Ranges, not points
A contribution without an interval is a guess with a confident face. Every figure shows its width.
Method
Bayesian media mix modeling and learned multi-touch attribution, fit on your own data. It refreshes nightly, refits weekly, and every figure carries its interval.
The pipeline
Every fit passes through the same six stages. Two of them are checks: one on your data before a fit starts, one on the fit before it is published. When a check stops a fit, you see why.
Weekly rows are the source of truth: one revenue column and one column per channel, from a connected source or a weekly file. Touchpoints and conversions land beside them in the same schema.
Identifiers resolve to subjects your organization owns. Suppression and erasure are built in, so every person-level row has a lawful basis.
Too few weeks, too little variation, a missing revenue column or a gap in the sequence is reported before a fit starts. Modeling begins when the data qualifies.
The media mix model refits weekly, with carryover and saturation estimated per channel from your data. With more than 60 weeks of history, it is first fit without the last eight weeks and scored on them.
A fit is published only when it reconciles to booked revenue, credits media with something, and converges. Otherwise you get the reason and the fix.
Every figure carries its interval, and the next dollar's return is ranked by channel. The simulator, the revenue graph and the forecast all read this same fit.
The output
A response curve for every channel, with its credible band. Read across it and you know what the channel returned, and what its next dollar will.
Response curves · adstock and saturation fitted per channel
Marginal return on next dollar
This ordering, not the average, is what should drive reallocation.
How much revenue comes from everything that is not media, with its interval.
Each channel's share of modeled revenue over the fit window. Never a point without a width.
How long each channel's effect lasts and where it flattens, estimated from your rows, not a benchmark.
Return on the next dollar by channel, ranked, so a reallocation has a reason before it has a slide.
The standard
Three conditions stand between a fit and your plan. When one is not met, the page names it and says what to change in the data.
The channels' modeled contribution can never add up to more than the revenue you actually booked. A fit that claims credit for money that doesn't exist is held back.
When the model hands every dollar to baseline, it has learned nothing about your channels, usually because spend barely moved. You see that, and how to vary spend so the next fit can learn.
A Bayesian fit runs several chains that have to land in the same place. Until they do, no single figure is shown, however confident it looks.
Model card
Priors, the carryover and saturation forms, the data thresholds and how each fit is scored. Nothing about how a number was produced is hidden.
| Outcome | Weekly booked revenue |
|---|---|
| Carryover | Geometric adstock, up to 8 weeks, per channel |
| Saturation | Logistic response curve, per channel |
| Seasonality | Yearly, with 52 or more weeks |
| Minimum data | 52 weeks; 104 recommended |
| Holdout | Last 8 weeks, with more than 60 weeks of history |
| Uncertainty | 80% credible interval on every figure |
Principles
A contribution without an interval is a guess with a confident face. Every figure shows its width.
Models are fit to your organization's own data, refreshed nightly and refit weekly. Never to a demo dataset.
The simulator reads the response curves, the revenue graph draws the fit, and the forecast projects it. No page quotes a number another disagrees with.
Priors, functional forms and per-fit diagnostics are published, so your data team can check the work.
Early access is by application. We set up each team ourselves.