MMM, multi-touch attribution or incrementality?

MMM vs MTA vs incrementality: the three methods in one table

MMM vs MTA is the comparison most measurement teams start with, and incrementality testing is the third method that settles what the first two disagree on. Whether you search for multi touch attribution vs marketing mix modeling, media mix modeling vs multi-touch attribution, MTA vs MMM or attribution modeling vs marketing mix modeling, the question behind the words is the same. The three methods answer different questions, at different levels, with different kinds of evidence. Choosing between them starts with the question you need answered.

	Multi-touch attribution (MTA)	Marketing mix modeling (MMM)	Incrementality testing
Unit	The user, along a tracked path	The channel, by week, in aggregate	A region, a user group or a store group, treated versus untreated
What it measures	Which tracked touchpoints preceded each conversion, credited by a rule	An estimate of what each channel added to sales over the period, and of the return of the next euro	What one change in advertising caused, at one spend level, over one period
What it needs	User-level tracking of the touchpoints	Two to three years of weekly spend by channel, sales and controls, with variation in spend	A clean intervention, comparable groups and enough size to detect the effect
Sees offline media	No	Yes	Yes, when the unit is geographic
Causal	No: correlation along a path	Observational: needs calibration	Yes, under the conditions of the test
Latency	Daily	A fit takes weeks; refits as experiments read out	A test runs for weeks and answers one question
Best use	Optimising inside a channel: keyword, audience, creative	Allocating the budget across channels	Measuring a channel’s incremental effect under test conditions, and calibrating the model
Typical failure	Overstating channels that reach people who were converting anyway	Crediting the wrong channel when channels move together	Testing too small, too short, or measuring a cut that was not designed as a test

The columns are not alternatives to pick one from. They are three instruments, and a measurement set-up that lacks one of them usually has a blind spot the other two cannot cover.

Multi-touch attribution: what it measures, what it misses

Multi-touch attribution follows individuals across the touchpoints its tracking can see and shares the credit for each conversion between those touchpoints by a rule: last click, linear, time-decay, position-based, or a data-driven weighting fitted on the paths. It is granular, available daily, and built into the platforms and analytics tools. For choosing between two keywords, two audiences or two creatives inside a channel, it is the usual instrument.

What it misses is structural. It only sees what it can track, which excludes television, radio, out-of-home and print, and a growing share of users as consent rules and tracking restrictions remove them from the path. It counts correlation along a path, not causation: a touchpoint that appears before a conversion gets credit whether or not the conversion would have happened without it, which is why retargeting and branded search look so good in attribution. And its rules share credit only among the touchpoints that were observed, never with the ones that were not, so the attributed sales of every channel add up to more than the sales the business made. Multi-touch attribution vs single-touch is a choice between rules; neither rule adds a counterfactual.

For the question a budget needs, marketing mix modeling vs attribution is not a close call: attribution modeling cannot say what a channel added, only what it touched.

Marketing mix modeling: what it measures, what it misses

A media mix model explains weekly sales as a baseline plus the contribution of each channel, after adstock for carryover and saturation for diminishing returns, with controls for price, promotions, seasonality and the rest. It works on aggregate data, needs no user tracking, and sees offline media alongside digital, which is what makes it the instrument for allocating a budget across channels. Its output per channel is a response curve, and the slope of that curve at the current spend, the marginal incremental ROAS, is the number a plan moves budget on.

What it misses follows from how it works. It is observational: a well-fitted model can still credit the wrong channel when two channels always move together, and everyone spends more in the fourth quarter. It needs history and variation, typically two to three years of weekly data, and a channel whose spend never changed is not identifiable at all. Its resolution stops at the channel and the week: it will not choose a creative or a keyword. And beyond the spend range it has seen, its curves are extrapolations. Media mix modeling vs attribution modeling is therefore not a contest but a division of labour, with one blind spot left: the model’s own attribution has to be checked against something that measures causation.

Incrementality testing: what it measures, what it misses

An incrementality test creates the counterfactual on purpose. Some units receive a change in advertising and comparable units do not, and the difference in outcome, with its interval, is what the change caused. The unit that works without user data is geography: a geo-lift experiment pauses or boosts a channel in some regions and reads the gap against a synthetic control built from the others. Platform lift studies, customer-list holdouts and A/B tests do the same at the user level, inside one platform or one list.

What it measures is the one thing the other two cannot: causation, under the conditions of the test. What it misses is everything outside those conditions. A test answers one question, for one channel, at one spend level, over one period; it is a point on the response curve, not the curve. It costs money and time, needs a clearly defined intervention and enough size to detect the effect the plan cares about, and a spend cut that happened for other reasons does not become a designed test because it is measured after the fact. And a platform measuring its own lift is measuring inside its own audience and rules. Lift vs incrementality, in that sense, is the difference between a platform’s number and a designed experiment’s number.

Which one, when: our opinion

Use all three, in a hierarchy, and do not let any of them do another’s job.

Experiments calibrate the model. The measured return of a geo test or a holdout becomes a prior on that channel’s coefficient in the next fit, and when the model and the test differ by more than a third, that flags an identification problem on that channel and the plan should say so. The next test goes where the model is least sure: the channel whose band is too wide to act on, or whose spend never varied.

The model allocates the budget across channels. Online and offline, weekly, on marginal incremental ROAS with its band, and with the benchmarks it still depends on listed and ranked. MMM vs incrementality is not a choice: one generalises what the other measured.

Attribution optimises inside each channel. Keywords, audiences, creatives, with its figures read as attributed, not incremental. Incrementality vs attribution is only a conflict when attribution’s numbers are used to move budget between channels; used inside a channel, they are the fastest signal there is.

The opinion, in one line: attribution vs incrementality is a question about what the number means, and a plan should never mix the two without saying which is which. That is why every figure in a Kuwalyst plan carries a tag, measured, projected or benchmark, and why the Media Planner reads platform figures as attributed and allocates on the model.

Three short cases

The three cases are fictional and simplified; the retailer, the direct-to-consumer brand and the agency are not Kuwalyst clients.

A retailer with television and search. Attribution credits most sales to branded search; the model, fitted on two years of weekly data, finds that television explains part of the branded-search volume and gives television a marginal return the dashboards never showed. The team runs a six-week geo holdout on television; the measured return sits within the model’s band, the model is refitted with it as a prior, and the next plan moves budget from search, which was saturated, to television. Attribution kept its job: choosing the search keywords.

A direct-to-consumer brand on paid social only. With one dominant channel and eighteen months of history, a media mix model would return its prior. The team skips the model and runs a platform lift study, then a customer-list holdout on retargeting; the retargeting lift is not distinguishable from zero, so the team stops increasing it and tests prospecting next. The model comes later, once a second channel and another year of data exist.

An agency with twelve clients on the same platforms. Attribution is what every client’s dashboard shows, and every client asks why the numbers do not add up. The agency fits a model per client where the history allows, starts the others on audited benchmarks with the tests to run first, and reports every figure with its tag. The question shifts from which dashboard is right to which figures are measured, which are projected, and what the next test is. Incremental lift testing vs A/B testing Facebook campaigns stops being a debate: the A/B test picks the creative, the lift test measures the channel.