Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

Marketing Analytics › Attribution

Marketing Mix Modelling Basics

What marketing mix modelling is, the aggregated inputs it uses, adstock and saturation, its outputs, how it complements attribution, and its limits.
Guideadvanced11 min readUpdated

What MMM answers

The question behind a mix model is the budget planner's question: if the total budget and its split across channels were different, what would the outcome have been? To answer it the model takes a time series of the outcome, decomposes it into a baseline that would have happened with no media plus a contribution from each channel plus the effects of price, promotions, season and other controls, and estimates how each channel's contribution responds to its spend. From those response curves follow the questions that matter for planning: which channel is closest to saturation, where the next unit of budget would earn most, and what the outcome would be under a different allocation.

It belongs to the same family of problems as Attribution but attacks it from the opposite end. Attribution starts from individual conversions and asks which recorded touches preceded them; the mix model starts from the total and asks which aggregate inputs explain its movement. That difference in grain gives the model its distinctive strengths. It can estimate media that leave no digital trace. It can include outcomes that never touch a website, such as store sales or call-centre orders. It can account for things that are not marketing at all, such as weather, pricing and competitor activity, which attribution has no way to hold constant. And it needs no identifiers, cookies or consent, because it only ever sees totals.

The price of working at that grain is that the model cannot say anything below it. It estimates a channel's contribution, not a campaign's, a creative's or a keyword's, and it estimates it from history, so a channel with no history or no variation in spend is invisible to it. The attribution hub positions mix modelling as the channel-level layer of a measurement stack, and that is the right way to think about it.

For the campaign-level question a model cannot answer, whether one radio burst or podcast series did anything at all, the guide to measuring TV, radio, podcasts and print covers response mechanisms, search lift and regional tests.

Inputs: aggregated spend, outcomes, controls

A mix model is built from three kinds of time series, each at one consistent grain, usually weekly, and usually covering a run long enough to contain more than one full seasonal cycle.

Media inputs. Spend, or a delivery measure such as impressions or gross rating points, per channel per period. The channel definitions are the modeller's first consequential decision: too coarse and the model cannot distinguish prospecting from retargeting; too fine and the series become sparse and the channels collinear. Each media series should also carry any structural change, such as a new agency, a new creative platform or a change in how spend is recorded, because a step in the series that is not explained will be attributed to something. Where television enters the model as rating points, record the audience the points were measured against and keep it the same for the whole series.

Outcomes. The variable the business is trying to move, in total: revenue, orders, sign-ups or store sales, from the system of record rather than from any single platform's reporting. Some models are built with several outcomes, for example an upper-funnel measure such as branded search volume and a sales measure, so that media that work through the first can be seen moving the second.

Controls. Everything else that moves the outcome and is not media: seasonality and holidays, price and promotions, product launches, distribution changes, competitor activity where it can be measured, weather for weather-sensitive categories, and macro factors. Controls are where most of the modelling judgement lives. Leave one out and its effect is absorbed by whichever media series happens to move with it; include a control that is itself caused by media, such as branded search volume used as a control rather than an outcome, and the model will give that control the credit the media deserved.

Everything enters as an aggregate. There is no user-level table anywhere in the pipeline, which is why the method has drawn renewed attention as individual tracking has become less complete.

Assembling the three kinds of series into one weekly table, with the gap checks and data log a reviewer will ask for, is the subject of the tutorial on preparing data for a marketing mix model.

Adstock and saturation

Two transformations turn raw spend into something a model can relate to sales, and understanding them is most of understanding the method.

Adstock, or carry-over. Advertising does not act only in the week it runs. Some of its effect arrives later, as people who saw it buy when they next need the product, and some of it accumulates, as repeated exposure builds familiarity. Adstock models this by giving each period's spend a lingering effect that decays over subsequent periods. The simplest form adds a fraction of the previous period's adstocked value to the current period's spend, so the effect tails off geometrically; other forms allow the effect to peak after a delay. The decay rate is estimated per channel and is itself informative: a channel with a long carry-over is doing something that a short-window attribution model will systematically under-credit.

