What is the difference between MMM and attribution?
MMM is a statistical model over time. Attribution is a set of rules over observed touches. MMM takes weekly spend or impressions per channel, plus control variables, and estimates how much of your result each channel caused. Attribution takes the links, codes and visits it can see, and gives credit to them.
So the two answer different questions. MMM asks how the budget split should change. Attribution asks which link, creator or ad did better than another one.
| Marketing mix modelling | Attribution | |
|---|---|---|
| Input | Weekly spend or impressions per channel, a KPI and control variables | Visits, clicks, codes and conversions per person or per link |
| Output | An estimated effect for each channel | Credit for each touch, link or campaign |
| Sees untracked effects | Yes, if the model is right | No, only observed touches |
| Data needed | Two to three years of weekly data | Tagged links from day one |
| Best question | How should I split the budget? | Which ad, post or creator worked better? |
Sources: Google Meridian documentation, Collect and organize your data; Manalens attribution documentation.
How much data does an MMM need?
Years of it. Google's Meridian guide says the history should be a minimum of two years of weekly data for geo-level models and three years for national-level models. With only monthly data, it recommends at least three years. Meta's Robyn guide says an MMM will need a minimum of two years of historical weekly data.
The model also needs enough data points per thing it estimates. Robyn recommends 1 independent variable for each 10 observations. Meridian gives an example: 12 media channels, 6 controls and 8 knots make 26 parameters, and two years of weekly data give 104 data points. That is 4 data points per parameter, and the guide then shows ways to raise that ratio.
| Requirement | Meridian (Google) | Robyn (Meta) |
|---|---|---|
| History | Two years weekly (geo), three years (national) | Two years weekly at minimum |
| Granularity | Weekly; geo level is best practice | Weekly as best practice; daily needs extra checks |
| Data points per variable | About 15 per parameter gives directional results in its example | 1 variable for each 10 observations |
| Inputs | KPI, media spend, an exposure metric, controls | Spend per channel, digital split by channel or publisher |
Sources: Google Meridian documentation, Collect and organize your data and Amount of data needed; Robyn documentation, Analyst's guide to MMM.
Why is MMM hard for a studio with one or two channels?
Because a small studio has too little history, too few channels with steady spend, and too many one-off spikes. A game often has a Steam page for one to two years before launch. Paid spend, if any, comes in short tests. Most wishlists come from beats such as Next Fest, a creator video or a sale.
Our arithmetic: one year of weekly data gives 52 data points. At the Robyn ratio of 1 to 10, that supports about 5 variables in total. You need a control for each festival, sale and big video, so the budget for channel variables is gone before you add one ad platform.
Meridian suggests ways out when data is short: use more years, combine channels, or drop a channel with low spend. For an indie studio, that leaves a model with one or two variables, which a simple lift test answers more directly.
Sources: Robyn documentation, Analyst's guide to MMM; Google Meridian documentation, Amount of data needed.
What should an indie studio use instead?
Attribution to compare options, and lift tests to size the effect. Both work from the first week a Steam page is live.
- Tag every link with UTMs. Steam counts wishlists within 72 hours of a click on a tagged link to an app or sale page, and finalizes them 4 days after the visit.
- Give each creator a code, so viewers who type it instead of clicking are counted.
- Keep a sheet with spend per channel per week. You need it for cost per wishlist anyway.
- Before each beat, record 7 to 14 days of daily wishlist adds as a baseline.
- After the beat, subtract the baseline from the wishlist adds in the window. That is the lift.
- Compare lift and UTM wishlists per channel every month, and move budget toward the channel with the lower cost per lifted wishlist.
Sources: Steamworks UTM Analytics and Wishlists documentation; Manalens attribution and Steam documentation.
Can I run a geo experiment with Steam data?
Sometimes, if your game gets enough wishlists per country. Meta's GeoLift best practices recommend daily data over weekly, at least 25 pre-treatment periods of 20 or more geo units, and a test of at least 15 days with daily data or 4 to 6 weeks with weekly data.
Steam gives the data GeoLift needs. The Steamworks wishlist report breaks wishlists down by region or country, and the data updates daily for the day before. The hard part is volume. If most countries get one or two wishlists a day, a change from the ads is hard to see. Group small countries into regions, or run the test longer.
Sources: GeoLift documentation, Best practices; Steamworks Wishlists documentation.
When does MMM start to make sense?
When several paid channels run all the time with steady spend, and you have two or more years of weekly data. That is more typical of a publisher with a portfolio, or of a live game with a constant user acquisition budget, than of a premium indie game before launch.
Even then, Robyn recommends calibrating the model with incrementality studies on an ongoing and regular basis. Robyn accepts a calibration result as one point estimate per test, such as the incremental response from a holdout. So lift tests are the base of an MMM, not a replacement for it.
Sources: Robyn documentation, Analyst's guide to MMM; Google Meridian documentation, Collect and organize your data.
What does each method miss?
Each has a blind spot, so use them together and say which number you quote.
- Attribution misses players who saw a video and searched Steam later, and players who were not signed in to Steam when they clicked.
- Lift tests miss slow effects that arrive after the window, and they break when two beats overlap.
- Geo tests need enough wishlists per country and countries that behave alike.
- MMM needs years of weekly data and steady spend, and its output is only as good as its controls.
Sources: Steamworks UTM Analytics documentation; Robyn documentation; GeoLift documentation, Best practices.
Where does Manalens fit?
Manalens is the analytics and growth engine built for games. It covers the attribution side: tagged short links, creator codes, six attribution models, the Steam UTM report and daily wishlist adds per country from a Steamworks Financial API key (the automatic Steam sync is in early access). It does not run MMM, import ad spend or compute lift. Keep spend in a sheet and divide.
Sources: Manalens product documentation.
Frequently asked questions
Is marketing mix modelling worth it for an indie game?
Usually not before launch. Google's Meridian guide asks for two to three years of weekly data, and Meta's Robyn guide asks for 10 observations per variable. A studio with one or two channels and a few beats a year gets more from tagged links and simple lift tests.
What is the difference between MMM and incrementality testing?
MMM estimates the effect of each channel from history with a statistical model. An incrementality test changes one thing, such as ads on in some countries and off in others, and measures the difference. Robyn uses incrementality results to calibrate its models, so the tests come first.
Can I use Meridian or Robyn for free?
Yes. Both are open source. The cost is the data and the time. You need weekly spend or impressions per channel, a KPI such as wishlist adds, control variables for sales and festivals, and two or more years of history. Most of the work is cleaning that data.
Does Manalens do marketing mix modelling?
No. It does not run MMM, import ad spend or compute lift. It records visits, clicks, codes and installs per channel, runs six rule-based attribution models, imports the Steam UTM report and syncs daily wishlist adds per country (the automatic Steam sync is in early access), which you can use in your own lift tests.
Sources and implementation
- Google Meridian documentation: Collect and organize your data (Google, read 18 November 2026) ↗
- Google Meridian documentation: Amount of data needed (Google, read 18 November 2026) ↗
- Robyn documentation: Analyst's guide to MMM (Meta Open Source, read 18 November 2026) ↗
- GeoLift documentation: Best practices (Meta Open Source, read 18 November 2026) ↗
- Steamworks documentation: Wishlists (Valve, read 18 November 2026) ↗
- Steamworks documentation: UTM Analytics (Valve, read 18 November 2026) ↗