Marketing Engineering

Marketing Mix Modeling:
End of Attribution Mythology

By ·December 2025·18 min read·Marketing Engineering
Share

Last-click attribution tells you which channel received credit for a purchase. It tells you nothing about which channel caused it. For most organizations, this distinction is the difference between confident allocation and structural mis-allocation. Last-click is comforting. It is also mythology.

Marketing Mix Modeling answers the question that attribution cannot: out of all the variables that could have driven this revenue, marketing channels, price, seasonality, competitive activity, macroeconomic conditions, which ones actually caused the result, and how much did each contribute?

The Attribution Trap

Last-click rewards the channel closest to conversion. Paid search, retargeting, and direct-response email systematically over-receive credit. Upper-funnel investment, brand campaigns, content marketing, social, systematically under-receives credit. The result is a budget that drifts toward last-click channels not because they cause the most revenue, but because they collect the most credit for revenue caused by others.

Multi-touch attribution does not solve the problem. It distributes the same observable touchpoints differently. It still cannot see the brand campaign that primed the customer, the content article that built consideration, the macroeconomic tailwind that drove demand. It is more sophisticated mythology.

Figure 01 · Credit vs cause
Where last-click sends the budget
Last-click creditTrue contribution (MMM)
Brand campaignsUpper funnel
6%
19%
Under-credited by 13 pts
Content marketingUpper funnel
5%
14%
Under-credited by 9 pts
SocialUpper funnel
9%
16%
Under-credited by 7 pts
Paid searchLower funnel
31%
21%
Over-credited by 10 pts
RetargetingLower funnel
22%
12%
Over-credited by 10 pts
Direct-response emailLower funnel
27%
18%
Over-credited by 9 pts
Last-click rewards the channel closest to conversion. The gap between the credit a channel collects and the contribution it actually drives is where budgets quietly drift toward the wrong places.
Illustrative decomposition

What MMM Actually Does

MMM uses econometric regression on aggregated, weekly data. Revenue is the dependent variable. Marketing spend by channel, price, promotions, seasonality, competitive activity, and external indicators are independent variables. The model decomposes revenue across drivers and produces the marginal contribution of each.

Critically, MMM works on aggregate data. It does not require user-level tracking. It survives privacy changes, cookie deprecation, and platform black boxes that have made user-level attribution increasingly fragile. This is one reason MMM has surged back into prominence over the last five years.

Figure 02 · Decomposition
One quarter of revenue, attributed to its true drivers
Total revenue100%
Base demand0%
Brand & upper funnel0%
Paid search0%
Video & TV0%
Social0%
Email & CRM0%
Price & promotion0%
Seasonality0%
MMM separates revenue that marketing caused from revenue that would have arrived anyway. Here, base demand explains 42 percent. The remaining 58 percent is the contestable territory, and 51 percent of total revenue traces to specific, marginal marketing investment.
Illustrative model output

Adstock: The Lagged Effect Attribution Cannot See

A TV ad seen on Monday influences purchase behavior on Wednesday, next week, and sometimes next month. Attribution models, observing only the touchpoint nearest conversion, are structurally blind to this carryover. MMM models adstock explicitly through decay functions, capturing both the immediate and the lagged effect.

For upper-funnel media, the majority of the commercial impact occurs after the initial exposure. Treating these channels as if they only worked at the moment of click systematically undervalues them and migrates budget toward downstream channels that depend on the upstream demand.

Figure 03 · Adstock
The effect attribution cannot see
2.5 weeks
Performance searchBrand & video
01234567891011weeks after exposure
25%
lands in week 0
75%
carries over later
A single burst of media keeps working for weeks. Drag the half-life to match a channel: performance search decays fast, brand and video carry for over a month. At this setting, 75 percent of the impact lands after the week of exposure, invisible to a last-click model.
Geometric adstock, normalised over 12 weeks

Bayesian MMM Over Classical OLS

Classical MMM uses ordinary least squares to fit the model to data. Bayesian MMM adds prior beliefs to the data, constraints derived from known media physics (TV spend cannot produce negative returns), the half-life of media effects, the shape of saturation curves. Priors regularize the model where data is sparse and produce probability distributions over parameters instead of single point estimates.

