- MMM regresses outcomes against spend history to estimate each channel's contribution — no pixels, no cookies, privacy-proof by construction.
- It needs what small brands don't have: 18–24+ months of history, real spend variation, and multiple meaningful channels. Feed it two channels of steadily-rising spend and it returns your inputs with confidence intervals.
- The honest ladder: reconciliation → incrementality tests → MMM. Most Indian D2C brands under ₹20–30L/month multi-channel spend should stop at step two.
Media Mix Modeling is the oldest measurement idea in the room — P&G was doing it before the internet — and it's back in fashion because iOS broke the pixel and MMM never needed one. The revival is real, the open-source tooling (Meta's Robyn, Google's Meridian) is genuinely good, and the vendor pitches have gotten far ahead of what the math can deliver at D2C scale. Here's the sizing truth.
What MMM actually does
Feed it weekly history — spend per channel, revenue, plus controls (seasonality, promos, price changes) — and it fits a model estimating how much each channel contributed, including saturation curves (diminishing returns) and adstock (delayed effects). Output: contribution shares, marginal ROAS per channel, and budget-reallocation suggestions. No user-level tracking anywhere.
What it needs to not hallucinate
| Requirement | Why | Where brands fall short |
|---|---|---|
| 18–24+ months of weekly data | The model learns from variation over time | Brand is 14 months old; tracking changed twice |
| Spend variation | If spend never moved independently, channel effects can't be separated | Spend only ever went up, on all channels together |
| 3+ material channels | Two channels = two knobs; the model mostly restates their correlation | 85% Meta, 15% Google |
| Stable definitions | Revenue must mean one thing across the whole window | Gross in the old sheet, net in the new dashboard |
The second row is the silent killer. MMM learns from differences — weeks where Meta rose while Google fell. If your history is "everything up and to the right together," the model cannot mathematically separate the channels, so it leans on its priors. And its priors were set by whoever configured it. That's how a mix model becomes a very expensive mirror.
The scale test, in one table
| Monthly ad spend | Channels | What measurement stack fits |
|---|---|---|
| < ₹5L | 1–2 | Reconciliation + weekly MER vs break-even. Anything fancier measures noise. |
| ₹5–20L | 2–3 | Add brand-pause and one geo holdout per quarter. iROAS deflators do MMM's practical job. |
| ₹20L–1Cr | 3+ | MMM starts to add signal — if the history and variation exist. Calibrate it with your geo tests. |
| > ₹1Cr | 4+ | MMM quarterly, incrementality to calibrate, reconciliation always-on. All three, each doing its own job. |
If you do run one: the three sanity checks
- Does it reconcile to the bank? Channel contributions must sum to something consistent with actual revenue. A model crediting channels with 130% of reality has already told you what it's worth.
- Does it disagree with anything? A model that confirms every existing belief wasn't identified — it was decorated. Real models produce at least one uncomfortable number.
- Was it calibrated against an experiment? Modern practice anchors MMM with incrementality results (Robyn and Meridian both support this). A model with no experimental anchor is a hypothesis, not a measurement.