Media mix modelling is a privacy-safe, top-down econometric method that estimates each channel’s incremental contribution to ecommerce sales. We recommend it when you need cross-channel budget decisions at scale rather than tactical, click-level optimisation. It works best for brands running multiple paid channels alongside promotions and seasonal pushes, where attribution tools alone cannot show the full picture.
TL;DR:
- Brands with three or more paid channels, regular promotions, and quarterly budget decisions have a stronger case; single channel advertisers can use platform reporting.
- Prepare at least two years of weekly sales, spend, promotion, price, and assortment data; missing controls can misattribute unrelated sales changes to media.
- Constrain adstock decay by funnel role; freely estimated decay on noisy weekly data can produce implausible results that fail out of sample.
- Test uncertain channels with holdouts or geographic experiments before trusting observational estimates, then pilot small budget shifts for a few months before broad reallocation.
- Across 2,226 Meta ad experiments, PIE reached out of sample R² 0.88 for incremental conversions per dollar, versus 0.19 for last click attribution.
Table of Contents
- What is media mix modelling and how does it differ from attribution?
- Why ecommerce and DTC brands are turning to MMM
- The building blocks: adstock, saturation, baseline and controls
- Data checklist and preparation before you build an MMM
- Model selection, diagnostics and identification risks
- Practical toolset: open-source modelling, experiments and PIE
- From elasticities to action: turning model output into budget decisions
- How a specialist ecommerce PPC agency applies measurement to paid media
- Who should invest in MMM now, and how to sequence the work
- Put your media mix model to work with Oxedent
- FAQ
- Sources
What is media mix modelling and how does it differ from attribution?
Media mix modelling, or MMM, is a statistical technique that uses aggregate time-series data, typically weekly sales, spend and external factors, to estimate how much each marketing channel contributes to revenue. It has roots in econometric demand modelling from decades before digital advertising existed, and it has resurfaced as a favoured measurement method since cookie restrictions made user-level tracking less reliable.
The core distinction between MMM and data-driven attribution is the direction of the analysis. Attribution works bottom-up: it tracks individual user journeys, click by click, and assigns credit to touchpoints along the path to purchase. MMM works top-down: it looks at aggregate outcomes over time and uses statistical relationships to infer how much of the variation in sales can be explained by variation in spend, promotions, pricing and seasonality.
This difference in direction means the two approaches answer different questions well:
- Attribution suits tactical, campaign-level optimisation: which ad set, keyword or creative is driving clicks and conversions right now.
- MMM suits strategic, cross-channel decisions: how to split next quarter’s budget across Google Ads, Meta, retail media and offline channels.
- Attribution struggles with upper-funnel and offline channels that do not generate trackable clicks.
- MMM captures halo effects, where one channel’s spend lifts demand that another channel appears to capture.
Neither method replaces the other outright. Attribution remains useful for day-to-day campaign management, while MMM gives you the aggregate view needed when you are deciding how much to spend on each channel over a quarter or a year. Many ecommerce teams end up running both, using attribution for in-platform bid and budget tweaks and MMM for the bigger allocation calls that attribution cannot answer reliably.
Why ecommerce and DTC brands are turning to MMM
Ecommerce and DTC brands face a measurement problem that attribution alone cannot solve: cookie deprecation and platform-side tracking restrictions have made user-level attribution progressively less complete. MMM does not depend on cookies, device IDs or pixel fires. It works from aggregate sales and spend data, which makes it resilient to the privacy changes that have eroded attribution’s accuracy.
A second reason is scope. Ecommerce brands increasingly spend across retail media, paid social, paid search, affiliate, and sometimes offline channels like direct mail or out-of-home. Attribution tools typically see only the digital, trackable slice of that mix. MMM can include every channel in one statistical frame, as long as spend and timing data exist for it, which makes it the only practical way to compare a retail media placement against a Meta campaign on the same footing.
