- Adstock and saturation curves per channel, so diminishing returns are visible before budget is wasted
- Budget-allocation scenarios: what shifting spend between channels does to modeled revenue
- Calibration against experiments and holdouts where available, so the model earns trust
- Automated retraining as new weeks of data land: a living model, not a one-off consulting PDF
How we work
How we run it.
Step 1
Audit data history first: MMM needs 2+ years of clean weekly data, and we build the pipeline that assembles it
Step 2
Start with a model your team can interrogate, not a black box
Step 3
Validate against known events (price changes, channel pauses) before trusting forecasts
Step 4
Operationalize: the model runs on your infrastructure and retrains on schedule
Typical stack
What it runs on.
- Google Meridian
- Meta Robyn
- Python / Bayesian modeling
- dbt
- Snowflake / BigQuery
- Airflow
FAQ
Common questions.
What is marketing mix modeling (MMM)?
MMM is a statistical technique that estimates how much each marketing channel contributes to revenue by analyzing historical spend and outcome data along with pricing, seasonality, and promotions. Because it works on aggregate data rather than user tracking, it is unaffected by cookie loss, iOS privacy changes, or walled gardens.
How is MMM different from attribution?
Attribution assigns credit for individual tracked conversions along a click/view path; it breaks as tracking degrades and can't see offline or brand channels. MMM measures channel contribution statistically from aggregate outcomes, capturing effects attribution misses at the cost of granularity. Mature teams run both: attribution for tactical in-platform decisions, MMM for budget allocation.
How much data does an MMM need?
Roughly two to three years of weekly data across spend, revenue, and the major business drivers (pricing, promotions, distribution, seasonality). Less history means wider uncertainty. Assembling and cleaning this dataset is usually half the project, which is why MMM done well is a data engineering problem before it's a statistics problem.
Should we use open-source MMM (Robyn, Meridian) or a vendor platform?
Open-source frameworks like Meta's Robyn and Google's Meridian are now credible foundations: transparent, auditable, and free of per-seat licensing. What they don't include is the data pipeline, calibration, and operationalization around them. We build on open source and invest the budget difference in data quality, which is what actually determines model accuracy.
Find your real channel mix
Tell us your channels and data history and we'll assess whether your data can support an MMM honestly, before you commit.