How to choose an MMM vendor: 12 questions to ask
by
Lauren Lauth, VP of Measurement
You've got three MMM proposals on your desk. They all promise better attribution, smarter budgeting and proof of ROI. None of them sound different until you start digging, and by then you've spent four hours in demos.
MMM vendors vary widely in methodology, in what they assume, and in whether they ever check their results against reality. Pick the wrong one and you'll reallocate budget on a model that overstates channel contribution by 30 to 40 percent. That's not a hypothetical. It happened to a DTC brand selling protein powder.
1. Do you calibrate your model against incrementality tests, or just run regression?
Start here. Regression detects correlation; causal testing confirms whether that correlation is incremental lift. Ask directly: do you run geolift or holdout experiments to validate model predictions? If the answer is no, or if validation is a paid add-on, nobody is checking the core assumptions.
2. What's your minimum monthly ad spend for reliable results?
Some vendors need $500K a month. Others model effectively below that using hybrid approaches. If you're under $500K, ask whether they'd recommend starting with incrementality testing on your largest channel and using MMM to predict the smaller ones. An honest vendor will tell you testing alone covers most of your allocation decisions.
3. How much data engineering is required on my end?
Ask whether they work with your existing CDP, warehouse and analytics stack, or whether they need months of cleanup first. The better vendors pull from several sources at once pixel,aggregatedchanneldata,first\-partypixel, aggregated channel data, first\-party pixel,aggregatedchanneldata,first\-party so they depend less on pristine inputs. Six months of preprocessing is a red flag.
4. Can your model actually forecast and optimize budget?
A model that only explains the past is half a product. Ask for a live demo of the optimizer: month-level budget scenarios, marginal ROAS by channel, saturation curves showing where diminishing returns begin. If they can't model those or show projected impact on sales and ROAS, they aren't ready to touch your budget.
5. Do you account for seasonality and business cycles?
Every brand has seasonal swings. Ask whether they estimate business seasonality separately from channel impact, using two years of history. A vendor that ignores seasonality will tell you to spend into your historical down periods and miss the peaks.
6. How do you prove your model predictions match reality?
This is the question that sorts the field. The answer you want: we run incrementality tests and compare predicted lift against actual lift. Without that loop, a vendor is operating blind. It's also fair to ask whether they'll run a head-to-head against a competitor on the same historical data.
7. Do I get a dedicated data scientist and account manager?
Support matters. Ask what the partnership actually looks like. Will they recommend which channel to test first? Do they prioritize quick wins in month one and shift to longer-term optimization after? Can they update the model monthly as spend changes? "You get dashboard access" is not support.
8. What's your actual implementation timeline?
Some vendors take four months to deliver first results. Ask whether they can get you initial recommendations in four to six weeks, with monthly refinements after that. Long timelines mean you miss seasonal moments and decisions pile up behind the model.
9. How do you handle cross-channel halo effects?
If TikTok awareness drives direct traffic or helps Meta convert better, does the model see it? Ask whether they use techniques that separate channel lift from halo, or whether they treat every channel as independent. Better models acknowledge halo exists and estimate its size.
10. What if I run a test that contradicts your model?
This one reveals how confident they really are. Ask: if we run a geolift test on Meta and your model predicted wrong, how do you update it? Would you recalibrate? Strong vendors expect this and have a documented process. Defensive vendors will start questioning the test.
11. How do you handle brand vs. performance spending?
Some brands put real money into awareness. Ask whether the model can estimate brand's incremental impact on performance channels, or whether brand gets written off as unmeasurable. With the right data, it can be modeled.
12. What does the pricing actually include?
Prices run $5K to $15K a month. Get specific: initial modeling, monthly updates, optimization scenarios, dedicated team, scenario modeling, training. Then ask what sits outside that number.
How vendors actually stack up
Vendor approach | Core methodology | Validation | Minimum spend | Data requirements | Optimization |
|---|---|---|---|---|---|
Regression only | Statistical modeling | Statistical diagnostics | $500K–$1M/mo | Clean aggregate data | Channel budgeting |
Incrementality-first | Testing + modeling | Geolift validation | $300K–$500K/mo | Pixel + aggregate | Full budget scenarios |
Continuous test integration | Real-time validation | $200K–$500K/mo | Pixel + aggregate | Month-level optimization + confidence intervals |
The shortcut: ask for a validation test
If you only ask one question, make it this: can you run a pilot on our largest channel and compare your model's prediction against a real incrementality test?
Whichever vendor lands closest to the actual test result is the one calibrated to your business. That's how you de-risk the decision without sitting through another round of demos.
Frequently asked questions
How long does implementation take?
Most vendors need four to eight weeks of historical data and two to four weeks of modeling. Good ones deliver initial insights inside six weeks and refine over the following quarter. Anyone promising results in two weeks is overselling.
Should I start with testing or MMM?
Under $500K a month, start with incrementality testing on your largest channel. That covers 60 to 70% of allocation decisions with high confidence. From there you can use MMM to predict smaller channels without running additional tests, which makes the whole measurement stack faster and cheaper.
Can different vendors use completely different methodologies on the same data?
They can, and they do. One vendor's regression model might put Channel A at 20% of revenue while another's causal model puts it at 12%. That's a methodology difference, not an arithmetic error, which is exactly why the validation question matters: which one predicts lift correctly?
What happens if my largest channel isn't where most spend goes?
Test your second-largest or highest-ROI channel instead. You want a channel with enough spend to move the needle and enough volume to reach significance. Your vendor should help you pick the first test, not leave you guessing.
Ready to find the right MMM vendor? Book a demo with WorkMagic to see how incrementality-calibrated modeling works in practice.