Sundar’s experiMENTAL
Hello experiMENTAList, it’s Sundar 👋
I’m a former Head of Marketing Science at Uber where I optimized $1Bn+ in spend across Brand, Performance, and Lifecycle. Now, I share weekly playbooks that help you prove and scale your Marketing ROI.
If you know a colleague or friend that might benefit from this newsletter, send them here.
Now let’s get experiMENTAL!
How Marketers Are Scaling With AI in 2026
61% of marketers say this is the biggest marketing shift in decades.
Get the data and trends shaping growth in 2026 with this groundbreaking state of marketing report.
Inside you’ll discover:
Results from over 1,500 marketers centered around results, goals and priorities in the age of AI
Stand out content and growth trends in a world full of noise
How to scale with AI without losing humanity
Where to invest for the best return in 2026
Download your 2026 state of marketing report today.
What your MMM is actually telling you
Every marketer has heard of an MMM, but few actually understand what it's doing under the hood. More importantly, it gets pitched as the holy grail of measurement, but that’s not the truth.
The reality is that an MMM is just one piece of a bigger puzzle.
What actually goes into an MMM
For MMMs “Garbage in, garbage out” isn't a cliché. It really can make or break one.
A good MMM needs:
Historical spend by channel → weekly or monthly, going back a few years
Note: How much data does an MMM need is a common question but there’s no one true answer. The rough estimate is at least 2 years of weekly data but more importantly you need volatility in the spend
Sales/revenue or whatever KPI you're modeling
Non-marketing variables that explain the "why" outside of marketing → Eg. seasonality, price, promotions, distribution, competitor activity, macro factors like weather
The quality of this input data is the ceiling on how good your output can be. Sorry, but you cannot model your way out of bad inputs.
What the model is actually doing
At its core, an MMM is a regression. Now for those that are not well versed in models / statistics, don’t let this scare you. Let’s go back to high school algebra and a simple formula you might remember:
y = mx + b
Of these 4 variables, in MMM world, we know 2 things:
Sales Variable or whatever KPI you picked (y)
Inputs (x)
We know these, because that’s what we’re giving to the model. It’s then the MMM’s job to spit out two things:
Importance (m)
Baseline (b)
Essentially, an MMM tells you how important each of your marketing channels ( inputs) are at predicting the sales variable (output). It also tells you what it believes the baseline amount of the KPI you’re looking to understand.
But, what exactly is a baseline?
There are multiple ways to explain what this means:
You could look at it as if you turned off marketing right now, what your organic demand would still drive.
Now, this is not completely accurate because your baseline is built up over all the actions you’ve taken over time. Baseline for a company 5 years old is different than 100 etc. So this definition of baseline works in the short time but not in the long run.
You can look at it as the error in the model.
This is not completely accurate because “error” in a model is highly dependent on the inputs. What might seem like error could be that you haven’t added the right variables.
A crazy example would be an MMM used to predict ice cream. Now, if for some reason you forget to included seasonality then your baseline is going to look pretty high.
Both are neither right nor wrong because an MMM is just a model and models have errors. But if an MMM were just a simple regression, then you wouldn't have a billion dollar industry where SaaS companies, agencies, and data scientists were all tryin to pitch you to use theirs.
The complexities in MMMs result from a few additional factors:
Adstock/carryover → marketing spend today has effects that decay over the following weeks, not just in the week it ran.
An MMM needs to understand this dynamic.
Saturation curves → aka diminishing returns. The same channel gets less efficient the more you pour into it.
An MMM needs to understand this dynamic.
Interaction effects → You’l hear terms like multicollinerarity and endogeneity and basicaly these are factors that you have to account for in the model to ensure the model isn’t spitting out a corrupt model.
For example, if two channels move together (you increase spend in Channel A and B at the same rate) then the model will have trouble figuring out which channel specifically is more impactful.
On top of all this, an MMM’s job is to also tease out the relationship with non Marketing variables as well so all of this creates a complex regression, but let’s not digress.
