Math matters.
Here is what we bet on it.

Identifiability fails because:

1. Cross-channel interactions are unmodeled.

2. Measurement bias propagates from the additive baseline.

3. Forecast cascade compounds at peak.

4. Halo stays invisible by construction.

5. Standard models cannot estimate the cost of stopping a channel.

One architectural cause beneath all of them: additive math, patched after the fact with attribution rules, cannot represent the structure of a real multi-channel business.

What we built

The model is the product of six years of work, PhDs in our research team, twenty-five million dollars invested. It is mechanistic. The math under it captures the actual structure of how marketing channels interact with each other and with the rest of a brand: seasonality, halo, funnel directionality, time-varying efficiency, cross-channel and cross-surface dynamics. Everything jointly modeled, with priors from preceding posteriors so the model learns continuously and stays stable as new data arrives.

The architecture pairs the channel model with a neural network for every campaign, trained jointly. That is how we get forecasts at the unit managers actually allocate against, beyond the channel level the additive math was confined to.

The system has been trained on more than seven billion dollars in observed ad spend across more than one hundred and fifty brands. Every refresh runs on the brand’s own data.

One model, built once, continuously learning.

One model. Built once.

Every standard measurement stack stitches three or four products together: an MMM, an attribution layer, a forecasting tool, an optimizer. Each one is calibrated separately. Each one carries its own structural assumptions.

We built one system that does all four, on the same equations, with the same priors. The map beats the patch.

What it does that nothing else does

The same architecture answers every question I walked through across these eight weeks, jointly, in the same equation.

Identifiability is anchored structurally. Confidence intervals run 10 to 20 percent around point estimates, where additive baselines typically run 50 to 100 percent.

Measurement is windowless and probabilistic. No click windows, no view-through pixels, no per-user identifiers. The model stays valid across privacy regimes, cookie deprecation, and platform reporting changes.

Forecasts are joint and structural. Campaign-level RMSE drops 14 to 21 times relative to standard additive baselines. The system predicts what the manager actually decides on, at the unit they actually allocate against.

Halo is captured by construction. Cross-channel and cross-surface effects live inside the equation, written into the structure from the start. The model can defend the bet to scale a channel and the cost of stopping it, with the same math.

Optimization runs on the same model. The Spend Optimizer is the planning surface for what the model already knows. There is no separate optimizer to calibrate against the measurement.

The proof

Across ten ground-truth datasets in the OMEN paper, the model cut causal-attribution RMSE by 70 percent versus open-source baselines and dropped peak-period optimization regret from 32 to 45 percent down to 5.6 percent.

Looking at joint structural campaign forecasting, RMSE dropped 14 to 21 times relative to additive baselines.

Over more than 150 brands and seven billion dollars in observed spend, we have run head to head against the rest of the category in data-science-led RFP evaluations. We are 8 for 8 undefeated.

The customer results across this arc share a pattern. Coterie cut Linear TV CAC 22% on the same spend. Global Healing inverted YouTube from #9 to #1 in modeled rank and scaled it 8.4x. Beekman 1802 scaled TV 2.5x with media ROAS up 11%. DECKED found its dashboard-ranked-last awareness tactic was 4x more efficient than retargeting. Catalina Crunch dropped CAC 24% in 30 days. MaryRuth scaled Pinterest 8x with CAC nearly 50% more efficient than the portfolio.

Same model, different brands. The architectural cause behind every win is the same.

The bet

This is what we bet on.

The category will move from additive math and patched attribution toward mechanistic, joint, campaign-level architectures. And the migration is already running. MMM adoption is up 212% year over year. Gartner has stood up a Magic Quadrant for the space. Forrester lists MMM in its top five technologies to deploy.

Brands that move first are the brands that get the budget defense, the forecast that holds, and the optimization that does not require translation.

We built the architecture before the category was ready for it. The category is now moving.

If you have read this far, you already see the math. The eight-week arc is closed. The next step is to see what the system looks like on a commerce mix like yours.

Book a demo