A marketing team sits down to decide next quarter's budget. Someone opens the marketing mix model, which says Meta is over-funded and the next dollar should go to connected TV. Someone else opens the attribution dashboard, which says Meta is the strongest performer in the account and deserves more. A third person remembers the geo test from March, which found that pausing Meta in eight states barely moved sales at all. Three numbers, three directions, one budget. The meeting stalls.
The instinct in that room is to decide which tool is right and trust it. That instinct is the mistake. Marketing mix modeling, multi-touch attribution, and incrementality testing are not three competing answers to one question. They are three different questions wearing similar clothes. When you understand which question each one is actually built to answer, the disagreement in the meeting stops being a problem to resolve and becomes information to use. This post, the first in a three-part series, walks through each method: what it can tell you, what it cannot, the data and time it needs, and the specific blind spot that makes it dangerous on its own.
Origin: two measurement traditions that never merged
Marketing measurement did not grow from one root. It grew from two, and they grew apart.
The older tradition is econometric. Marketing mix modeling grew out of regression work in the 1960s, when statisticians at consumer giants like Procter and Gamble related sales to spend on television, print, and radio. It moved into full commercial use around 1989 and 1990, through specialist firms such as Hudson River Group and Marketing Management Analytics, and it became standard practice for consumer packaged goods brands because those companies already had the raw material: syndicated sales data, shelf prices, and media spend, all measured in aggregate. MMM never looked at a person. It looked at a market.
The younger tradition is digital. The web could log a click, and a click could be tied to a browser through a cookie. That made a different kind of measurement possible: a person-level, touch-by-touch record of how someone reached a purchase. Last-click attribution came first, then in the mid 2000s multi-touch attribution, which promised to spread credit fairly across every touch in the journey rather than dumping it all on the final click.
For a decade these two traditions were treated as rivals, and the digital one looked like the winner. It was granular where MMM was coarse, daily where MMM was quarterly, precise to the dollar where MMM gave a range. Then a third idea arrived from a different field entirely. Incrementality testing borrows the logic of the randomized controlled trial, the same design medicine uses to test a drug. Instead of modeling who deserves credit, it runs an experiment and measures cause directly.
By the mid 2010s, analysts had a name for the recognition that no one method was enough. Forrester began arguing for a unified measurement standard that merged attribution and marketing mix modeling, a view it pushed from 2015 onward and built into its 2016 measurement research. The deliberate combination of all three methods is now more often called triangulation. The name matters less than the shift behind it, from picking a method to operating a set.
Present: three questions, not three answers
Here is the core idea, stated plainly. Each method answers a question at a different altitude, and the questions do not overlap.
MMM answers the strategic question
Marketing mix modeling asks: across the last year or two, looking at every channel including the offline and brand spend that no pixel can see, where did the budget produce the most outcome? It is a top-down statistical model. It takes weekly or daily sales, relates them to weekly or daily spend per channel, and accounts for the things that move sales regardless of marketing: price, promotions, seasonality, a competitor's launch, the weather for some categories. Google's own modern measurement playbook treats this as one leg of a tripod, the strategic view that accounts for every factor driving revenue. Singular, reading that same playbook, labels MMM the macro view: the longest-range, broadest-scope method of the three.
What MMM can do that nothing else can: it sees the whole picture. Television, radio, out-of-home, sponsorships, organic demand, all of it, because it works on aggregate data and never needs a user identity. That same design makes it immune to the privacy changes that broke cookie-based tracking. It also handles two facts about advertising that tactical tools ignore. Adstock captures how an ad keeps working for weeks after it runs. Saturation curves capture diminishing returns, the point where the next dollar into a channel buys less than the last one. Those let MMM answer the real budget question: not "did this work" but "what happens to sales if I shift a million dollars from one channel to another."
What MMM cannot do: tell you which keyword, which creative, or which audience to change this afternoon. It operates at the channel level on a weekly grain. It needs a long runway of historical data, usually around two years, and it needs the spend in that history to have actually varied, because a model learns by watching change. Its deepest limit is that it is correlational. MMM observes that spend and sales moved together. It did not run an experiment, so when a brand raises spend across several channels at once, or spends hard exactly when demand was already climbing, the model can credit a channel for sales it did not cause.
MTA answers the tactical question
Multi-touch attribution asks: inside the trackable digital journey, which touches and campaigns show up most often on the paths that end in a conversion? It is bottom-up. It stitches together the clicks and views it can observe, then assigns each one a share of the credit, either by a fixed rule or, in the data-driven version, by an algorithm trained on observed conversion patterns. In Singular's reading of the same tripod, attribution is the micro view: the most immediate level, focused on connecting a click or view to a conversion.
What MTA can do: move fast and stay granular. It can tell you within a day or a week that one creative is outpacing another, that a particular audience has gone cold, that a keyword's cost per conversion has drifted. That is precisely the signal a media buyer needs on a Tuesday to decide where the next thousand dollars goes. Funnel frames the split well: MTA answers which touchpoints weigh most heavily in the conversion process, where MMM answers which activities move revenue overall.
