Summer changes
When summer arrives, many brands observe the same pattern: some metrics fall, connection times…
Never have we had so much data available on user behaviour. Every campaign can fill a dashboard with clicks, impressions, conversions, ROAS, CPA, attributed sales, reached audiences, frequency, repeat purchases or average order value. At first glance, marketing seems to have entered an age of near-absolute precision: every action leaves a signal, every interaction can be measured and every outcome can be assigned to a channel, campaign or touchpoint.
But that sense of precision can be misleading.
The great challenge of measurement today is not simply gaining access to more data, but understanding what it really means. A campaign may show a good ROAS and yet have generated no incremental growth. A channel may receive many attributed conversions while, at the same time, capturing demand that already existed. A dashboard may present highly orderly results without answering the most important question: what would have happened if that campaign had not been activated?
That is where the problem begins. Many metrics explain what happened after an advertising exposure, but do not always demonstrate that this exposure was the cause of the result. The difference may seem technical, but it is strategic. Assigning a sale to a campaign is not the same as proving that the campaign created an additional sale. Measuring activity is not the same as measuring impact.
That is why one of the most common mistakes in marketing is confusing apparent precision with strategic truth. The fact that a figure is accurate does not mean it explains reality well. The fact that a report is sophisticated does not mean its conclusions are correct. And the fact that a platform attributes a conversion to a campaign does not necessarily mean that conversion would not have happened anyway.
In an increasingly fragmented ecosystem, with greater privacy restrictions, more channels connected to commerce and more pressure to demonstrate business results, measuring well requires accepting an uncomfortable idea: no single model solves everything on its own. Attribution, marketing mix modelling (MMM) and incrementality testing do not necessarily compete with one another. Each answers different questions, with its own advantages and limitations.
The issue, then, is not to choose a single model as though it were the definitive answer. True maturity lies in knowing how to combine them: using attribution to read tactical signals, MMM to understand aggregate patterns, and incrementality testing to validate whether a campaign has generated real value. In an imperfect environment, measuring better does not mean having absolute certainty, but building more robust decisions from different forms of evidence.
Why no single model solves everything on its own
The problem begins when a tool designed for a specific question is expected to answer every business question. In marketing measurement, this happens often: attribution is asked to prove real impact, MMM to explain every tactical decision, and incrementality tests to resolve every investment-related doubt. But each approach has a different role.
Attribution can be very useful for optimising campaigns in the short term. It helps identify which adverts, channels or audiences play a role in the journey to conversion. However, it also has a clear limitation: it tends to favour touchpoints closest to purchase. As a result, it may give too much credit to campaigns that capture existing intent and push into the background actions that have previously helped generate that demand.
Marketing mix modelling offers another perspective. Rather than tracking users individually, it works with aggregate data and makes it possible to observe how investment, channels, seasonality, promotions, price and the competitive context interact. Its value lies in helping to understand the overall impact of marketing, but it is not always useful for making immediate tactical decisions about a particular creative, audience or keyword.
Incrementality testing, for its part, attempts to get closer to causality. Its main question is not “which campaign do we attribute this conversion to?”, but “what additional result has this campaign generated?”. But neither is it a magic solution. It requires sound experimental design, a sufficient sample, time, discipline and a careful reading of the results.
That is why the question is not which model is “the right one” and which ones can be dispensed with. The key is to understand which question each one answers. Attribution helps optimise, MMM helps provide context and incrementality testing helps validate. When combined well, they do not remove all uncertainty, but they do allow for more robust decisions than interpreting a single model as though it were the whole truth.
What attribution contributes
Criticising the limitations of attribution does not mean dismissing it. Attribution remains a useful tool, especially when a team needs to read quick signals and make operational decisions. It makes it possible to compare campaigns, analyse journeys, identify which channels are involved before a conversion and optimise specific elements such as creatives, audiences, formats, messages or landing pages.
Its main value lies in speed and granularity. Compared with more aggregate models or experiments that require more time, attribution offers an immediate reading of what is happening within a campaign. It can help identify which adverts generate more interaction, which audiences appear to respond better, or which touchpoints appear most frequently on the path to purchase.