The third driver is planning. Promotional calendars, seasonal spikes and price changes all interact with media spend in ways that single-channel dashboards obscure. MMM explicitly models promotions, seasonality and baseline demand alongside channel spend, which means you can see how much of a Black Friday spike came from media versus price cuts versus simple seasonal demand.
A retailer case study using joint elasticity estimation and budget simulation found that channel elasticities and simulated budget allocations tended to favour a majority share of spend on advertising, with a smaller but material share retained for promotions, when funnel-differentiated adstock constraints were applied. That kind of output, a defensible split between advertising and promotional spend, is exactly the decision MMM is built to support.
Use MMM when you are facing one or more of these situations:
- You run paid media across three or more channels and need to compare their incremental value.
- You rely on promotions or seasonal campaigns and want to isolate media’s contribution from price and timing effects.
- You are planning budgets a quarter or more ahead, rather than optimising bids week to week.
The building blocks: adstock, saturation, baseline and controls
Every media mix model rests on the same underlying structure, even when the statistical method used to fit it varies. Understanding the three transforms that turn raw spend into modelled sales contribution makes the rest of the modelling process far easier to interpret.
Adstock captures the idea that advertising effects persist and decay over time rather than vanishing the moment spend stops. A television ad or a video view does not convert a customer instantly; its influence fades over subsequent weeks. Adstock applies a decay parameter to each channel’s spend so that this week’s effect includes a weighted share of prior weeks’ spend. Decay rates differ sharply by funnel role. One applied study suggests funnel-differentiated decay ranges around 0.6 to 0.8 for video, 0.4 to 0.6 for display, and 0.3 to 0.5 for search, reflecting how upper-funnel formats linger in the mind longer than a search ad clicked moments before purchase.
Saturation, often modelled with a Hill curve or similar diminishing-returns function, captures the fact that each additional pound of spend on a channel generates less incremental return than the pound before it. A channel that is far from saturated can absorb more budget efficiently; one that has flattened out needs either a cap or a different creative and targeting approach to keep working. Saturation curves are what let an MMM tell you not just whether a channel works, but how much further you can push it before returns fall away.
Baseline, trend, seasonality and controls account for everything that is not media. Baseline demand is what you would sell with no advertising at all. Trend captures gradual growth or decline. Seasonality captures recurring calendar effects such as a pre-Christmas lift. Controls include price changes, assortment shifts, competitor activity, weather, and anything else that moves sales independently of your media spend. Without properly specified controls, the model risks crediting media with movements that price or season actually caused.
These three elements compose into the standard MMM structure: raw spend for each channel first passes through adstock to capture carryover, then through saturation to capture diminishing returns, and the resulting transformed variable is weighted by a coefficient and added to baseline, trend, seasonality and controls to produce modelled sales.
- Raw weekly spend per channel is collected and cleaned.
- Each channel’s spend series passes through an adstock transform reflecting its funnel role.
- The adstocked series passes through a saturation curve to capture diminishing returns.
- The transformed series is combined with baseline, trend, seasonality and controls to estimate total sales.
- Coefficients from this fitted relationship become the channel elasticities used for budget decisions.
Pro Tip: Constrain adstock decay by funnel role rather than letting the model estimate it freely; unconstrained decay parameters on noisy weekly data often produce implausible results that collapse under out-of-sample testing.
Data checklist and preparation before you build an MMM
The quality of an MMM depends far more on the inputs than on the sophistication of the statistical method. Before any modelling begins, you need a clean, consistent dataset covering enough history to separate media effects from everything else moving sales.
At minimum, gather the following at weekly granularity:
- Sales or revenue, ideally net of returns, broken out by the geography or segment you intend to model.
- Ad spend by channel and format, separated enough to distinguish, for example, Meta prospecting from retargeting, or Google Shopping from Performance Max.
- Promotions and discounting, including start and end dates, discount depth and which products were included.