The result is a model that is more robust with limited data, better at quantifying uncertainty (which matters for budget decisions), and better suited to optimization under uncertainty. Google's Meridian, Meta's Robyn, and PyMC-Marketing have democratized the technical infrastructure. The methodological consensus is firmly Bayesian.

Figure 04 · Uncertainty as an output
A distribution beats a decimal
90% credible: 2.3x to 3.7x
1x2x3x4x5x6xOLS 3.6x3.0x
Ordinary least squares returns one number and hides its own fragility. Bayesian MMM returns a probability distribution: the most likely return, and an honest range around it. Priors drawn from media physics regularize the estimate where data is thin, which is exactly where a single point estimate is most dangerous.
Posterior over channel ROAS, 90% credible interval

The Data Requirements

A well-specified MMM needs at least 2 years (ideally 3+) of weekly data covering revenue or sales volume, marketing spend by channel, pricing history, promotional activity, and external variables such as economic indicators and seasonal indices. Data quality is the most common limiting constraint. Models are only as good as the data that feeds them.

Going deeper

The Saturation Curve

Every channel has a point where the next dollar earns less than the one before it. The saturation curve maps exactly that. Early spend buys incremental customers cheaply. As reach saturates, each additional dollar competes for attention you have already paid for, and the marginal return bends toward zero.

MMM estimates the shape of this curve for every channel, which is what makes it more than a backward-looking report. It tells you not just what worked, but how much room each channel still has before it stops working. A channel at 40 percent of its saturation point is an investment. The same channel at 90 percent is a leak. The budget allocator above runs on exactly these curves: it moves the next dollar to wherever the slope is still steep.

The Validation Discipline

MMM is statistical inference, not measurement. Its results require validation. Incrementality testing (geo-lift, holdout cohorts) provides causal validation by running controlled experiments in real markets. MMM identifies the expected contribution of each channel; incrementality tests measure the actual lift by turning a channel off in one geography and comparing against a matched control.

The two methods are complementary. MMM provides strategic direction across the entire media mix. Incrementality testing validates specific causal claims. Organizations that use both achieve the highest confidence in budget allocation decisions.

Figure 05 · The decision MMM enables
Allocate the budget against diminishing returns
$2.7M
Total spend
$5.4M
Modeled contribution
2.00x
Blended ROAS
Paid search$0.90M spend · $2.00M contribution
0.79x marginal
Social$0.60M spend · $1.04M contribution
1.09x marginal
Video & TV$0.50M spend · $0.96M contribution
1.62x marginal
Email & CRM$0.30M spend · $0.82M contribution
1.28x marginal
Brand$0.40M spend · $0.57M contribution
1.28x marginal
At the optimal allocation every channel returns the same on its next dollar. Channels in red sit below the portfolio average: move budget to the ones with headroom.
Every channel saturates. The model knows the shape of each curve, so it can answer the only question that matters at the margin: where does the next dollar earn the most? When a channel’s marginal return falls below the blended return, it is over-funded. Move the sliders and watch the trade-off.
Illustrative saturation model · figures in $M

Who Should Build MMM

Historically, MMM required large budgets and specialist teams. That has changed. Open-source Bayesian frameworks have brought the technical infrastructure within reach of any organization spending $5M+ annually on media. The constraint is no longer the model. It is the data infrastructure that feeds the model and the organizational discipline to act on what the model finds.

An honest limit

What MMM Cannot Do

MMM is inference, not omniscience. It cannot see a channel it was never given data for. It struggles to separate two channels that always move together, a problem called collinearity, which is why media plans with more variation produce better models. It captures the spend behind creative, not the quality of the creative itself. And it answers questions at the level of weeks and channels, not individual customers, which is precisely why it survived the collapse of user-level tracking.

Naming these limits is not a weakness of the method. It is the difference between a model used well and a model trusted blindly. The strongest measurement programs pair MMM with controlled experiments for exactly this reason: each covers the other’s blind spot.

The takeaways
Interactive · Scorecard
How mature is your marketing measurement?

Five questions. You will get a maturity score and where to focus next. Answer honestly, the result is only useful if it reflects how decisions are actually made today.

