Multi-touch attribution and marketing mix modelling are answering different questions

There is a well-worn slide in this industry showing multi-touch attribution and marketing mix modelling on opposite sides of a comparison table, with a recommendation at the bottom that happens to match whatever the presenter sells. The framing is wrong at the root. These are not competing solutions to one problem; they are different instruments pointed at different questions, and a business can genuinely need both, either, or neither.

Multi-touch attribution measures the journeys you can see. Marketing mix modelling infers the effects you cannot. Asking which is better is asking whether a microscope beats a telescope.

This article is what each actually does, where each breaks, why they disagree, and what a mid-market business should realistically do given that both are expensive and one of them may be unbuildable on your data.

What multi-touch attribution does

Multi-touch attribution stitches together the observable interactions of individual customers and applies a rule to divide credit among them. Its unit of analysis is the person, and its reach is limited to whatever it can observe about that person.

The appeal is granularity. It can tell you that a particular campaign, at a particular time, appeared in the journeys of customers who spent a particular amount. That specificity is genuinely useful for optimisation within digital channels, and it is the level at which most day-to-day marketing decisions are made.

The structural limitation is that it can only reason about what it observes. Anything happening off-platform — a conversation, a recommendation, a billboard, an article read without clicking, an AI answer summarising your position — is invisible, and invisible touches receive no credit. The method is not neutral about this; it systematically transfers credit to the channels that happen to be trackable.

That bias has grown rather than shrunk. Browser tracking prevention, mobile platform changes, consent requirements and the rise of zero-click search have all reduced observability over the last several years, which means multi-touch attribution now sees a smaller and less representative slice of the journey than it did when the technique became popular.

What marketing mix modelling does

Marketing mix modelling uses statistical regression on aggregate time series — spend by channel, sales, and external factors like seasonality and price — to estimate how much each channel contributed. It never looks at an individual and therefore never loses one.

The appeal is coverage. Because it works on aggregates, it can include television, radio, out-of-home, sponsorship, print and anything else with a spend figure, alongside digital. It is unaffected by cookie loss, consent rates or platform changes, which makes it more durable than tracking-based methods.

The limitation is resolution and data appetite. A model needs several years of history and, crucially, variation in spend: if a channel's budget has been flat for three years, there is no signal from which to estimate its effect. It produces broad estimates with genuine uncertainty ranges, and it cannot tell you anything about a campaign that ran for three weeks.

It is also easy to build badly. The number of modelling choices — how to handle diminishing returns, how long effects persist, which external factors to include — is large, and different reasonable choices produce materially different answers. A model whose assumptions are not visible is not evidence.

The comparison that matters

The useful comparison is not accuracy but applicability: what each can include, what decision each supports, what it costs, and how quickly it answers.

Read down the rows and the complementarity is obvious. Each one's weakness is the other's strength, which is precisely why the vendor framing of them as alternatives is so misleading. Large advertisers that can afford both run both, and use the disagreement between them as information.

Why they disagree, and what the disagreement tells you

Multi-touch attribution typically credits trackable digital channels more highly than marketing mix modelling does, because it can only see those channels. The size of the gap is a rough measure of how much of your demand generation is invisible to tracking.

This is the single most useful thing to do with both outputs, and it is rarely done because the two are usually owned by different teams and presented in different meetings. Putting the two channel rankings side by side and examining the largest divergences is more informative than either ranking alone.

A channel that scores well in tracking and poorly in modelling is often intercepting demand rather than creating it — branded search and retargeting are the classic examples. A channel that scores poorly in tracking and well in modelling is usually creating demand that surfaces elsewhere, which is the normal signature of upper-funnel media.

Neither reading is definitive, and both are hypotheses worth testing with a holdout. The value is in generating the hypothesis, which neither instrument does alone.

The third method nobody puts on the slide

Incrementality testing withholds marketing from a comparable group and measures the difference. It is the only one of the three that observes causation rather than inferring it, and it answers one narrow question at a time.

A geographic holdout is the usual form: suspend a channel in a set of matched regions, keep it running elsewhere, and compare. The result is a genuine causal estimate for that channel, in that period, at that spend level, which is both more trustworthy and considerably narrower than what the other methods produce.

The costs are real and should not be minimised. You give up revenue in the held-out regions, the test takes weeks, and it needs enough volume for the difference to be detectable. In a business with modest volume the test may be unable to resolve anything, which is worth establishing before running it rather than after.

Used sparingly on the largest line items, it is the highest-value measurement most businesses are not doing. One well-designed holdout on the biggest channel each year will catch the errors that both other methods are structurally prone to.

What a mid-market business should actually do

Match closed sales to your own source records for the operational picture, and run occasional holdout tests on the largest channels for causal evidence. That combination costs a fraction of either formal method and answers most of the questions a mid-market business faces.

The reason to start with matching is that it is verifiable. You can point at a sale, point at the enquiry it came from, and show both records. Neither multi-touch attribution nor mix modelling produces anything you can check that way, and for a business establishing trust in its numbers for the first time, checkability is worth more than sophistication.

It is also achievable in a week from data you already have, which matters more than it should. Measurement programmes that produce a finding quickly survive; those that require a year of build compete for patience they have not earned.

CloseRev is deliberately the simple end of this spectrum: it reconciles closed sales against your own lead and call records and shows what matched and what did not. It is not a mix model and does not pretend to establish causation.

When you genuinely need mix modelling

Consider marketing mix modelling when a material share of spend is on channels that cannot be tracked at all, when you have several years of varied spend history, and when the budget at stake justifies months of work.

