Key takeaways
- Marketing mix modeling estimates each channel's contribution from aggregate weekly data. An incrementality test measures one channel's contribution directly with a holdout. One is a model, the other is a measurement.
- MMM is worth building when spend is spread across many channels, some of them hard to track, and you have 2 to 3 years of weekly data with real variation in it.
- I treat holdouts as the ground truth and use them to calibrate whatever model a team trusts, MMM or attribution.
- Below that scale, holdouts plus customer-level or deal-level data answer the budget question faster.
01What marketing mix modeling is
Marketing mix modeling (MMM) is a statistical method that estimates how much each marketing channel contributed to sales, using aggregated weekly or monthly data on spend, sales and outside factors like price, seasonality and promotions. Teams use it to split budget across channels and to forecast what a different mix would return, without tracking individual users.
MMM, also called media mix modeling, is old. Consumer brands have used it for decades to plan TV and print. It came back because privacy changes left user-level tracking less complete, and MMM never needed user-level tracking in the first place.
Under the hood it's a regression of your sales on spend by channel, with 2 adjustments that make it marketing-shaped. Carryover, often called adstock, lets an ad keep working for weeks after it runs. Saturation lets each extra dollar in a channel return less than the one before it. The output is a contribution and ROI by channel, a response curve for each and a suggested budget split.
02MMM vs multi-touch attribution vs incrementality
These are 3 different instruments. Most arguments about which one is right come from asking one of them to do another's job.
| Marketing mix modeling | Multi-touch attribution | Incrementality testing | |
|---|---|---|---|
| Question it answers | How much did each channel contribute to sales over time | Which tracked touchpoints came before each conversion | What did this channel cause, compared with not running it |
| Data | Aggregate weekly spend, sales and controls | User-level clicks and views | A test group and a matched holdout |
| Method | Statistical model fit to history | Rules or models that split credit | Controlled experiment |
| Causal | Estimated, and only as good as the model | No, it assigns credit | Yes, within the test's scope |
| Sees untracked channels | Yes, including TV, audio and offline | No | Yes, if you can hold the channel out |
| Granularity | Channel, sometimes campaign | Campaign, ad and keyword | Whatever you held out |
| Speed | Weeks to build, refreshed monthly or quarterly | Daily | Weeks per test |
| Weak spot | Channels that always move together, missing controls, false precision | Credits conversions that would have happened anyway | One channel and one period at a time |
| Best for | Annual and quarterly budget split across channels | Daily optimization inside a channel | Settling a specific budget decision and calibrating the other two |
03Where I stand
My measurement work has been on the experiment side. At Shopify I ran paid merchant acquisition, $20M+ a year across 6 channels, and scaled it 4x at a 3.2x ROAS. With Data Science I ran geo holdouts across paid social that found $2M+ of spend that wasn't adding merchants, and the budget moved to channels that were.
That kind of finding is hard to get from a model alone. A holdout compares 2 groups where the only planned difference is the spend. A model infers the same answer from history, and history is full of things that moved at the same time.
So my rule is plain. Any model is a hypothesis about what caused sales. A holdout is a measurement. I use holdouts to calibrate whatever model a team already trusts, and I trust the model more for each channel where it agrees with a test. The mechanics of running one are in how I run incrementality tests, so I won't repeat them here.
04When MMM is worth building
I look at 5 things before recommending a model. If most of them land in the right column, skip the model for now.
| Signal | Build an MMM | Use holdouts and customer data instead |
|---|---|---|
| Channel count | 5 or more channels, some of them hard to track, like TV, audio, creators or paid social views | 2 to 4 channels that are mostly click-trackable |
| Data history | 2 to 3 years of weekly data on a business that hasn't changed shape | Less than 2 years, or a recent change in product, pricing or market |
| Spend variation | Channels moved independently, and some went dark for weeks | Every channel flat, or all of them rising and falling together |
| Budget size | A 10% shift between channels is worth more than building and maintaining the model | A 10% shift wouldn't pay for an analyst's quarter |
| Sales cycle | Many transactions a week, short to medium cycle | Few, long B2B deals you can trace one by one |
The last row matters more than people expect. At Pipe17, a B2B company with long sales cycles, the right tool was the CRM. A six-month deal-level audit proved Google Ads at a 4.0x ROAS and cut the keywords that lost money. Deal-level evidence also showed a six-figure LinkedIn line had produced no closed revenue, and it was cut. A model would have needed years of data to say less, less confidently.
05What an MMM needs
- An outcome that pays. Revenue, paying customers or gross profit by week. A model fit to leads or signups will tell you which channels buy cheap leads.
- Spend by channel by week. Impressions too where you have them, and by region if you can. Regional data gives the model many more data points from the same weeks.