What MTA cannot do, and this list is long now. It cannot see anything untracked: a television spot, a podcast read, a billboard, a friend's recommendation, an Instagram impression that was viewed but not clicked. It cannot account for seasonality or a product change. It struggles with marketplace revenue and the upper funnel. Its data requirement, user-level tracking, is exactly the thing that privacy rules have degraded since 2022. And like MMM, it is correlational, with a sharper version of the same flaw. A retargeting ad shows almost entirely to people who already visited your site, so it appears right before conversions it did not create, and the model hands it credit anyway. Part 2 of this series goes deeper on why these correlational tools disagree; the contested obituary of the method is covered in multi-touch attribution is dying.
Incrementality answers the causal question
Incrementality testing asks the question the other two structurally cannot: did this specific spend actually cause outcomes that would not have happened anyway? It does not model. It experiments. The most common form for media is the geo test: you split markets into a group that gets the campaign and a comparable group that does not, run for a defined window, and measure the difference. The counterfactual, what would have happened without the campaign, is estimated from the control markets, in modern tools by building a synthetic control out of untreated geographies. The gap between what actually happened and that counterfactual is the incremental lift. Singular places incrementality between the other two as the meso view, and it tends to give the most conservative estimate of effectiveness, because it strips out the baseline sales a channel was getting credit for.
What incrementality can do: prove cause. It is the only one of the three that produces a real causal measurement, and because a geo test compares aggregate markets rather than tracking individuals, it works without user-level data and survives the privacy shift. Measured describes it as the most accurate form of media measurement for that reason.
What incrementality cannot do: cover everything, cheaply, all the time. Each test answers one narrow question about one channel or tactic in one window. Its result applies to that period and those geographies and does not automatically generalize. It is the slowest method to set up, it needs enough scale that a real effect can clear the natural noise in the data, the minimum detectable effect, and a poorly designed test, with control markets that do not match or a holdout too small to register a change, produces a confident-looking number that means nothing. You cannot run a continuous experiment on every line of the media plan.
Future and impact: why no single one is the source of truth
Lay the three side by side and the pattern is hard to miss. MMM is broad but correlational and slow. MTA is fast and granular but blind to half the funnel and also correlational. Incrementality is causal but narrow and expensive. Every method's strength sits exactly where another's weakness sits. That is not a coincidence to be optimized away. It is the reason triangulation works at all.
The market has converged on this. In a July 2025 survey by EMARKETER and TransUnion, 27.6 percent of US marketers named marketing mix modeling their single most reliable measurement methodology, the top answer, and 46.9 percent planned to invest more in it. In the same research, 52 percent already run incrementality experiments and 36.2 percent planned to spend more on them. The center of gravity has moved toward the methods that survive privacy rules and prove cause. Nobody is abandoning attribution; they are demoting it to the tactical job it is good at.
The honest version of triangulation is not "average the three numbers." It is a workflow with a direction. Fospha describes it as a sequence rather than a one-time check: model with MMM to find the strategic winners, activate with MTA to optimize day to day, validate with incrementality to confirm the causal story. The methods also calibrate each other. An incrementality result becomes a Bayesian prior that anchors the MMM to experimental ground truth. The MMM, in turn, tells you which channels are worth the cost of a test. Google's modern measurement playbook gives the cleanest rule for combining them: in-house attribution tends to run generous, incrementality runs conservative, and a healthy MMM result should land between those two bounds. When it does, you can act. When it does not, you have found a problem worth investigating.
Scale decides how much of this a brand needs. A common framing in 2026, from ClickZ among others, is that under roughly two million dollars in annual paid media, attribution plus the occasional test is enough; the full three-layer stack earns its complexity past about twenty million, where a wrong budget call costs more than the measurement does. Below that, MMM in particular struggles, a threshold covered in is your budget big enough for MMM.
The risk to name plainly: triangulation does not hand you one clean number. It hands you three estimates that will not perfectly agree, and the work is reconciling them into a decision you can defend. That reconciliation is a real skill, and getting it wrong, by quietly trusting whichever number is highest, is worse than using one method honestly. That is the subject of part 2, reconciling three measurement numbers into one you can trust.
For now, the takeaway is the reframe. Stop asking which measurement method is correct. MMM is correct about strategy. MTA is correct about tactics. Incrementality is correct about cause. None of them is correct about all three, and the brands that measure well are the ones that stopped looking for a single source of truth and started running a set of methods that each cover for the others. The MMM renaissance that made this practical is covered in marketing mix modeling is back.
Council summary
This post argues that marketing mix modeling, multi-touch attribution, and incrementality testing are not three contenders for a single truth but three instruments tuned to different questions: strategy, tactics, and cause. Its strongest move is the fair can-and-cannot treatment of each method, naming the specific blind spot, the correlational trap for MMM and MTA and the narrow scope for incrementality, that makes any one of them dangerous when trusted alone. The reader's takeaway is the reframe: stop asking which number is right, and start reading disagreement between the three as information rather than error. The piece sets up the harder problem honestly, which is the reconciliation work itself, and hands it cleanly to Part 2. Accurate, well-sourced, and free of filler, it earns a pass.
Comments