Media mix modelling (MMM) is a statistical method that estimates how much each channel, plus outside forces like seasonality and price, contributed to sales. It is worth buying once you spend across several channels, have about two years of weekly data and can act on the answer. Under roughly $10 million in annual revenue, you will usually get more from simple incrementality tests than from a full model.
That is the blunt version. The rest of this guide covers the mechanism, the data, the traps and the order of operations. We run paid media for e-commerce brands, so we see where measurement helps a budget decision and where it just produces a nice deck.
Key takeaways
MMM works on aggregate data (spend and sales by week), so it needs no cookies and no user-level tracking. It also needs history: vendors typically ask for two years of weekly data.
A model that fits your past sales well can still give wrong ROI numbers. A March 2026 simulation showed lower error on prediction can degrade ROI accuracy when priors are not calibrated.
MMM, attribution and experiments answer different questions. IAB Australia’s Ad Effectiveness Council (December 2025) says MMM should sit inside a wider measurement framework, not stand alone.
Incrementality tests, such as geo holdouts, are the cheaper first step and later become the calibration for your model.
Below about $10 million in revenue, heavy incrementality tooling often costs more than it returns, according to operators on the Blue Sense podcast (October 2024). Treat that as a rule of thumb, not a law.
Table of Contents
What media mix modelling actually does
An MMM is a regression. You feed it weekly spend per channel, weekly sales, and the things that move sales without you touching them: seasonality, promotions, price, weather, economic conditions, competitor activity. It then estimates how much of each week’s sales came from each input.
Two pieces of maths matter. First, adstock: an ad seen this week still nudges purchases next week, so the model spreads each channel’s effect forward in time. Second, saturation: the hundredth thousand dollars in a channel buys less than the first. The model fits a curve to that, and the curve is what tells you where to move the next dollar.
The output is not a report card on the past. It is a response curve per channel, a baseline (sales you would get with zero media) and a scenario tool: if we move $50,000 from channel A to B, what happens?
Media mix or marketing mix?
The terms get swapped constantly. Strictly, media mix models focus on advertising channels, while marketing mix models also cover pricing, distribution and promotion. For a brand deciding where to put ad dollars, the practical difference is small. For a brand that discounts heavily, it is huge, because promo depth can explain more of a sales spike than any ad did.
What you need before you start
Most failed models fail at the data stage, not the maths stage.
History. Vendors typically want two years of weekly data. Daily data can shorten that. Fewer than two seasonal cycles means the model cannot separate Christmas from your Christmas campaign.
Variation. If you spent the same $20,000 every week for two years, the model has nothing to learn from. It needs weeks where spend went up, down or off. Brands that scale steadily in one channel are the hardest to model.
More than spend. Expect to supply external factors too. Without economic indicators, seasonality and competitor activity, baseline demand swings get credited to your marketing.
Offline and owned channels. TV, streaming, radio and out-of-home never show up in a click-based tool. MMM is one of the few ways to put them next to paid social on one page.
A decision to make. If no one can move budget based on the result, skip it.
A useful sanity check: can you list the last six times your weekly spend in any channel changed by more than 30%? If not, your data may not support a channel-level model yet.
MMM vs attribution
Both tell you which channels “work”. They do it in opposite ways and fail in opposite ways.
| Multi-touch attribution (MTA) | Media mix modelling | Incrementality tests | |
|---|---|---|---|
| Data | User-level click and view paths | Aggregate weekly spend and sales | Treatment vs control groups |
| Strength | Fast, granular, creative-level | Covers every channel, offline included, no cookies | Direct causal evidence |
| Weakness | Rewards the last touch; blind to people it cannot track | Slow, needs history, can be confidently wrong | One channel at a time, costs some sales during the test |
| Cadence | Daily | Monthly or quarterly refresh | As needed per channel |
Attribution tells you who clicked before buying. It does not tell you who would have bought anyway. Brand search is the classic case: it harvests demand that other channels created, then takes the credit. We cover the column-level version of this problem in our Google Ads attribution guide.