- Price and assortment changes, since a price cut or a stockout can easily be mistaken for a media effect if left unrecorded.
- External controls, such as competitor promotions where known, or macro factors like seasonality indices.
A useful way to see how these inputs map to their modelling role is to lay them out against the transform or control they feed:
| Data input | Granularity | Role in the model |
|---|---|---|
| Sales/revenue | Weekly | Dependent variable |
| Channel ad spend | Weekly, by channel/format | Input to adstock and saturation |
| Promotions | Weekly, with start/end dates | Control variable |
| Price changes | Weekly | Control variable |
| Seasonality index | Weekly | Baseline/trend component |
On timeframe, most practitioners recommend at least two full years of weekly history, which gives enough cycles through major seasonal events to separate recurring calendar effects from one-off media pushes. Shorter histories make it difficult to distinguish a genuine channel effect from a seasonal coincidence.
Preprocessing work typically includes applying the adstock transform to each channel’s raw spend, encoding promotions as dummy or intensity variables rather than leaving them implicit in the sales series, and checking for missing weeks or inconsistent channel definitions across the history. Collinearity between channels that tend to move together, such as search and shopping spend rising and falling in step, needs explicit attention, since it is one of the most common causes of unstable or counter-intuitive coefficient estimates.
Model selection, diagnostics and identification risks
Once the data is prepared, the choice of model family shapes what the output can tell you. Linear models are the simplest and most interpretable, treating each channel’s contribution as additive once adstock and saturation have been applied. Log-log specifications, where both sales and spend are log-transformed, are common when you expect constant elasticities, meaning a given percentage change in spend produces a consistent percentage change in sales. Time-varying models allow channel coefficients to shift over the modelled period, which can capture genuine changes in channel effectiveness but also introduces more parameters to estimate from the same amount of data.
Whichever family you choose, a proper diagnostic routine is non-negotiable before you trust the output for budget decisions:
- AIC/BIC to compare model fit while penalising unnecessary complexity.
- Out-of-sample RMSE to check whether the model predicts weeks it has not seen, not just the weeks used to fit it.
- VIF (variance inflation factor) to flag collinearity between channel spend series that could be inflating or destabilising coefficients.
- Durbin-Watson and residual ACF to check for autocorrelation in the residuals, which signals the model has not fully captured the time-series structure.
These checks matter because MMMs face a genuine identification problem that no amount of data volume alone fixes. Recent research on nonlinear and time-varying effects in marketing mix models shows that conflation between a channel’s saturation curve shape and genuine changes in its effectiveness over time is common, particularly when spend series are autocorrelated and when adstock variables are involved. In plain terms: a model can mistake “this channel is becoming saturated” for “this channel is becoming less effective over time”, and from observational spend data alone, it is often impossible to tell the two apart.
The practical remedy is exogenous variation. Recovering a channel’s true response curve from spend data that moves smoothly and predictably, as most always-on budgets do, is what the same research describes as an ill-posed problem. Deliberate holdouts, step-changes in spend, or geo-level experiments introduce the kind of variation that lets the model separate saturation from time-varying effects with far more confidence than observational data alone permits. This is one reason experiments remain central to credible MMM programmes rather than a nice-to-have add-on.
Practical toolset: open-source modelling, experiments and PIE
Choosing how to implement MMM is as important as understanding the theory. Three broad approaches dominate current practice, and most credible programmes end up combining them rather than picking just one.
Meridian is an open-source media mix modelling framework that supports geo-level modelling and reach and frequency inputs, giving teams full transparency into the modelling process and the ability to customise it for their own data structure. Its openness suits teams that want auditability over a black-box vendor tool, particularly when geo-level data is available to strengthen identification.
Running randomised controlled trials gives the most credible incrementality estimates available, since a true RCT isolates a channel’s causal effect from everything else moving sales. RCTs are resource-intensive and slow to run at scale across every channel and market, which is where their practical limit shows up: no ecommerce brand can run a continuous RCT on every campaign, every week.
