Reading an MMM
Contribution and ROI.
Reading an MMM is a free Digital Marketing Academy lesson on CoddyKit — lesson 3 of 4. You can read the complete lesson below for free — then practise it hands-on in the browser with a built-in code editor and a 24/7 AI tutor. It is part of the Digital Marketing Academy learning path, one of 4 lessons in the course, and your progress syncs across the web and the CoddyKit app.
From Coefficients to Decisions
A fitted MMM hands you coefficients, contribution charts, response curves, and fit statistics. Reading them well turns math into budget decisions.
This lesson walks through the outputs an analyst actually scrutinizes before trusting a model.
The Contribution Decomposition
The headline output splits total sales into base and each channel's incremental contribution over the period.
Read it as: of all units sold, how many would have happened anyway (base) and how many each channel added. Sum of parts should reconcile to actual sales.
channel contribution % of sales
base 1,400,000 70.0%
TV 280,000 14.0%
Search 180,000 9.0%
Social 100,000 5.0%
OOH 40,000 2.0%ROI / ROAS per Channel
Divide a channel's incremental revenue by its spend to get ROI (or ROAS). This is the comparable currency across channels.
Watch the difference between average ROI (whole period) and marginal ROI (the next dollar). Allocation decisions hinge on the marginal figure.
channel spend inc_rev ROAS
Search 60,000 300,000 5.0
TV 140,000 560,000 4.0
Social 50,000 150,000 3.0
OOH 40,000 80,000 2.0Average vs Marginal ROI
Because of saturation, a channel with high average ROI may be nearly maxed out, so its next dollar returns little.
Marginal ROI is the slope of the response curve at current spend. Reallocate toward channels with the highest marginal, not highest average, ROI.
Search avg ROAS = 5.0
but spend sits near saturation:
marginal ROAS = 1.8 (next $ underperforms)
Social avg = 3.0, marginal = 3.4 -> grow hereReading the Response Curve
Each channel's response curve plots incremental sales against spend. The steep early region is efficient; the flat region is saturated.
Find current spend on the curve. If you're on the flat part, more budget is wasted. If on the steep part, there's headroom.
spend -> 10k 20k 40k 80k 160k
sales -> 50k 90k 140k 165k 175k
slope -> high ... ... low ~flat (saturated)Coefficient Signs and Magnitudes
Sanity-check signs first. Media coefficients should be positive; price usually negative; promo positive. A negative TV coefficient signals collinearity or a model error.
Magnitudes should be plausible against priors and benchmarks, not just statistically significant.
b_TV = +0.42 (ok, positive)
b_search = +0.31 (ok)
b_price = -1.20 (ok, demand falls w/ price)
b_social = -0.05 (red flag: investigate)Fit and Validation Stats
R-squared shows in-sample fit, but a high value can mean over-fitting. Trust out-of-sample metrics more: MAPE on a holdout period and time-series cross-validation.
Also inspect residuals: they should look like noise with no leftover trend or seasonality.
R2 (train) = 0.94
MAPE (holdout) = 7.8%
Durbin-Watson ~ 2.0 (no autocorrelation)
residuals : random, mean ~0Uncertainty Intervals
Point estimates lie. Bayesian MMMs report credible intervals; frequentist ones report confidence intervals. A channel whose interval crosses zero is not reliably positive.
Present ranges to stakeholders so decisions account for how much the model actually knows.
channel ROAS 90% interval
Search 5.0 [4.1, 5.8] confident
TV 4.0 [2.9, 5.1] confident
OOH 2.0 [-0.3, 4.2] weak / uncertainDecomposition Over Time
Beyond the period total, view contributions week by week. This reveals whether a channel's effect is steady, spiky, or fading.
Spikes that align with flights build confidence; contributions that appear with no spend suggest leakage from a missing control.
Calibrating Against Experiments
The strongest validation is comparing MMM channel lift to a randomized geo or conversion-lift experiment.
If MMM says Search ROAS is 5 but a geo test says 3, recalibrate, often via a prior. Models that match experiments earn trust for the channels never tested.
channel MMM_ROAS experiment_ROAS
Search 5.0 3.0 -> recalibrate down
Social 3.0 3.2 -> well calibratedRed Flags Summary
Distrust a model with implausible signs, a base that is too small, ROIs that defy benchmarks, wide intervals presented as certainty, or great in-sample fit but poor holdout.
Reading an MMM is as much skepticism as arithmetic.
Quick Check
Apply the average-vs-marginal distinction to an allocation call.
Recap
Read an MMM through its decomposition (base vs incremental), per-channel ROAS, and especially the response curves that reveal marginal returns.
Check coefficient signs, holdout fit, residuals, and uncertainty intervals, and calibrate against experiments. Decisions follow marginal ROI, not headline averages.
Frequently asked questions
Is the “Reading an MMM” lesson free?
Yes — the full text of “Reading an MMM” is free to read here on the web, and the Digital Marketing Academy course includes 4 lessons in total. To practise it interactively (a built-in code editor and a 24/7 AI tutor) and unlock the rest of the Digital Marketing Academy course, upgrade to CoddyKit PRO.
What will I learn in “Reading an MMM”?
Contribution and ROI. You practise Digital Marketing Academy with hands-on code you run directly in the browser, and a 24/7 AI tutor answers your questions as you work through the lesson.
Do I need any experience to start Digital Marketing Academy?
No prior experience is required. Digital Marketing Academy on CoddyKit is structured for beginners through advanced learners; this is — lesson 3 of 4, so you can start here or from the beginning and move at your own pace.
How long does the “Reading an MMM” lesson take?
Most CoddyKit lessons take about 5–10 minutes. Each one is bite-sized and interactive, so you make steady progress and pick up exactly where you left off across the web and the app.
Can I write and run code in this Digital Marketing Academy lesson?
Yes. Every Digital Marketing Academy lesson includes a built-in code editor, so you write and run real code right in your browser and get instant AI feedback — no local setup required.