1How do you currently credit revenue across channels?
2How many years of clean weekly data can you assemble (revenue, spend, price, promotions)?
3Do you run controlled experiments such as geo-tests or holdouts?
4Can someone actually reallocate budget based on what measurement finds?
5Do you account for the lagged, long-term effect of brand and upper-funnel media?
0/5 answered
Indicative self-assessment, not a substitute for a diagnostic
Continue the series
The Doctrine
The Stochastic Doctrine
Read →
Decision Science
Decisions Under Uncertainty
Read →
Insight
Marketing Effectiveness: Measuring What Actually Works
Read →
The next step

Bring us the decision that will not hold still.

A Strategic Diagnostic is a focused working session, not a sales call. You leave with a clear read on whether our models can resolve your friction, and the first move if they can.

Request a Strategic Diagnostic
Frequently Asked

Questions on This Topic

What is marketing mix modeling (MMM)?+

MMM is an econometric technique that uses statistical regression to isolate the marginal contribution of each marketing input to revenue, controlling for all other factors including price changes, seasonality, macroeconomic conditions, and competitive activity. Unlike attribution models that track individual customer journeys, MMM analyzes aggregate patterns over time to establish causal relationships between marketing spend and business outcomes.

Why is last-click attribution still so widely used if it is flawed?+

Three reasons: it is easy to implement (any analytics platform produces it), it produces definitive-looking numbers that feel actionable, and the channels it over-credits, paid search, retargeting, have strong commercial interests in maintaining it as the standard. The organizations that have moved to MMM typically did so after a significant budget decision went wrong that they could trace back to attribution mythology.

What data does a Marketing Mix Model require?+

A well-specified MMM requires a minimum of 2 years (ideally 3+) of weekly data including: revenue or sales volume, marketing spend by channel, pricing history, promotional activity, and relevant external variables (economic indicators, competitive spend where available, seasonal indices). Data quality is the most common limitation, models are only as good as the data that feeds them.

How long does it take to build an MMM, and how often should it be updated?+

A first-time MMM build typically takes 6 to 10 weeks, including data assembly, model specification, validation, and interpretation. Models should be recalibrated annually to account for changing media dynamics, and major structural changes (new channels, market entry or exit) should trigger interim updates. Bayesian MMM frameworks enable faster incremental updates than classical OLS approaches.

What is the difference between Bayesian MMM and classical MMM?+

Classical MMM uses ordinary least squares (OLS) regression to fit historical data. Bayesian MMM incorporates prior knowledge, known constraints about how media works, such as that TV spend cannot have negative returns, and produces probability distributions over parameters rather than point estimates. This makes Bayesian MMM more robust with limited data, better at quantifying uncertainty, and better suited to budget optimization under uncertainty. It is now the recommended approach by Google, Meta, and academic consensus.

Can small and mid-sized companies benefit from MMM, or is it only for large advertisers?+

Historically, MMM required large budgets and specialist teams, limiting it to enterprise advertisers. This has changed substantially. Open-source Bayesian frameworks (Meridian by Google, Robyn by Meta, PyMC-Marketing) have democratized the technical infrastructure. Organizations spending $5M+ annually on media now have viable access to MMM-quality insights, particularly with the support of a specialist implementation partner.

What is adstock and why does it matter?+

Adstock captures the carryover effect of advertising, the idea that a TV ad seen today continues to influence purchase behavior for days or weeks afterward. Classical attribution models cannot measure this because they only observe the touchpoint at the moment of conversion. MMM explicitly models adstock through decay functions, capturing both the immediate and lagged effects of each channel. This is particularly important for upper-funnel media like TV and brand campaigns, where the majority of the commercial impact occurs after the initial exposure.

How does incrementality testing complement MMM?+

Incrementality testing (such as geo-lift experiments) provides causal validation of MMM findings by running controlled experiments in real markets. While MMM identifies the expected contribution of each channel through statistical modeling, incrementality tests measure the actual lift by turning a channel off in one geography and comparing against a matched control. The two methods are complementary: MMM provides the strategic direction, incrementality testing validates the specific causal claims. Organizations that use both achieve the highest confidence in their budget allocation decisions.