Saturation, or diminishing returns. The first unit of spend in a channel reaches the people most likely to respond; each further unit reaches people less likely to, until additional spend produces almost no additional outcome. Saturation models this with a curve that rises steeply at first and flattens, typically an S-shaped or concave function of adstocked spend. The shape and the point where it flattens are estimated per channel, and they are what make the model useful for planning: the marginal return at the current spend level, which is the slope of the curve at that point, is what decides whether the next unit of budget should go to this channel or another.

Together the two transformations mean that the model's view of a channel is not "spend in, sales out" but "spend, carried over time, through a curve that flattens, out". The estimated contribution of a channel is the area under that transformed series, and its average return and its marginal return can be very different numbers. A channel can show a healthy average return on ad spend (ROAS) over the history and a poor marginal return at its current level, which is exactly the case where attribution, which only ever reports averages, gives the wrong budget signal.

Outputs and what they mean

A finished model produces a small set of artefacts, and each needs to be read with its uncertainty attached.

OutputWhat it isHow to read it
DecompositionThe outcome split into baseline, each channel's contribution and each control's effect, per periodThe baseline is what would have happened with no media; a large baseline is normal and not a failure of marketing
Channel contributionThe total outcome attributed to each channel over the historyA contribution for unobserved media is the model's distinctive value; compare it with what attribution could see
Response curvesOutcome as a function of spend per channel, after adstock and saturationThe slope at current spend is the marginal return; the flattening point is the saturation level
Average and marginal returnContribution over spend, and the slope of the curve at the current levelPlan on marginal; report on average; never confuse the two
Carry-over estimatesThe decay rate per channelLong carry-over argues for a longer attribution window and for patience in tests
Optimised allocationThe split that maximises the outcome under budget and channel constraintsA suggestion to test, not an instruction; it extrapolates the curves beyond observed spend
UncertaintyIntervals on every estimate aboveWide intervals on a channel mean the data could not pin it down; do not move budget on its midpoint

The decomposition is the first thing to inspect and the one most often misread. A baseline that accounts for most of the outcome is the expected result for an established brand, since existing customers, distribution and habit drive most sales whether or not media runs this week. The model's job is the movement around that baseline. The optimised allocation is the last thing to act on, because it is produced by pushing each curve into regions where spend has rarely been and the curve is least certain; it should be read as a direction to test with an experiment, and the budget allocation guide describes how such a test result feeds the marginal-return method.

Illustrative example — not a real result

A furniture retailer reads its first model

A furniture retailer with stores and a website builds a weekly model on 156 weeks of history across television, radio, paid search, paid social, display and catalogue mailings, with controls for price promotions, bank holidays and a store-opening programme. The decomposition puts 71 per cent of revenue in the baseline, 9 per cent on promotions and 20 per cent across media. Television carries the largest media contribution at 7 per cent of revenue, with an estimated carry-over of about five weeks, a result the retailer's attribution reports had no way to show because television produces no clicks. Paid search shows an average ROAS of 6.8 but a marginal ROAS of 2.1 at current spend, with the curve almost flat; paid social shows an average of 3.4 and a marginal of 3.9, still rising. The optimiser suggests moving around £180,000 a year from search to social and television. The team does not do that outright. It moves a third of the amount, runs a geo holdout on the extra social spend, and uses the measured lift to recalibrate the social curve before the next planning round.

MMM, attribution and incrementality together

The three methods answer three questions at three grains, and a measurement stack uses each where it is strongest.

Row labelAttributionIncrementality testingMarketing mix modelling
QuestionWhich recorded touches preceded conversions?What did this channel cause, at this spend, in this period?How does each channel contribute to the total, and how would a different split perform?
GrainIndividual paths, down to keyword and creativeOne channel or one change, one periodChannel by week, across years
Sees unobserved mediaNoYes, if the outcome measured is totalYes
CausalNoYes, by designPartly: a model of history, strengthened by calibration
LatencyImmediateWeeks to months per testMonths to build; refreshed periodically
NeedsTagging, identity, consentRandomisable units, patience, a power calculationLong consistent history with spend variation
Best useTactical optimisation and channel rolesBudget decisions on large or contested linesAnnual and quarterly channel-level planning