Predicted Incrementality by Experimentation (PIE) addresses that limit directly. Rather than relying solely on a handful of expensive experiments, PIE learns a mapping from campaign features to causal effects using a sample of RCTs, then applies that mapping to predict incrementality at scale across campaigns that were never directly tested. In the underlying study using 2,226 Meta ad experiments, PIE achieved out-of-sample accuracy of R²=0.88 for incremental conversions per dollar, compared with R²=0.19 for seven-day last-click attribution on the same data.
The strongest programmes stack these approaches: build an MMM for the strategic, cross-channel view, run targeted experiments where identification is weakest, and use PIE-style scaling to extend experimental learning across the campaigns you cannot test directly.
- Meridian suits teams wanting geo-level transparency and control over the modelling pipeline.
- RCTs suit high-stakes channel questions where the budget at risk justifies the cost of testing.
- PIE-style methods suit scaling a limited experimental programme across many untested campaigns.
From elasticities to action: turning model output into budget decisions
A finished MMM is only useful once its coefficients become decisions. The elasticities it produces, the percentage change in sales for a percentage change in spend on each channel, need translating into a working budget plan.
- Run budget simulations by feeding different spend allocations through the fitted saturation curves to see predicted revenue at each scenario, rather than relying on a single point estimate.
- Compare scenarios against current spend to identify channels still on the steep part of their saturation curve, which can typically absorb more budget efficiently, against channels flattening out, which cannot.
- Stagger upper and mid-funnel activity ahead of major promotions so that awareness and consideration spend builds demand before a sale period begins, rather than competing for attribution credit with lower-funnel spend during the promotion itself.
- Convert elasticities into incremental ROAS (iROAS) by dividing the modelled incremental revenue from a channel by its spend, giving a figure that reflects true causal contribution rather than last-click credit.
- Translate iROAS into contribution margin targets so budget decisions reflect profitability, not just revenue, particularly where product margins vary significantly by category.
Calendar-aware planning matters as much as the maths. A model that treats every week identically will misallocate budget around Black Friday, January sales, or a product launch, because it cannot anticipate that upper-funnel spend needs to land weeks before a promotion to build the demand that the promotion then converts. Rolling forecasts and weekly review cycles, rather than a static annual plan, let you adjust allocations as actual performance diverges from the model’s predictions.
Pro Tip: Treat every budget simulation output as a range, not a single number: report the uncertainty around each scenario so stakeholders understand the confidence behind a reallocation recommendation, not just its headline figure.
For brands converting these outputs into blended performance targets across paid and organic channels, blended ROAS offers a useful complementary framework for interpreting what an MMM’s channel-level contribution means at the whole-business level.
How a specialist ecommerce PPC agency applies measurement to paid media
We built our agency around a focus on eCommerce PPC, managed with profitability as the measure of success rather than clicks or impressions. That focus shapes how we treat measurement work like MMM. We are not trying to prove that every channel deserves more budget; we are trying to find the allocation that grows revenue without eroding margin.
In practice, that means we lean on a few consistent operational habits:
- Feed optimisation, ensuring product data feeding Google Shopping and Performance Max campaigns is structured to support the kind of granular channel comparison that MMM and attribution both depend on.
- Fluid campaign structures, built to be reorganised quickly when a budget simulation or incrementality read suggests a channel’s role should change.
- Waste reduction, treating spend that is not converting at an acceptable contribution margin as a cost to cut, not a metric to defend.
Our engagement path typically starts with a free Google and Facebook Ads audit, which gives us and the client a shared, evidence-based view of where current spend is working and where it is not. From there, clients who fit our focus move into managed optimisation across Google Ads, Facebook Ads, Google Shopping and Performance Max. We qualify prospects deliberately: we focus on brands ready to scale profitably rather than early-stage or purely price-driven accounts, because the data-led work we do needs a budget and a sales history substantial enough to act on.