- Controls. Price changes, promotions, launches, seasonality, holidays and anything else that moved sales without marketing.
- Variation. If 2 channels always rise and fall together, no model can separate them. Sometimes the first step is to change spend on purpose so the next model has something to learn from.
- An owner and a refresh schedule. A model built once and never refit drifts away from the business within a few quarters.
- Experiments to calibrate it. At least one holdout on each big, uncertain channel.
Open-source options
2 open-source packages are worth knowing. Robyn comes from Meta's Marketing Science team and is written in R, with a beta Python interface. Meridian comes from Google, is Bayesian and supports regional data. Both are built to take experiment results: Robyn scores candidate models on how closely they match your lift and geo tests, and Meridian turns past experiments into priors on channel ROI. Google's documentation for Meridian suggests at least 2 years of weekly data for regional models and 3 for national ones.
Both come from companies that sell ads. The code is open and inspectable, which helps, but I'd still check any model's answer for a big platform against a test you designed yourself.
06How holdouts and MMM work together
- Find where the model is least sure.Rank channels by spend times uncertainty. A big channel with a wide range on its ROI is the first test.
- Run a holdout built for that decision.A geo holdout usually, because it measures the whole channel effect and matches the way an MMM sees the world.
- Feed the result back.Load it into the model as a calibration input or a prior, then refit.
- When they disagree, believe the test and debug the model.The usual causes are a missing control, 2 channels that moved together or a test window too short to catch carryover.
- Use each tool at its own altitude.The model splits budget between channels. Inside a channel, platform bidding does the work, and it works better when you feed it the right events, which is what value-based bidding covers.
- Retest when the inputs change.A new spend level, new creative or a new market can change a channel's return. Retest the channels that moved and refit the model each quarter.
07Point every model at the customer who pays
Whichever method you pick, the outcome you feed it decides the answer. When I moved bidding and reporting to paid conversions on Shopify's $1 trial, signup CPA rose about 20% and cost per paying merchant fell about 30%. A model fit to signups would have scored that change as a loss.
The dashboard we allocated budget with went further. It joined ad data to merchant GMV at 30, 60, 90 and 180 days by campaign. A channel that brings in merchants who stop selling by day 90 looks fine in any model fit to first conversions. That's also why I ran 40+ structured experiments a quarter, each changing one variable: a clean test on the right outcome beats a sophisticated model on the wrong one.
08Mistakes I see with MMM
- Shipping a model nobody has checked against a test. Different model settings can fit history equally well and disagree on channel ROI. A holdout is how you choose between them.
- Reading average ROI as marginal ROI. The average says what the channel returned. The response curve says what the next dollar will return, and that's the number a budget decision needs.
- Too many channels for the data. Splitting paid social into 8 line items with 2 years of national data gives you confident noise.
- Treating it as a replacement for attribution. MMM won't tell you which keyword or ad to cut this week.
- Building it once. The model is a process with an owner, a refresh date and a test calendar.
09What I'd do at most companies
For a company spending on 2 to 4 trackable channels, I start with clean conversion data that reaches the customer who pays, bidding on that data, and one holdout a quarter on the biggest line nobody has proved. I add an MMM when untracked spend and channel count grow past what tests alone can cover, and I calibrate it with the tests already run.
That sequence is the core of my work as a performance marketing consultant, and the wider version, across paid, search and lifecycle, is what I do as a growth marketing consultant.
FAQQuestions people ask
- What is marketing mix modeling?
- Marketing mix modeling (MMM) is a statistical method that estimates how much each marketing channel contributed to sales, using aggregated weekly data on spend, sales and outside factors like price, seasonality and promotions. Teams use it to split budget across channels and forecast what a different mix would return.
- What is MMM in marketing?
- MMM stands for marketing mix modeling, also called media mix modeling. It works on aggregate data instead of tracking individual users, so it can measure channels like TV, audio and paid social views that click-based attribution misses.
- What is the difference between marketing mix modeling and multi-touch attribution?
- Multi-touch attribution splits credit for each tracked conversion across the clicks and views before it, using user-level data. Marketing mix modeling estimates each channel's contribution from aggregate spend and sales over time, so it covers untracked channels but answers at the channel level, not the ad level.
- Is MMM or incrementality testing better?
- They do different jobs. An incrementality test measures what one channel caused, in one period, with a controlled holdout, while MMM estimates every channel at once from history. The strongest setup uses holdout results to calibrate the model.
- How much data do you need for marketing mix modeling?
- Google's Meridian documentation suggests at least 2 years of weekly data for geo-level models and 3 years for national models. The data also needs real variation, with channels whose spend moved independently of each other.
