Costs
Part of Building a marketing mix model that survives scrutiny, step by step
What separates the strong marketing mix modeling examples from the weak
Strong marketing mix modeling examples name a decision, a data source and a limit; weak ones reverse-engineer a target ROI and hide the assumptions.
What to take away
- A strong example names the decision, the KPI source system, the period and the baseline. A weak one shows a result and stops.
- Reverse-engineering a model to hit a target return on investment is the weakest pattern, because the target sets the answer.
- Meta Robyn, Google Meridian and Google LightweightMMM are open-source software, not benchmarks.
- Nielsen, Analytic Partners, Ekimetrics, Circana and Gain Theory publish results too, and their definitions travel with them.
- Ask for diagnostics before you cite a case, then replace every borrowed figure with your own inputs.
The four fields that decide the comparison
Read any published example as a decision record, not as a performance claim. Strong ones state the decision, the KPI source system, the period and the baseline. Weak ones omit them one at a time.
Building a marketing mix model that survives scrutiny shows how those four fields get filled from raw spend and sales files, which helps before you judge someone else's.
Drop the baseline and a return-on-investment multiple stands alone. Drop the period and seasonality carries the number. Without a named KPI source, nobody can audit the figure.
Wikipedia's marketing mix modeling entry explains why this matters. The technique estimates response from observed spend and outcome history, so the inputs decide the output.
Example: one beverage launch, two write-ups
One beverage launch, two write-ups
Strong write-up
- Decision
- Shift 15 percent of Q3 national TV budget to retail media
- KPI source
- Weekly retail scanner units joined to the spend file
- Period
- 104 weeks to March, plus a 12-week holdout
- Baseline
- Pre-period average and the holdout comparison
- Promotions
- Modeled separately from media
- Diagnostics
- Convergence checks and posterior predictive fit published
Weak write-up
- Decision
- Optimize the media mix
- KPI source
- Revenue, dashboard unnamed
- Period
- Last year
- Baseline
- One ROI multiple, no comparison
- Promotions
- Folded into media
- Diagnostics
- None published
A reader can argue with the left column, which is the point. The right column can only be believed or rejected.
Campaign measurement examples worth adapting collect cases in the left column's shape across several industries, which makes them easier to reuse than a single vendor write-up.
Named examples and who publishes them
| Example | Publisher | Detail you can check |
|---|---|---|
| Meta Robyn | Meta, open source on GitHub | R package, free to install; a case index names partners reporting use |
| Google Meridian | Google, released 2025 | Python; fluege.de is its published advertiser case |
| Google LightweightMMM | Python on NumPyro, aimed at smaller datasets | |
| PyMC-Marketing | PyMC Labs | Bayesian MMM and customer lifetime value in one package |
| fluege.de | German flight search site, written up by Google | Query volume plus reach-and-frequency inputs |
| Nielsen | Measurement company | Mix models sit beside audience panels |
| Analytic Partners | Analytics firm | ROI Genome research across categories |
| Ekimetrics | Data science consultancy, Paris origins | Method published beside results |
| Circana | Retail data firm formed from the IRI and NPD merger | Retail panels paired with mix work |
| Gain Theory | WPP effectiveness agency | Runs mix models for advertisers across categories |
The Robyn index lists organizations that report using the package. The fluege.de write-up states a 3.8 percent incremental-revenue result. That figure belongs to that project and its inputs, so it does not travel.
Weak examples, described and named
The weakest pattern builds a model backwards from a target ROI. The next shows a result produced on a shipped demo dataset and calls it client evidence. Robyn ships a simulated weekly dataset for its demo script, and Meridian ships example data. Neither is a client.
A third pattern is a vendor write-up that calls its client a global beverage brand and publishes no diagnostics, no period and no baseline. Nothing there can be checked, so nothing can be reused.
If you cannot tell whether a case ran on real client data, treat it as a method demonstration. Questions to settle before you start covers the ones that recur when a team weighs published examples against its own constraints.
Verifying a vendor case before you cite it
- Ask for the period, the data grain and the holdout window. No holdout means no credible incrementality claim.
- Ask for convergence checks, posterior predictive fit and a stated uncertainty range. Statistical significance tells you an effect is unlikely to be chance, not that it is large enough to act on.
- Ask which system holds the KPI. If the outcome is GA4 sessions rather than orders or finance revenue, note it in your own write-up.
- Ask whether price, promotion and distribution were separated from media.
- Ask what decision followed, and what the next measurement showed.
Turning a borrowed example into a testable brief
- Name the KPI and the system that holds it
- State the budget decision and its date
- Map paid, owned and non-media treatments
- List conditions that moved spend and outcome together
- Set the geographic and time grain
- Define spend limits and review criteria
Testable Brief Checklist
- Name exact KPI and source system
- State budget decision and action date
- Map paid, owned, non-media treatments
- List spend and outcome conditions
- Define geographic and time grain
- Set feasible spend limits and review criteria
Reject any example that starts with a target ROI. Use examples to borrow variables, checks and delivery formats instead. Proving the numbers hold up sets out the inputs and diagnostics worth recording before you sign off.