For brands weighing attribution tools against a fuller measurement stack, our overview of ecommerce attribution tools covers the trade-offs worth understanding before committing budget to any single method.
Who should invest in MMM now, and how to sequence the work
MMM earns its cost fastest for brands already spending meaningfully across three or more paid channels, running regular promotions, and facing genuine quarterly or annual budget decisions rather than daily bid tweaks. If you are spending on one channel, attribution and in-platform reporting will tell you most of what you need.
For everyone else, I would sequence the work rather than attempt it all at once. Start with quick wins: a data audit, a cleaned spend and sales history, and one or two small, targeted experiments on the channels where incrementality is least certain. Use those experiments to sense-check whatever attribution is currently telling you. Only then build the full model, because a model built on unreliable inputs will produce confident-sounding numbers that are wrong.
Once the model runs, treat its first output as a hypothesis to test with a small reallocation, not a mandate for a sweeping budget overhaul. Expect a pilot cycle of a few months before you trust the model enough to drive a full quarter’s planning.
— Biplab
Put your media mix model to work with Oxedent
Building a model is only half the job. The harder part is turning elasticities and saturation curves into campaigns that actually hit a better contribution margin, week after week, which is where our day-to-day management work earns its keep.
We run eCommerce PPC exclusively: Google Ads management, Facebook Ads and Google Shopping management, and landing page optimisation, all structured around the profitability targets your measurement work points to rather than vanity metrics. If your MMM output suggests a channel is under-saturated or a promotion window needs upper-funnel spend staggered ahead of it, our fluid campaign structures are built to act on that quickly, without locking you into a long-term contract.
Most engagements begin with a free Google and Facebook Ads audit, which gives us a concrete read on where your current spend sits against the allocation your model recommends. From there, our PPC management service starts from £350 per month for ongoing, data-led optimisation across your paid channels. If you have model outputs ready to act on, or want a second opinion on what they mean for your budget, request your audit and we will take it from there.
FAQ
What is a media mix model?
A media mix model is a statistical technique that uses aggregate sales, spend and external data, such as promotions and seasonality, to estimate how much each marketing channel contributes to revenue. It works top-down from historical time-series data rather than tracking individual customer journeys, which makes it useful for strategic budget decisions across channels.
What are four types of ecommerce business models?
The main ecommerce business models are business-to-consumer (B2C), business-to-business (B2B), consumer-to-consumer (C2C) and direct-to-consumer (DTC), each describing who sells to whom rather than how media is measured. DTC brands, which sell directly to shoppers without a retail intermediary, are the group most likely to rely on media mix modelling because they typically manage their own multi-channel paid media budgets.
What is the 3-3-3 rule in marketing?
Definitions of the 3-3-3 rule vary across marketing sources, and no single authoritative version applies universally across industries. In general usage it refers to structuring content or campaigns around three audiences, three channels or three messages, but this falls outside the scope of media mix modelling and should not be treated as a fixed measurement standard.
What are the main types of digital marketing channels?
Digital marketing commonly spans paid search, paid social, display advertising, retail media, affiliate marketing, email marketing, search engine optimisation and content marketing. A media mix model can incorporate any of these channels as long as consistent spend and timing data exists for each one, which is what allows it to compare their incremental contribution on equal footing.
How much does ecommerce PPC management cost?
Our PPC management service starts from £350 per month, covering ongoing, data-led optimisation across Google Ads, Facebook Ads, Google Shopping and Performance Max. Pricing for Google Ads, Facebook Ads and Google Shopping management beyond the base tier depends on account scope and is available on request.
Sources
- Predicted Incrementality by Experimentation for Ad Measurement | Management Science
- Your MMM is broken: Identification of nonlinear and time-varying effects in marketing mix
- Journal case study on joint elasticity estimation and budget simulation