The problem arises when this tactical reading is interpreted as absolute proof of impact. An attributed conversion does not always equal a generated conversion. In many cases, attribution indicates which touchpoint a result is assigned to, but does not necessarily show that this touchpoint was the main cause of the sale.
That is why attribution works best when it is used to optimise, rather than to close the discussion entirely on a campaign’s real value. It is a valuable tool for managing the short term, but it needs to coexist with other approaches when the question is no longer simply “which channel received the credit?”, but “which part of growth would have been lost had we not invested?”.
What MMM contributes
Marketing mix modelling provides a different perspective because it does not begin with the individual journey of each user, but with aggregate data. Its aim is to analyse how investment, sales, channels, seasonality, promotions, price, distribution and even the competitive context are related. Rather than asking which touchpoint received a specific conversion, it seeks to understand how each variable contributes to overall business performance.
This makes it particularly relevant in an environment where user-by-user tracking is becoming increasingly limited. When fewer individual signals are available, aggregate models make it possible to observe patterns that do not depend solely on cookies, identifiers or platform data. That is why MMM is gaining importance again in the current measurement debate, alongside attribution and incrementality testing, as approaches that should coexist rather than exclude one another. eMarketer, for example, continues to place incrementality and advanced measurement among the central themes in marketing and commerce media in 2026.
Its value lies in providing a broader view: it can help identify long-term effects, points of saturation, cumulative impacts, dependence on seasonality or the relative contribution of different channels. It also makes it possible to elevate the conversation beyond daily optimisation and connect marketing investment with planning, budgeting and growth decisions.
But MMM should not be interpreted as an automatic truth either. Its usefulness depends on data quality, the variables included, the model’s assumptions and a cautious reading of the results. An aggregate model can provide context and direction, but it does not always explain precisely which creative, audience or format performed best in a specific campaign.
That is why MMM does not replace attribution or incrementality testing. It adds another layer of interpretation. It helps us understand the business with greater perspective, but it needs to be complemented by more tactical tools and experiments that make it possible to validate specific hypotheses. Its greatest value emerges when it is not used as a single answer, but as part of a broader measurement system.
What incrementality testing contributes
Incrementality testing brings a question that neither attribution nor MMM can always answer on its own: which part of the result would not have happened if the campaign had not been activated. Its value lies precisely there: in attempting to separate the results a campaign can claim from the results it has actually caused.
This approach can be applied in different ways: control groups, holdout groups, geo-testing, A/B tests, audience experiments or comparisons between similar markets. In every case, the logic is similar: compare what happens when a campaign is active with what happens in a scenario that is as similar as possible in which that campaign does not intervene.
That is why incrementality testing is particularly useful for validating hypotheses. It can help establish whether a channel is generating additional growth, whether a promotion is attracting new buyers, or whether a campaign with good attributed performance is creating real value. It also serves to challenge assumptions and calibrate other models, because it introduces an experimental layer that requires dashboard conclusions to be tested against more robust evidence.
But it should not be idealised either. Not everything can always be tested, nor are all tests equally reliable. A sufficient sample, careful design, time and discipline are needed to interpret results. A poorly designed experiment can generate conclusions just as misleading as poorly interpreted attribution.
That is why incrementality testing should not be seen as a magic solution, but as a validation tool. Its greatest contribution is not to eliminate all uncertainty, but to bring measurement closer to a more demanding question: not only which results we see, but which results we can reasonably attribute to a specific marketing decision.
How the three approaches coexist in a mature measurement strategy
A mature measurement strategy does not treat attribution, MMM and incrementality testing as rival approaches. It understands them as complementary layers of interpretation. Each contributes part of the answer, but none should become the only source of truth.
Attribution helps optimise day to day. It makes it possible to read quick signals, compare campaigns, adjust budgets, review creatives and identify which touchpoints are involved closest to conversion. It is useful for operational management, provided it is not mistaken for definitive proof of impact.
MMM provides a broader view. It helps us understand the overall contribution of channels, the weight of investment, the influence of seasonality and the relationship between marketing and business results. Its role is not to decide which advert to change tomorrow, but to support better planning and interpret performance within a wider context.
Incrementality testing, finally, makes it possible to validate hypotheses and correct biases. It helps establish whether what appears to work in attribution or in the aggregate model is actually generating additional value. In this sense, it brings measurement closer to a more causal question: what would have happened without that campaign, channel or investment?
