Funnel-Aware MMM: A Bayesian Architecture for Full-Funnel Marketing Optimization
On this page
An additive, funnel-blind marketing mix model (MMM) can omit a path that matters for budget decisions: upper-funnel activity creates demand that lower-funnel channels later capture. Additive response functions do not imply that observed channel spends are statistically independent. Check the model's graph and controls before interpreting attribution as a causal budget recommendation.
- Identify direct and mediated paths before cutting a channel with weak direct returns.
- Record budget caps and channel activation dates; observed spend can conceal unmet demand.
- Compare predictions with held-out periods and, where feasible, live incrementality evidence.
- Keep historical spend modeling separate from experiments on a new offer or message.
What does the documented architecture add?
PyMC Labs' May 24, 2026 case, by Niall Oulton, Timothy McWilliams, and Teemu Säilynoja, describes a funnel-aware model for Nürnberger Versicherung. Lower-funnel spend is both an outcome of upper-funnel activity and a predictor of leads. Upper-funnel activity also has a direct path to leads.
The implementation combines separately estimated adstock and saturation transformations, a time-varying HSGP baseline, lower-funnel-specific controls, and a censored likelihood for cap-hitting days. The authors report higher upper-funnel estimates after censoring and masking periods before lower-funnel channels became active. That direction is a result of this engagement, not a guarantee for other datasets. Their optimizer jointly varies channel allocations and lower-funnel budget caps. The PyMC-Marketing upper-funnel guide is a related implementation reference.
What should the likelihood mean on a capped day?
Treat potential spend as a latent quantity. An uncapped day provides an observed spend value. On a day when the cap binds, the observation instead says potential spend is at least the cap. The likelihood must use that distinction; replacing the cap with a guessed total invents data.
Check whether a cap actually bound, rather than labeling every day with an administratively configured cap as censored. Preserve changes in cap policy and channel availability. A zero before a channel launches is not evidence that an active channel attracted no demand.
Which assumptions make the recommendation credible?
A graph encodes assumptions; fitting the graph does not prove them. A mediated causal interpretation needs assumptions about confounding of spend, demand, and leads, including the mediator-to-outcome relationship. Promotions, seasonality, bidding changes, and business conditions can affect several variables together.
A flexible baseline can absorb real media response, or media variables can absorb baseline variation. When campaigns coincide with seasonal peaks, inspect sensitivity to kernels, priors, lag structure, and controls. Posterior intervals describe uncertainty under the specified model; they do not automatically cover omitted confounding or a changed market.
| Check | Why it matters | Evidence to request |
|---|---|---|
| Mediation structure | An omitted downstream path can distort attribution | Graph, variable definitions, and competing specifications |
| Censoring | Observed spend may understate potential spend | Cap-hit records and likelihood definition |
| Channel activation | Pre-launch zeros have a different meaning | Launch dates and treatment of inactive periods |
| Baseline and priors | Flexible components can exchange explanatory weight | Sensitivity results and posterior diagnostics |
| Budget intervention | Historical association may fail under a new allocation | Held-out forecasts and a matched live test where feasible |
How should a team act on the output?
Ask the model for alternative allocations under the same total budget and operational constraints. Report expected leads and uncertainty for each scenario, the assumed cap changes, and how far the recommendation moves beyond observed spend. Choose a bounded live test when the allocation is costly or poorly represented in training data. Define the lead outcome and attribution window before inspecting results.
A message or offer choice experiment answers a different question. A randomized comparison can estimate an assigned contrast on task choices. Generated choices are modeled stated choices inside the configured simulation; they do not establish the response to changing channel spend or reach. Transfer to live buying requires matched human or field evidence. The research record and methods archive explain related validation questions.
Before cutting upper-funnel spend, establish whether the weak direct return survives a model that represents plausible mediation and cap constraints. If the answer changes across credible specifications, report that uncertainty rather than presenting a single allocation as settled.