The Methodology
how marginal works — from data to decision.
a plain-english walkthrough of the methodology, the model, and what you get at the end of each run.
two methodologies. one measurement system.
there are two ways to measure what your marketing caused. bayesian marketing mix modelling (mmm) gives you contribution at scale — decomposing revenue across your full media mix over time, accounting for brand, upper funnel, and channels that last-click makes invisible. incrementality testing gives you causation — controlled experiments that tell you what your spend actually drove, not just what was present when a conversion happened.
most brands use one or the other. marginal runs both, and connects them. incrementality test results calibrate the model. model outputs surface where the next test should run. over time the system compounds.
bayesian, open-source, and independent.
the mmm engine is google meridian v1.5.3 — an open-source bayesian model built by google's marketing science team. we use it because it's academically rigorous, publicly auditable, and has no black box. you can inspect every prior, every likelihood function, every posterior distribution.
bayesian mmm works by combining your historical data with prior beliefs about how marketing works — then updating those beliefs with your actual results. the output isn't a single attribution number. it's a probability distribution: a range of plausible contribution estimates, with a confidence level attached to each.
this is a more honest output than a point estimate. it reflects the uncertainty that's actually in your data, rather than hiding it behind false precision.
what you bring. what we handle.
to run an mmm, marginal needs: spend data by channel by week (typically 2+ years of history for reliable estimates), revenue or conversion outcome data at the same cadence, and any external factors you want the model to account for — seasonality, promotions, macro events.
you don't need a data warehouse. you don't need a data science team. we work with exports from your existing tools — google analytics, your mmp, your finance system.
for incrementality testing, we work with you to define the test design: which channel or campaign to test, which geos to use as treatment and control, what duration gives sufficient statistical power, and what outcome to measure.
the outputs — and what to do with them.
each model run produces: a channel contribution decomposition (what share of revenue each channel drove over the measurement period), a saturation curve for each channel (where you are on the diminishing returns curve), a budget optimisation scenario (what the model suggests would happen if you reallocated spend), and a recommendation on where to run your next incrementality test.
each incrementality test produces: a causal estimate of incremental revenue or conversions driven by the tested channel or campaign, a confidence interval, and a calibration update for the next model run.
the outputs are presented as a structured report plus a working file. there's no proprietary dashboard to learn. you take the findings into your existing planning and reporting workflow.
the calibration loop
mmm and incrementality testing connected in a compounding measurement system.
before and after marginal
| dimension | before marginal | after marginal |
|---|---|---|
| time to first model | months of setup with a data science team or consultancy | 2–3 weeks from data receipt to first output |
| data requirements | custom data pipelines, warehouse infrastructure, analyst resource | csv exports from your existing tools — no warehouse needed |
| budget decisions | gut feel, last-touch attribution, platform dashboards with conflicted incentives | channel contribution decomposition with credible intervals and saturation curves |
| model refresh | annually at best — static outputs that age out of date | ongoing — each run incorporates new data and incrementality results |
| accessibility | data science team required to run and interpret | guided configuration for marketing teams — no statistician needed |
frequently asked questions
the initial setup and first run typically takes 2–3 weeks from data receipt — one week for data preparation and model configuration, one to two weeks for the model run and output review. subsequent runs are faster once the pipeline is established.
the model is most reliable with 2+ years of weekly spend and revenue data. we can run with less, but we'll be transparent about what that means for confidence intervals. shorter histories produce wider uncertainty ranges — which is the honest answer, not a flaw.
not in the current version. you export data from your existing tools and we handle the rest. native connectors are on the roadmap.
your mmp tells you which touchpoints were present before a conversion. marginal tells you what caused the conversion — and what drove revenue even when no conversion was tracked (brand, upper funnel, channels that don't pixel well). they answer different questions.
geo-holdout tests require withholding spend from a defined geographic region for the test duration — typically 4–8 weeks. conversion lift tests can often run within a platform without pausing campaigns. we'll work with you to design a test that minimises business disruption while achieving statistical power.
we'll present pricing options based on your measurement cadence and scope. there's no obligation to continue — but most brands find the first run surfaces enough to make the next one an easy decision.
see it in practice — start with a free pilot run.
the first three model runs are free. bring your spend data, we'll run the model and walk you through what it finds.
start your pilot — first 3 runs free