Attribution vs MMM vs Incrementality: Which Actually Measures Marketing?

TL;DR — There are three legitimate ways to measure marketing's effect: attribution (which tracks who clicked), marketing mix modelling (which estimates what spend moved the total), and incrementality testing (which proves causality with experiments). None of them is correct on its own. The honest answer is that they measure different things, each is biased in a known way, and mature measurement stacks use all three — attribution for tactical optimisation, MMM for budget allocation, incrementality to keep the other two honest.

Attribution, marketing mix modelling (MMM), and incrementality testing are the three families of measurement that serious marketing organisations rely on. They are often presented as rivals — as if you must pick a champion. They are not. They answer different questions, they fail in different ways, and the practitioners who get measurement right use them in combination. This review lays out what each one actually does, where each one misleads, and how to choose — or, more often, how to combine them — based on the size and maturity of your business.

What is marketing attribution? (and why last-click misleads)

Marketing attribution is the practice of assigning credit for a sale or conversion to the marketing touchpoints a customer was exposed to before buying. In its most common form — last-click attribution — 100% of the credit goes to the final click before the conversion, regardless of everything that came before it. Multi-touch attribution (MTA) spreads that credit across several touches using rules (linear, time-decay, position-based) or statistical models.

The appeal is obvious: it is granular, real-time, and ties spend to an outcome at the keyword and campaign level. It is what most digital dashboards default to.

The problem is that last-click attribution is systematically wrong in a known direction. It over-credits the bottom of the funnel and starves the top. A buyer may see a brand campaign on television, hear about it from a friend, search for it three weeks later, and then convert on a branded search term. Last-click hands the entire win to that final branded click — the cheapest, most bottom-funnel touch — and records nothing for the work that actually created the demand. Over time, optimising to this signal quietly defunds the brand building that drives most of the value, a pattern Binet and Field have documented across the IPA Effectiveness Databank as one of the great effectiveness failures of the digital era.

There are two further, structural problems that have made attribution steadily less reliable:

Walled gardens and fragmentation. Attribution only "sees" touches within a single platform's logs. A journey that spans Google, Meta, a retailer site, and an offline store is invisible to any one of them. Each platform reports as if it were the whole story.
Privacy and signal loss. As third-party cookies, mobile ad identifiers, and cross-site tracking have been restricted, deterministic user-level attribution has degraded. Platforms have responded by moving to modelled conversions — estimates, not observed clicks — which imports modelling assumptions into what looks like hard data.

The honest summary: attribution is an excellent tactical tool for optimising within a single, lower-funnel, digital channel. It is a poor basis for budget allocation across channels, and a dangerous basis for deciding whether marketing as a whole works.

What is Marketing Mix Modelling (MMM)?

Marketing mix modelling is a top-down, econometric method. Rather than tracking individual users, MMM takes aggregate time-series data — sales, media spend by channel, price, promotion, seasonality, macro factors — and uses regression to estimate how much each input contributed to the outcome. It has been used since the 1960s, originally for television-heavy packaged-goods advertisers, and it answers a question attribution fundamentally cannot: how much did each channel and the campaign as a whole move total sales, including the sales no one clicked on?

This is MMM's decisive strength. Because it works on totals, it captures offline media, brand effects, and the long, lagged payback of brand building — exactly the things attribution is blind to. Binet and Field's effectiveness work treats econometric modelling as one of the most rigorous ways to prove commercial return, which is why it has long been accepted evidence in the IPA Effectiveness Awards.

MMM also has well-understood weaknesses. It is not real-time — a robust model needs months or years of history and typically reports with a lag, so it cannot steer next week's bidding. Its accuracy depends entirely on the quality and granularity of the input data and the skill of the modeller: feed it bad spend data or omit a key driver (a competitor price cut, a distribution change) and it will confidently produce the wrong answer. And because it estimates averages over a period, it is poor at the granular, campaign-level detail where attribution shines.

After years in the shadow of digital attribution, MMM has staged a genuine revival — driven precisely by the signal loss that undermined attribution. Two open-source contributions matter here: Meta released Robyn (an open-source MMM package) and Google released Lightweight MMM, with Google later introducing Meridian. By open-sourcing the methodology, the major platforms effectively conceded that aggregate, privacy-resilient modelling is the future of cross-channel measurement. That is as clear a signal from the industry as you are likely to get.

What is incrementality testing? (the gold standard)

Incrementality testing measures causality directly, by running experiments. The core idea is a counterfactual: what would have happened without the marketing? You establish that by holding the marketing out for a comparable group — typically through geo-experiments (running the campaign in some markets and withholding it in matched control markets), holdout tests (randomly withholding ads from a subset of users), or ghost-bid/ghost-ad studies.

Where attribution assigns observed credit and MMM estimates statistical contribution, incrementality proves whether the spend caused additional outcomes that would not have happened otherwise. That is why it is described as the gold standard for causal claims: a well-designed geo-test answers "did this money actually create sales?" in a way neither of the other methods can.

The trade-off is cost and scope. Real experiments are expensive to design, slow to run, and hard to scale across every campaign and channel. They also measure one intervention at a specific moment — they do not, on their own, produce a continuous allocation model. The most powerful use of incrementality is therefore as a calibration layer: run periodic holdout or geo-tests to measure the true incrementality of a channel, then use those results to correct the biases in your attribution and MMM. When your attribution says a channel delivered 1,000 conversions but a holdout test shows only 400 of them were incremental, you now have an honest number to work with.