What the outputs mean
Every MMM should spit out a few key outputs:
Channel contribution
ROI Curves
Baseline
Channel contribution → how much of your KPI each channel is estimated to have driven historically
ROI Curves → what marginal return looks like at different spend levels. This is the actionable part for budget decisions
Baseline → The baseline is as I mentioned above what the model can not explain through modeling.
So, what do we do with these numbers?
How to use an MMM
The most important thing to remember is that an MMM is part of a broader measurement strategy generally called triangulation (MMM, attribution, incrementality testing) that gets you to the holy grail.
Step 1: Roadmap
Once you get the outputs (what channels the MMM thinks are underperforming or overperforming), then, your job as a Marketer is to take that output and build a testing roadmap from it.
Effectively, you must validate what your MMM says about channel performance and ROI curves with a real world test where possible.
Note: Often you can run a test where you cut budget from the channels it flags as underperforming and reinvest into the ones it flags as overperforming. You stay budget neutral, but you're building a more optimized portfolio mix.
Step 2: Test
Self explanatory
Step 3: Triangulate
Remember where I keep bringing up triangulation? This is where it comes in. An MMM will tell you a channel's true performance, but you can't use an MMM to make daily decisions. For daily operations, most teams use some form of attribution.
So the third step after the MMM is to take the outputs of the MMM and have those results align with your attribution, usually through the form of some multipliers.
A simple example would be: say the MMM says Meta drove 5,000 orders, but your last click says 10,000 orders. You would then put a 0.5x multiplier on the last click attribution to align what your attribution says to what the MMM believes is true.
Step 4: Refresh
This is where the MMM actually becomes powerful. Once you run these tests, you can feed them back into the model to make it smarter. Remember, an MMM is a historical view. Yes, it can be used for forecasting, but that forecast assumes the relationships it learned will largely hold. The more you test and feed those results back in, the more current the model stays, and the tighter the gap between the model and your forecast.
A question that often comes up is “How often should I update my MMM?” How often you refresh it comes down to how quickly you make decisions and how quickly your business moves.
If you're not running promotions frequently, not changing channels often, not experimenting constantly, refreshing every six months is fine. If you're a more frequent e-commerce type business with constantly changing promotions and stock, look at quarterly. Uber did quarterly, and it was enough for our business, even with daily and weekly promotions running.
Aim for 1x every 6 months and quarterly if your operations can sustain it.
Where it goes wrong
There's 2 things Marketers most often forget:
An MMM is a model
It is not running an experiment. It needs to be triangulated against experiments, geo-lifts, incrementality tests, to actually be trusted.
The model is fitting a curve to history
It is a snapshot of history and not a set it and forget it.
In addition, an MMM is a model and that means confidence intervals matter. They’re often the first thing stripped out when a crappy agency does an MMM but a channel's contribution range can be wide enough to completely change the decision you'd make off it.
For example, your MMM could tell you that Meta is incrementally driving 12 ± 6%. That incrementality is very different at the lowest vs highest range and could have a very different impact on your business.
Always ask for numbers with confidence intervals.
Here are a few other pitfalls:
Agencies tuning the model to confirm the story the CMO wants to hear, or that happens to justify the agency's own channel
Overfitting is a real thing. This one is hard to spot without a data scientist but also question the methodology. Anytime the answer is a pure black box you’re likely in trouble.
Treating a single MMM run as gospel instead of a directional input that needs updating and validating
An MMM is not a crystal ball. It’s more like a crystal vase. Just kidding. That’s a terrible analogy but I’m just seeing if you’re paying attention.
An MMM only gets better with time so use it to set a general direction but then validate.
Not asking what variables were excluded. An incomplete model will misattribute lift to whatever's left in it
So remember, an MMM is above all a model and just part of a broader measurement strategy.
Was this article helpful?
Need 1:1 help solving your Marketing Science problem? Let’s chat >
That’s it for this week!
Stay experiMENTAL,
Sundar
