Most MMM vendors and internal data teams are better at building models than explaining them. You get a presentation with 30 slides, charts with unfamiliar axes, and a summary that says "Facebook has a 2.1x ROI and TV has a 1.4x ROI." What does that actually mean? How confident should you be in it? What should you do differently based on it?
Here's how to read the key outputs, and what to push back on.
MMM output has three key components: contributions, response curves, and saturation points
The decomposition chart
This is the most important chart in any MMM output. It breaks your total sales during the modeled period into components:
- Baseline: Sales that would have happened without any advertising, driven by organic demand, repeat customers, word of mouth, brand equity, and other factors the model can't attribute to media
- Media contributions: What each channel added above baseline
- Other factors: Seasonality, promotions, price changes, competitor activity, macroeconomic factors
A healthy baseline is typically 40–70% of total sales. If the model says your baseline is 90%, your advertising might not be contributing much, or the model may be miscalibrated. If the baseline is 20%, something is likely wrong with the model specification: it's overcrediting media and undercounting organic demand.
Ask your vendor: "What is our baseline contribution, and what does that tell us?" If they can't answer this clearly, probe further before accepting the other numbers.
ROI vs. mROI (marginal ROI)
These are two different numbers, and they're not interchangeable.
Total ROI: The return generated by your entire spend in a channel over the modeled period. If you spent $1M on Facebook and the model attributes $3.2M in sales to that spend, your total ROI is 3.2x.
Marginal ROI (mROI): The return on the last incremental dollar you spent in that channel. Because of saturation, the marginal dollar in any channel always returns less than the average dollar. If you've been heavily investing in Facebook, your mROI might be 0.8x even if your total ROI is 3.2x.
The number that matters for budget decisions is mROI. Total ROI tells you how a channel has performed historically at its current spend level. mROI tells you what would happen if you added (or removed) another dollar.
If Facebook's total ROI is 3.2x but its mROI is 0.8x, each additional dollar in Facebook is returning 80 cents. That's not a reason to scale Facebook, it's a signal that you're past the efficient range and should be reallocating budget.
Most good MMM outputs will show you both, or at least a budget optimization curve that shows how ROI changes at different spend levels.
Saturation curves
Saturation curves show the relationship between spend level and contribution for each channel. The y-axis is contribution to sales; the x-axis is spend. The curve flattens as the channel becomes saturated.
What to look for:
Channels with steep curves where you're underspending: The curve is still rising sharply at your current spend level, which means additional spend in this channel would return well. This is an opportunity.
Channels with flat curves where you're already at high spend: The curve has leveled off, meaning you're in diminishing returns territory. These channels are candidates for budget reallocation.
A channel can have a high total ROI and still be flat on its saturation curve, because all that historical spend happened when the curve was steeper. The curve tells you about the marginal efficiency now, which is what matters for your next budget decision.
Adstock and carryover
Most MMM outputs show a carryover (adstock) rate for each channel, how long the effects of advertising persist after you stop spending.
A TV adstock of 6 weeks means that if you stop spending on TV today, the sales lift from last week's TV investment will continue (at a declining rate) for roughly another five weeks. The effect doesn't cut off immediately when the spend stops.
This is important for two reasons:
First, it explains why turning off a channel doesn't immediately collapse sales. If your MMM shows TV has a 6-week carryover and you pause TV in January, you might not see the full impact until March. Teams that pause a channel and see "no change" in the first few weeks often incorrectly conclude the channel wasn't driving anything. The carryover delays the impact.
Second, it explains why a short-duration media blitz can still generate value over weeks afterward. The efficiency of your timing decisions depends on understanding the carryover of each channel.
Ask your vendor how they estimated the carryover rates, and whether these are informed by external research (academic literature on adstock by channel type) or derived purely from your data. Carryover estimates from sparse data can be noisy.
Confidence intervals
A well-specified MMM output will show uncertainty around each estimate, not just a single point estimate.
"Facebook ROI is 2.1x (95% CI: 1.4x–3.2x)" is meaningfully different from "Facebook ROI is 2.1x." The first tells you how confident the model is. The second presents false precision.
Wide confidence intervals mean the model is uncertain about that channel's contribution, usually because there isn't much variation in spend for that channel over the historical period. If you spent roughly the same amount on YouTube every week for two years, the model has limited ability to separate YouTube's effect from everything else that was constant. The confidence interval will be wide.
Narrow confidence intervals on a channel where you've varied spend significantly are more trustworthy.
If your vendor gives you point estimates with no uncertainty ranges, push back. Ask to see the confidence intervals. If they tell you "our model doesn't produce confidence intervals," that's a frequentist vs. Bayesian distinction, and it's a signal to ask harder questions about model specification and reliability.
Budget optimization recommendations
Most MMM vendors will include a recommended budget allocation based on the model. This shows where you're over- and under-invested relative to the point of efficient returns on each channel's saturation curve.
Treat these recommendations as directionally useful, not prescriptive. The model is based on historical data under historical conditions. Market dynamics, creative quality, and audience saturation all change. A recommendation to double your YouTube spend assumes the incremental yield will match what the historical data suggests, which is not guaranteed.
Use budget optimization outputs to identify the direction of change (less Facebook, more YouTube, test podcast) and the rough magnitude. Validate the directional shift with incrementality tests on the channels you're scaling.
Red flags in an MMM output
A few things should give you pause:
A channel you know was running shows zero or negative ROI. This can happen due to multicollinearity (the channel's spend moved in lockstep with another variable, so the model can't separate them), poor data quality, or model misspecification. It's worth investigating before acting on.
Implausibly high ROI for any single channel. ROI above 10x for a digital channel should raise questions. Either the channel is genuinely exceptional, or the model is overcrediting it due to correlation with another driver.
No confidence intervals. See above.
A baseline contribution far outside the 40–70% range without a clear explanation (very new brand with no organic demand, or a business where brand equity does almost no work).
Poor model fit. Ask for the in-sample fit statistic, typically expressed as R-squared or MAPE (mean absolute percentage error). An R-squared below 0.85 or a MAPE above 10–15% suggests the model isn't explaining your sales patterns reliably.
Questions to ask your vendor or data team
- What are the confidence intervals on each channel's ROI estimate?
- How does the model handle multicollinearity between channels that move together?
- What external factors did you include, seasonality, promotions, pricing, macroeconomic variables?
- What is the model's in-sample fit (R-squared or equivalent)?
- How were the adstock/carryover rates estimated, and what is the uncertainty on those estimates?
- Have these outputs been validated against any incrementality data, and if so, where do they agree and where do they diverge?
The last question is the most important. An MMM output that hasn't been calibrated against any external evidence is an unchecked model. Good vendors will run incrementality studies alongside the MMM and reconcile the two. If yours hasn't done this, it's worth raising.