Marketing Mix Modelling vs Attribution in B2B: Two Instruments, Two Different Questions

A CFO asks one question and gets the wrong instrument almost every time. The question is "if I cut 200k from paid next year, what happens to revenue". Back comes a chart showing paid search touched 38 per cent of closed-won deals. Those are not the same subject. The chart describes what happened around deals that already closed. The question is about a counterfactual: a version of next year that has not happened yet.

What each instrument can and cannot say

Attribution is record-keeping at the level of the individual deal. It reads touch data on a person or account and assigns credit by a rule: first, last, linear, position-based, or a learned weighting. Its output is descriptive. It says which channels appeared in the path of closed deals, useful for seeing where demand comes from and for feeding conversions back to ad platforms, as closed-loop attribution sets out.

What it cannot do is say what would have happened otherwise. If a buyer would have found you through a referral anyway, last-touch still hands credit to the branded search click before the demo. Attribution has no concept of the deal you would have won anyway.

Mix modelling works at the aggregate level. It regresses revenue or pipeline against spend by channel over time, with terms for seasonality, adstock and diminishing returns. Its output is a response curve: spend this much more, expect roughly this much more, with a confidence interval. That is the instrument that answers the CFO, and the one almost nobody in B2B has enough data to fit.

Why MMM is structurally hard for most B2B companies

Four reasons, and they compound.

Not enough observations. A mix model wants two to three years of weekly data, which is 104 to 156 rows. Most B2B companies have run their current channel mix, product and pricing for eighteen months, leaving perhaps 70 usable weeks against which you fit a coefficient per channel plus seasonality plus adstock. The model will fit. It will fit noise.

Long and variable lags. In e-commerce the effect of spend lands within days. In B2B, March spend produces a September deal, and the lag differs by segment. Adstock parameters exist to model this, and estimating them well needs more data than you have, so they get assumed rather than learned: the model's core mechanism is a guess you supplied.

Spend that barely varies. Regression learns from variation. If your LinkedIn budget has been 40k a month, plus or minus 10 per cent, for two years, there is almost nothing for the model to learn from, and its estimate of what 80k would do is an extrapolation far outside the observed range. The channels you most want to reason about usually have the flattest spend history.

A few large deals dominate the series. One 900k contract in week 31 outweighs everything else in the quarter, and the model attributes that spike to whatever spend preceded it, which is coincidence. Modelling pipeline rather than revenue helps; modelling opportunity counts helps more, at the cost of ignoring deal size. Neither fixes the real problem: the outcome series has fat tails and the regression assumes it does not.

A ten-minute test before anyone builds anything. Put the last 24 months of monthly spend per channel in a spreadsheet and compute the coefficient of variation for each. If a channel varies by less than about 15 per cent around its mean, no model can say anything credible about its response curve. You never ran the experiment.

What to borrow from MMM without building one

The discipline is portable even when the model is not.

Start with the response curve's shape. MMM's central claim is that response is concave: the tenth thousand of spend does less than the first, and every channel saturates. You do not need a regression to look for that. Plot cost per qualified opportunity against monthly spend per channel for two years and see where it bends. Crude, confounded, still better than a flat efficiency average.

Second, model a baseline. MMM separates what would have happened anyway from the part marketing drove. Asking "what would have come in with zero paid spend" forces an estimate of organic, referral and existing-customer demand, the missing term in most efficiency arguments.

Third, take seasonality and carryover seriously. Pipeline in a business with European buyers drops in August and late December, and spend does not stop working when the month ends. A channel judged only on same-month conversions will always lose to one with a shorter cycle. That is how brand work gets defunded by an accounting convention.

Geo holdouts and incrementality tests are the practical middle path

If you cannot model the counterfactual, create one. This is available to companies far smaller than those that can fit an MMM, and it is the measurement that ends arguments.

Split your addressable market into two comparable sets, by country, region or metro. Where it is too small for that, hold out a matched list of target accounts from an ABM campaign instead. Keep spend running in one set, switch it off in the other, hold long enough to cover the sales cycle, and compare pipeline creation per addressable account. In DACH, splitting Germany by Bundesland gives usable cells for most B2B advertisers. Allow a pre-period of the same length to establish that the sets behaved alike beforehand.

Two rules keep holdouts honest. Fix the window and the threshold in writing before switching anything off, because criteria set afterwards are not a test. And run it long enough: with a three-month cycle, a four-week holdout measures nothing but click behaviour. Most teams never run one because switching off working spend feels like burning money. It is the only number in the building that owes nothing to a model.

They have limits. A holdout measures only the channel you switched off, in the period and geography you chose. Long-run brand effects escape a twelve-week window, and a null result is usually an underpowered test, not proof the channel does nothing.

When neither instrument is worth the effort

If you spend 30k a month across three channels and close forty deals a year, stop. There is no model to fit, no holdout with enough power, and measuring costs more than any decision it could improve. The right instrument at that scale is a clean, well-joined record of what happened in the signal ledger, an audit of whether the data flows at all, and asking buyers how they found you. Self-reported attribution on the demo form is unfashionable and, at low volume, more informative than anything computed. The same holds when your mix is not a mix: a company taking 80 per cent of pipeline from outbound and referrals has no allocation problem worth modelling.

Escalate deliberately: attribution once the joins are reliable, holdouts once a channel is large enough that being wrong costs six figures, mix modelling only with three years of history and real spend variation. Skipping ahead buys no sophistication. It buys a confident number with nothing underneath it.

Frequently asked questions

What is the difference between marketing mix modelling and attribution?

Attribution works at the level of individual deals, assigning credit to the touches that preceded them. It is descriptive: it says which channels appeared in the path of closed deals. Mix modelling works at the aggregate level, regressing revenue or pipeline against spend over time to estimate a response curve, which makes it counterfactual: it estimates what happens if spend changes. Attribution feeds conversions back to ad platforms; mix modelling answers budget allocation, and needs far more data.

Can a B2B company build a marketing mix model?

Most cannot do it credibly. A mix model wants two to three years of weekly observations, real variation in spend per channel, and an outcome series not dominated by a few large deals. Typical B2B conditions break all three: the channel mix is younger than the model needs, budgets have been flat, and one contract can outweigh a quarter. Incrementality tests answer more trustworthily for less money.

What is a geo holdout test?

A geo holdout splits the addressable market into comparable geographic sets, keeps a channel running in one and switches it off in the other, then compares pipeline creation per addressable account over a period long enough to cover the sales cycle. It measures incrementality directly rather than inferring it from a model, because the untreated set is a real counterfactual. Fix the window and the threshold in writing first.

When is marketing measurement not worth the investment?

When spend and deal volume are too low for any instrument to change a decision. A company spending 30k a month across three channels and closing forty deals a year has neither the observations for a model nor the power for a holdout. At that scale the useful work is making the data join correctly and asking buyers how they found you. Sophisticated measurement on thin data produces confident numbers with nothing behind them.