The platforms now offer this functionality directly — Google and Meta both provide lift-testing and conversion-lift products — which has made causal measurement far more accessible than when it required bespoke econometricians.

Attribution vs MMM vs Incrementality (comparison)

Approach	What it measures	Best at	Where it misleads	When to use it
Attribution (last-click / MTA)	Observed touchpoints per user journey	Granular, real-time, within-channel optimisation	Over-credits bottom-funnel and branded search; blind to offline and brand effects; degraded by privacy loss	Tactical bidding and creative testing inside a single digital channel
Marketing Mix Modelling	Aggregate contribution of all inputs to total sales	Cross-channel budget allocation; captures offline and brand-building returns; privacy-resilient	Slow (lags of weeks/months); depends on data quality and modeller skill; reports averages, not granular detail	Annual or quarterly budget setting and total marketing ROI
Incrementality testing	Causal lift — what the spend actually created	Proving causality; calibrating the bias in the other two	Expensive, slow, one experiment at a time; does not produce a continuous model	Periodic validation of channel value and major budget decisions

Read the table as complementary, not competitive. The three methods sit at different levels of the measurement stack and have different failure modes — which is exactly why combining them beats choosing between them.

Which should you use? (by maturity and size)

The honest answer is all three, eventually — but staged by what your business can actually support.

If you are small or early-stage. You almost certainly cannot afford a full MMM or a programme of geo-experiments. Your best measurement is disciplined incrementality in its cheapest form: structured holdout tests and pre/post comparisons on your most expensive channels, combined with platform attribution for day-to-day optimisation. Do not over-invest in attribution granularity you cannot act on. The bigger risk at this stage is measuring the wrong thing precisely.

If you are a mid-scale advertiser spending meaningfully across channels. This is where MMM starts to pay for itself. Once you have a year or more of clean spend and sales data and enough budget that allocation errors cost real money, a periodic MMM (rebuilt annually, refreshed quarterly) gives you a defensible basis for the cross-channel split. Layer in occasional incrementality tests to sanity-check the channels where MMM is least certain.

If you are a large, mature advertiser. Run the full stack: real-time attribution for tactical optimisation, MMM for strategic allocation and long-term ROI, and a rolling programme of incrementality testing to calibrate both. Feed the experiment results back into the MMM as priors and into the attribution as correction factors. This is what the best marketing organisations do, and it is the only configuration that holds up under scrutiny from a sceptical finance team.

A note on the temptation to declare a winner. The recurring failure mode in this field is a measurement team falling in love with one method — usually whichever is newest — and dismantling the others. Attribution-only shops misallocate to the bottom of the funnel. MMM-only shops lose tactical agility. Incrementality-only shops run beautiful experiments they cannot scale into a plan. The discipline is in holding the tension between them.

Where people get it wrong

Treating attribution as ground truth. It is an allocation heuristic with a known bias, not a measurement of causality. Decisions that hinge on "this channel has the best ROAS" are usually decisions about last-click noise.
Expecting MMM to be real-time. It isn't, and forcing it to be produces unstable, unreliable models. Use it for the questions it can actually answer.
Running incrementality tests but never acting on the results. An experiment that contradicts your dashboard is the point, not a problem to be explained away.
Optimising to the metric instead of to marketing effectiveness. The deepest error is measuring what is easy rather than what matters. No measurement method fixes a business that is tracking the wrong outcome.

Key takeaways

Attribution, MMM, and incrementality answer different questions and fail in different ways — they are complements, not substitutes.
Attribution is granular and real-time but systematically over-credits bottom-funnel clicks and is degrading with privacy change.
MMM is the only one that captures cross-channel, offline, and brand-building returns; it is slow and depends on data quality.
Incrementality testing is the only one that proves causality; its best use is as a calibration layer for the other two.
Mature organisations use all three: attribution for tactics, MMM for strategy, incrementality to keep both honest.
The question is never "which is best?" — it is "what can my business reliably run, and am I measuring causality or just activity?"

Marketing Effectiveness — the discipline these methods exist to serve
Marketing ROI — how honest return is actually calculated
Multi-Touch Attribution — the deeper dive on attribution's mechanics and limits
Marketing Mix Modelling — the econometric method in detail
Incrementality Testing — causal measurement, properly explained

Evidence & sources

Binet, L. & Field, P. (2007). Marketing in the Era of Accountability. IPA — treats econometric modelling as a rigorous basis for proving commercial return.
Binet, L. & Field, P. (2013). The Long and the Short of It. IPA — the evidence base for long-term brand-building returns and the failure of short-term, bottom-funnel optimisation.
Sharp, B. (2010). How Brands Grow. Oxford University Press — mental availability and the mechanisms that make top-funnel, broad-reach brand building effective.
Ehrenberg-Bass Institute publications — on how brands grow and the measurement of advertising effects.
WARC — ongoing coverage of measurement practice, MMM methodology, and effectiveness benchmarks.
Meta Robyn (open-source MMM) and Google Lightweight MMM / Meridian — platform-published, attributed Tier-2 sources on modern MMM methodology and the move to privacy-resilient modelling.
Google and Meta conversion-lift / geo-experimentation product documentation — Tier-2, attributed, on incrementality testing as a platform capability.