The methods also correct one another. A test on a single channel gives a point on that channel's response curve that the model should pass through, and modern practice calibrates the model to such points rather than letting it estimate every curve from history alone; the incrementality testing guide covers how to produce them. The model, in turn, says which channels are worth testing: those with the widest uncertainty, the largest spend or the most surprising marginal return. Attribution fills the gap below the model's grain, deciding which campaigns and keywords inside a channel deserve the budget the model assigned to it. The choosing an attribution approach guide turns this division of labour into a decision framework, and the channel mix guide is where the model's channel-level outputs become a plan.

Data requirements and limits

The method's limits follow from what it is: a statistical model fitted to a history of aggregates.

  • It needs variation. A channel whose spend was flat has no relationship for the model to find; it will be given a contribution that reflects the prior, the modeller's assumption, rather than the data. Planned variation, including deliberate pauses and bursts, is the cheapest way to make a channel estimable.
  • It cannot separate channels that move together. When search and social spend rise and fall in step, the model cannot tell which one moved sales, and it will split the credit on grounds that are not evidence. This collinearity is the most common reason for implausible results.
  • It depends on the controls. An omitted driver of sales lends its effect to whatever media series it correlates with. Modellers spend more time on the control set than on the media set for this reason.
  • It is only as current as its history. A channel that changed character recently, a new creative platform, a new audience, is represented by its old self.
  • It is sensitive to specification. Choices of channel grouping, adstock form, saturation form and priors change the answer. Two competent modellers given the same data will not produce identical curves, and the honest way to present results is with intervals and with the specification stated.
  • It extrapolates poorly. Response curves are well estimated only within the range of spend observed. An optimiser's recommendation to triple a channel is a guess about an unobserved region.

None of these makes the method unreliable in the hands of someone who knows them. They make it a model, to be calibrated against experiments, refreshed on a schedule, and used for the channel-level questions it was built to answer. Open-source modelling code published by large advertising companies has lowered the technical barrier considerably; the data, the controls and the judgement remain the work.

When the choice is between building a model and running an experiment, the comparison of marketing mix modelling and incrementality testing sets out which questions each answers and how tests are used to calibrate the model.

Common mistakes

  • Reading the baseline as a failure. Most sales of an established brand are not driven by this week's media.
  • Planning on average returns. The marginal return is the planning number.
  • Omitting price and promotion controls. Their effect becomes someone's media contribution.
  • Letting channels move in lockstep. The model cannot separate what the plan never separated.
  • Acting on the optimiser without a test. It extrapolates beyond observed spend.
  • Treating the model as a one-off. Channels change; refresh and recalibrate on a schedule.
  • Building one when a geo test would do. The simpler instrument answers many single-channel questions.

The incrementality testing guide covers the experiments that calibrate a model, and the attribution hub places mix modelling in the wider measurement stack beside attribution and testing.

Frequently asked questions

How much data does a marketing mix model need?

A long run of consistent weekly or daily history, long enough to contain more than one full seasonal cycle so that seasonality can be separated from media, and with enough variation in each channel's spend that the model has something to learn from. A channel whose spend never changes cannot be estimated, however long the history, and a channel that is always switched on and off together with another cannot be separated from it.

Does marketing mix modelling replace attribution?

No. They answer different questions at different grains: attribution describes which recorded touchpoints converting people passed through and is the right tool for tactical, within-channel decisions; mix modelling estimates each channel's aggregate contribution, including unobserved media, and is the right tool for channel-level budget planning. Mature teams run both and use experiments to calibrate the model and check the attribution story.

Can a small business use marketing mix modelling?

Usually not well, and usually not necessarily. The model needs several channels with varying spend and a long outcome history, which small businesses rarely have, and the budget questions it answers are often answerable with a simple geo test or a holdout. It becomes worth the effort when a meaningful share of spend goes to media that tracking cannot see, such as television, radio, out-of-home or podcasts, and when there is enough history to model.

Terms used on this page

Frameworks

Where this page is used

Part of these learning paths

Also related

Related