MMM sees the whole picture but only at the resolution of a week. It cannot tell you which creative won. Experiments sit in the middle: they answer one causal question cleanly.
The honest answer to “MMM or attribution” is neither alone. Platform reports for daily steering, experiments for truth on individual channels, MMM for the budget split. IAB Australia’s council says the same: models are not “set and forget” and need continuous refinement as consumer behaviour and media shift.
Incrementality testing: the cheaper first step
An incrementality test splits your audience or your geography into a group that sees ads and a group that doesn’t, then compares outcomes. Measured’s primer gives the standard formula: test conversion rate minus control conversion rate, divided by test conversion rate, gives incrementality as a percentage.
Worked example with made-up numbers, labelled as such: your test group converts at 2.0% and your holdout at 1.5%. Incrementality is (2.0 − 1.5) / 2.0 = 25%. Three in four conversions the platform claimed would have happened anyway. That does not make the channel bad. It means its reported ROAS overstates the real return by a factor of four, and you should budget against the real number.
Three ways to run one
Conversion lift test. The platform randomises users into exposed and held-out groups. Easy to run inside the ad manager, but the platform grades its own homework, and it only covers that platform.
Geo holdout. You switch ads off (or cut them) in some regions and compare against matched regions. Because it works on regions, not individuals, it does not depend on tracking. In Australia, state or metro splits are the obvious design, with the catch that Sydney and Melbourne behave differently from Perth or Adelaide, so matching matters.
Spend-pulse test. Cut or raise spend for a few weeks in one channel and watch total sales. Crude, but it costs nothing to set up. Pair it with a MER view. Our MER vs ROAS workflow shows how to read the blended number without fooling yourself.
Why geo tests suit smaller brands: they need only sales by region, which most Shopify stores already have. What they cost is some lost sales in the holdout regions, and a few weeks of patience. We walk through the Meta-specific version in our two-perspective piece on incrementality for Meta ads.
Caveat. Test length and size depend on your volume. A store doing 40 orders a day cannot detect a 5% lift in a two-week split. Run a power calculation before you pause anything, or you pay the cost of the test and learn nothing.
The trap: a model that fits is not a model that is right
Most buyers judge a model on how closely it traces past sales. That sounds reasonable. It is the wrong test.
MMM is a causal question, not a forecasting one. In a simulated Bayesian MMM built in March 2026, pushing prediction error (MAPE) down made ROI estimates worse when priors were not calibrated. The mechanism: spend is not random. You raise budget when demand is rising, so the model sees spend and sales climbing together and gives the channel credit that belongs to demand. A flexible model happily fits that pattern perfectly and gets the cause backwards.
A benchmark dataset paper posted to arXiv in August 2026 makes the same point from the other side: real-world MMMs are rarely validated against ground truth, because in real data the truth is unobservable.
So what do you do? Calibrate. Practitioners writing in 2026 recommend feeding the results of geo experiments into the model as Bayesian priors, instead of trusting the fit alone. Your experiment says paid social is 25% incremental. The model has to respect that. That loop, test then calibrate then test again, is the whole discipline.
If a vendor leads with “our model has 95% accuracy”, ask: accuracy against what? Against past sales, or against a holdout test you ran? Those are different claims.
Who sells MMM in Australia
The market is crowded. IAB Australia’s December 2025 landscape report profiles twelve vendors, with additions in the latest update. Methods range from classical econometrics to Bayesian and machine learning approaches. Two broad types of option exist:
Open-source frameworks from large platforms. Free software, but you or an analyst supply the data pipeline, the priors and the interpretation. Sensible for teams with a data scientist.
Commercial vendors and measurement consultancies, from global research firms to local specialists. You pay for data engineering, modelling and a refresh cycle. Sensible when no one in-house owns statistics.