A simple way to summarise it would be this: attribution guides, MMM provides context and testing validates.
For example, a commerce media campaign may show high ROAS within the platform. Attribution will indicate that the campaign has been involved in many sales. MMM can help reveal whether that channel is contributing to overall business growth or simply concentrating sales that would have occurred anyway through other routes. And an incrementality test can establish whether additional sales have actually been generated or whether the campaign has captured demand that already existed.
When these approaches are combined well, measurement ceases to be a competition between models and becomes a learning system. The aim is not to find one single figure that closes the discussion, but to bring together different pieces of evidence in order to make better decisions.
Common mistakes in data interpretation
One of the greatest risks in advanced measurement is not a lack of data, but reading it badly. When an organisation has dashboards, models, reports and real-time metrics at its disposal, it can easily fall into the temptation of treating every figure as a definitive answer. But data does not speak for itself: it always needs context, judgement and a well-formulated question.
The first common mistake is to confuse attributed sales with incremental sales. The fact that a platform assigns a sale to a campaign does not necessarily mean that sale would not have occurred without the campaign. The advert may have influenced the decision, but it may also simply have captured purchase intent that already existed.
Another common mistake is believing that the channel with the highest ROAS always deserves more budget. A high ROAS may indicate efficiency, but it may also conceal investment that is too conservative, audiences very close to conversion or a heavy dependence on customers who were already going to buy. Without a broader reading, optimising solely for ROAS can reduce real growth potential.
It is also risky to make strategic decisions based solely on platform data. Each platform tends to measure from within its own environment, using its own attribution windows, methodologies and criteria. If this data is interpreted in isolation, it is easy to overestimate the role of certain channels and undervalue others that contribute in less visible ways.
Another problem is added to this: ignoring external factors. Seasonality, promotions, price changes, distribution, product availability or competitors’ activity can alter campaign results. If they are not taken into account, marketing may claim credit that actually comes from other business variables.
It is also worth avoiding comparisons between models as though they all answered the same question. Attribution, MMM and incrementality testing do not measure exactly the same thing. That is why their results will not always align, and that difference does not necessarily mean that one is “right” and another is “wrong”. It may simply indicate that they are observing the problem from different angles.
Finally, there is one particularly relevant cultural mistake: using data to justify decisions that have already been made, rather than using it to learn. Measurement should help challenge hypotheses, correct biases and improve future decisions. When data becomes merely an internal validation tool, it loses much of its strategic value.
Measuring well requires accepting that models work with uncertainty. They will not always provide an exact answer, but they can help reduce the margin for error. The aim is not to find a perfect figure, but to interpret reality better so that we can decide more responsibly.
Conclusion
Maturity in measurement is not about having more reports, more metrics or more dashboards. Above all, it is about asking better questions. It is not enough to know how many conversions have been attributed to a campaign, what ROAS a platform shows or which channel appears at the end of the journey. The important question is what we can learn from this data and which decisions it helps us make with better judgement.
In an imperfect ecosystem, the aim is not to eliminate uncertainty altogether. That would be unrealistic. There will always be variables that cannot be fully controlled: changes in demand, seasonality, price, competition, product availability, consumer behaviour or the limitations of the models themselves. The key is to reduce that uncertainty enough to make better decisions.
That is why marketing today needs to combine data, context, experimentation and human judgement. Models can identify patterns, organise information and provide valuable signals, but they do not replace strategic interpretation. Good measurement does not merely answer what has happened; it helps us understand why it may have happened, which part of the result can genuinely be attributed to marketing and what needs to be checked before making the next decision.
In this sense, measurement should no longer be seen solely as a control tool. Its value does not lie only in justifying budgets, evaluating campaigns or demonstrating results after the fact. It can also become a learning tool: it allows us to challenge hypotheses, correct biases, discover opportunities and build an organisation better prepared to make decisions in changing environments.
In an imperfect ecosystem, the advantage does not lie in finding the perfect model, but in building a more intelligent way of learning. Measuring better does not mean seeing everything; it means knowing what each piece of data can say, what it cannot say and which decisions it makes possible with greater responsibility.
Is your data helping you learn better, or is it merely confirming what you already wanted to believe?
Sources:
Comments