Marketing Mix Modeling for evaluating the brand's contribution to sales

Why attribution and MMM argue with each other

Marketers often encounter this issue precisely at the moment of conflicting numbers: Facebook pixel shows ROAS 8x, while MMM for the same campaign shows 3x. It's not that one model is wrong. They answer different questions.

Platform attribution systematically understates the role of media that create demand in advance — according to different estimates, the ROI of performance channels is overestimated by 1,5–2 times, because the system assigns conversions to people who have already made a decision thanks to TV or outdoor advertising. MMM sees the entire market and can separate these effects. The correct practice is not to choose one tool, but to use MMM for strategy and attribution for tactical optimization within the digital budget.

What MMM actually calculates and where the contribution figures come from

The model is a regression that explains fluctuations in sales through advertising pressure, price, seasonality, distribution, and competitor actions. The media's contribution depends heavily on the category: in the financial sector, advertising may constitute a significant share of sales contribution, while in telecom this role is traditionally higher. The wide variation is due to the fact that categories differ in the strength of base demand, purchase frequency, and the brand's role in the decision-making process.

A separate line that many models miss — competitors. Their activity can contribute a negative impact of up to 30% to brand sales, and this effect is growing year by year. If the model does not include media suppression of competitors, it shifts their influence onto random noise or, worse, onto the brand's own channels, distorting conclusions about what works.

How MMM captures the long-term brand effect — and why it usually doesn't

Here is where it gets interesting — and what few review articles write about. Advertising does not only work at the moment of exposure: its effect accumulates and fades gradually — this is called carryover effect. The problem is that typical modeling windows are often insufficient to capture the full long-term advertising effect. The long-term brand effect often significantly exceeds the short-term effect, but the short window simply cannot see it — the model captures only the activation spike, not the foundation that advertising builds over months.

The Dom.ru case illustrates this clearly: a TV campaign launched in 2022 continued to drive growth in incoming sales long after its airing — the effect was spread far beyond the standard attribution window.

Why 'doubling carryover' doesn't solve the problem

A common but mistaken approach is to artificially double the carryover window, hoping to capture the long-term effect. This doesn't work and creates a new problem: the extended carryover begins to correlate with seasonal and trend components of sales that are modeled as separate variables. As a result, the model attributes growth to advertising that is actually explained, for example, by a market organic trend — false correlation emerges, and final brand contribution figures become an artifact of setup, not a reflection of reality. The correct approach is not a crude doubling, but separate modeling of the short-term activation curve and long-term brand curve with different decay parameters, or using structural time series models that explicitly isolate the baseline brand level as a separate component.

How to embed brand metrics into the model

Most MMM models stop at media metrics and sales, completely ignoring what happens in the consumer's head between the advertising contact and the purchase. To see the full contribution of branding, tracking metrics — awareness, preference, consideration — must be added to the model as intermediate variables, not as a side report.

• Awareness is included as a mediator between advertising pressure and sales — part of the media effect is realized precisely through the increase in brand awareness, not directly;
• Preference and consideration help explain why the same media pressure produces different responses in sales at different times — the base of consumers ready to choose the brand changes;
• The two-stage model (media → brand metrics → sales) allows separating the effect of 'brand building' from the effect of 'demand activation' mathematically, not through expert judgment;
• Tracking data is needed with the same weekly or at least monthly regularity as media data — jagged waves of surveys break the model.

How to prepare for MMM and not ruin the results

The standard is a minimum of 104 weeks, that is, two full years of weekly data. Less than 52 weeks only give direction ('the channel works / does not work'), but not the accuracy sufficient for budget reallocation — at such a volume, the model will not distinguish seasonality from the effect of advertising.

1. Collect weekly data for at least 2 years: sales, media investments by channels, price, promotions, distribution, weather if necessary;
2. Include competitor activity — without it, the model will attribute their influence to noise or to your own channels;
3. Add brand tracking metrics as a separate layer, not as reference information;
4. Do not use a single ad stock for all formats — TV, digital, and OOH have different speeds of effect accumulation and decay, and this needs to be calibrated separately for each creative and flight;
5. Check the model on historical shifts in the media plan — if the model does not explain known sales failures or spikes, it cannot be trusted for budget decisions.

Common mistakes that ruin even good data

The model can still produce an incorrect result even with a perfect dataset — the problem is not with the data, but with the specification. The most dangerous mistake is the already mentioned mechanical doubling of the adstock for 'capture' of the long-term effect: it causes multicollinearity with the trend. The second mistake is ignoring competitive pressure, which leads to overestimating or underestimating the performance of own channels. The third one is using a single decay curve for all advertising formats, whereas TV spots and banners accumulate effects differently. The fourth, and the most frustrating one, is building the model on data shorter than a year, when the client requires precise figures for budget rebalancing, but the model physically cannot provide them. And the last, less obvious one: complete disregard for brand metrics, which makes MMM see only half the picture and constantly undervalues the contribution of brand building in favor of short-term activation.