Before you sign with either, ask five things: How many weeks of data do you need? How do you handle seasonality and promotions? Do you calibrate with experiments? How often does the model refresh? Can I see the response curves, not just a summary slide?
Do you need it yet?
Operators on the Blue Sense podcast (October 2024) argue most businesses under $10 million in revenue do not need incrementality tooling. We agree with the direction and would add a number of our own: no hard cut-off exists, so treat $10 million as a prompt to think, not a rule.
A rough ladder:
Under about $2 million, one or two channels. Skip MMM. Watch blended MER, run a spend-pulse test each quarter.
Growing, three or more paid channels, some offline. Run geo holdouts on your biggest channel. Start storing clean weekly data now, so you have two years when you are ready.
Multi-channel with TV, streaming or out-of-home. Build a model, and use your experiment history to calibrate it.
Why this order? An early MMM built on thin, flat data gives a confident answer you cannot check. A single holdout test gives a narrow answer you can.
One more thing to fix before any of this: the incentives. The same podcast makes the case for aligning your agency around contribution margin instead of platform-reported metrics or a percentage of ad spend, since a fee that rises with spend rewards spending, not profit. Our view is that this applies to every agency, us included, so ask the question of whoever you hire.
What you can do this month
Export 24 months of weekly sales, spend by channel, promo calendar and price changes into one sheet.
Mark every week where spend moved by more than 30% in any channel. Count them.
Pick your largest channel and design a geo holdout with matched regions. Run the power calculation first.
Compare the lift result with the platform’s reported ROAS. Write the gap down. That ratio is your first calibration factor.
Decide which budget line the result would change. If none, stop and save the money.
Our perspective
Most brands we meet are not short of dashboards. They are short of a number they trust enough to move budget on. MMM is one route to that number. It is not the first route.
We run strategy, creative, paid media and conversion work as one team, so we see the loop that measurement vendors miss. A model might say paid social is under-invested. Then someone has to make creative that holds up at higher spend, fix a landing page that leaks, and watch the next test. If those jobs sit with four different suppliers, the insight dies in a hand-off.
Our approach to measurement is plain. Use platform data to steer daily. Use blended MER to keep ourselves honest. Use holdout tests to settle arguments about single channels. Reserve a full model for brands whose channel mix and data history justify it. We would rather tell a client they do not need a model yet than sell one that cannot be checked.
And a model that cannot be checked against an experiment is a very expensive opinion.
Getting the measurement and the media under one roof
If you want results measured against revenue and not platform vanity numbers, we handle Meta and TikTok ads, integrated campaigns across TV, streaming, radio and out-of-home and conversion optimisation as one team. See the full range on our services page, or look at how we worked with Bonaire.
FAQ
What is the difference between MMM and multi-touch attribution?
MMM uses aggregate weekly spend and sales to estimate each channel’s contribution, offline channels included. Multi-touch attribution follows individual user paths and depends on tracking. MMM is slower and coarser, but it survives cookie loss and covers channels attribution cannot see.
How much data does a media mix model need?
Vendors typically ask for two years of weekly data, though daily data can shorten that. You also need variation in spend and outside factors such as seasonality and promotions. Flat, unchanging budgets give the model little to learn from.
What is a geo holdout test?
You stop or reduce ads in a set of regions and compare sales with similar regions where ads keep running. The gap is the incremental effect. It works without user-level tracking, but needs enough volume and well-matched regions to detect a real difference.
Is incrementality testing better than MMM?
Neither is better. Incrementality tests give clean causal answers for one channel at a time. MMM gives a whole-mix view and forecasts. The strongest setup uses test results to calibrate the model, which is the approach IAB Australia’s council points toward with its call for MMM inside a wider measurement framework.
Sources
IAB Australia, Market Mix Modelling Landscape (December 2025, PDF)
PPC Land, IAB Australia unveils market mix modelling vendor landscape