The first condition is the important one. If ninety percent of your spend is digital and trackable, a mix model will spend enormous effort estimating things you can observe directly. If forty percent is television, radio, sponsorship and trade, no tracking-based method will ever see it and the model is the only instrument available.

The second condition disqualifies more businesses than they expect. Flat spending produces no variation, and no variation produces no estimate. If your budget has been broadly constant for three years, the honest answer is that a model cannot be built usefully on your data yet, and a deliberate programme of varying spend is the prerequisite.

The organisational reason these projects fail

Both methods fail more often for organisational reasons than technical ones: no single owner, no agreed decision the output will inform, and no agreement in advance about what result would change behaviour.

The pattern is consistent enough to predict. A measurement project is commissioned because the numbers are unsatisfactory, it takes six to twelve months, and it delivers an answer that contradicts somebody influential. Because nobody agreed beforehand what would follow from that answer, the finding is debated rather than acted on, and the programme quietly loses its sponsor within two quarters.

The protection is to write down, before commissioning anything, the specific decision the output will inform and the threshold at which it changes. If the model says this channel returns less than a stated figure, we will move a stated amount of budget. Committing to that in advance is uncomfortable, which is exactly why it works.

It also disciplines the scope. A great many measurement programmes are commissioned without any decision attached at all, purely because the current numbers feel unsatisfactory. Those programmes produce reports rather than changes, and the reports get shorter every quarter until they stop.

Where the two methods are converging

The serious work in this field is moving toward calibrating one method with the other: using experimental results to constrain a mix model, and using the model to sense-check journey-level attribution. That is a research direction, not a product you can buy today.

The idea is sound and worth understanding even if you cannot implement it. A holdout test produces a trustworthy causal estimate for one channel; that estimate can be used as a prior or a constraint when fitting a mix model, which makes the model's other estimates more credible. Several large advertisers and the open-source modelling projects are working along these lines.

What this means practically for a mid-market business is modest but real: any holdout test you run has value beyond its immediate answer, because it becomes a fixed point that future modelling work can be anchored to. Recording the design and results of tests carefully is therefore worth more than it appears at the time.

It is also a reason to be sceptical of unification claims from vendors. The organisations closest to this problem describe it as difficult and partially solved. A product asserting that it has been fully solved is not reporting from the same frontier.

A realistic three-year sequence

Establish matched revenue reporting in year one, add annual holdout tests on the largest channels in year two, and consider mix modelling in year three only if a large share of spend remains untrackable and spend has varied enough to model.

The sequencing matters because each stage produces the evidence that justifies the next. Matched reporting reveals how much revenue is unattributable, which tells you whether the invisible portion is large enough to warrant modelling. Holdout tests reveal whether your matched picture is systematically biased, which tells you how much to trust it.

Skipping to the sophisticated end is the common error and it fails predictably. A mix model commissioned by a business that cannot yet reconcile its own revenue to its own sales system will produce an answer nobody can validate against anything, and it will be disputed by whoever it makes look bad, on grounds nobody can settle.

Three years sounds slow to anybody who has just sat through a vendor pitch. It is considerably faster than the actual observed timeline of businesses that start at the sophisticated end, because those programmes usually restart at least once.

The vendor conversation

Ask any vendor which of the three methods their product implements, what it cannot see, and what would have to be true for its answer to be wrong. A vendor who cannot describe their own blind spot is selling confidence rather than measurement.

The question about blind spots is the diagnostic one. Every method here has a well-known and easily articulated weakness, and a competent vendor will name theirs immediately and explain how they mitigate it. Hesitation or a claim to have solved it is the signal to be careful.

Be particularly cautious about products that claim to unify all three. That is a genuine research problem, several serious organisations are working on it, and nobody has a clean answer. A confident claim to have one from a mid-market vendor is a marketing position rather than a technical achievement.

Every measurement method here is wrong in a specific, known, describable way. The ones worth buying are sold by people who will tell you which way theirs is wrong.

Questions people actually ask

What is the difference between MTA and MMM?

Multi-touch attribution follows individual customer journeys and divides credit among the touches it can observe. Marketing mix modelling ignores individuals entirely and uses statistics on aggregate spend and outcomes over time to estimate each channel's contribution. One is a tracking method, the other is an econometric one.

Which is more accurate, MTA or MMM?

They are accurate about different things. MTA is precise about the journeys it can see and blind to everything else. MMM can include offline and unmeasurable channels but produces broad estimates with wide uncertainty. Neither is a more accurate version of the other.

Why do MTA and MMM give different answers?

Because MTA only sees trackable digital touches and so overstates them, while MMM captures broad effects and cannot resolve short-term or narrowly targeted activity. A systematic disagreement is expected and is itself informative, usually about how much of your demand is invisible to tracking.

Is marketing mix modelling only for large advertisers?

Traditionally yes, because it needs several years of history and meaningful variation in spend. Open-source implementations have lowered the cost, but the data requirement has not changed, and a business with two years of flat spending cannot produce a useful model at any price.

What should a mid-market business use instead?

Match closed sales to your own source records for the operational picture, and run occasional holdout tests on your largest channels for causal evidence. That combination is cheaper than either MTA or MMM and answers most real questions.

Does incrementality testing replace both?

It answers the causal question better than either and answers only one question at a time, slowly and at the cost of foregone revenue. It is the audit, not the reporting system.

Found this useful?

• Share on X
• Share on LinkedIn
• Share on Reddit
• Share on Email