Insights · Data
Incrementality goes beyond the last click to measure how much advertising truly adds, instead of rewarding whoever was already there. A more honest way to set budget.
For years the implicit question behind every advertising report has been: which channel gets this sale? Last-click attribution answers confidently, but it answers a bookkeeping question, not a causal one. Knowing who touched the customer last says nothing about what would have happened without that touch. It is the difference between recording an event and explaining it.
Many of the clicks we reward come from people who would have bought anyway. A search on the brand name, a retargeting ad served to someone with a full cart, an email to someone already decided: all of it is counted as generated revenue, when it is often merely intercepted revenue. Budget inflates exactly where the conversion was already written.
Incrementality flips the perspective. It does not ask who brought the sale, but how many additional sales exist because we invested. It is a counterfactual question: what would have happened without the ad? You cannot observe it directly, which is why it must be built through a measurement design, not read off a dashboard.
Lift is the difference between the results of people exposed to advertising and the results of a comparable group that was not. It is the net added value, stripped of what would have happened on its own.
Without a point of comparison there is no incrementality, only correlation. The control is a set of people or areas deliberately left unexposed, acting as a mirror of what the campaign would have achieved by standing still.
A share of sales arrives regardless: loyal customers, spontaneous demand, word of mouth. Confusing this baseline with the campaign's merit is the most expensive error, because it means paying for results that do not depend on the spend.
Each extra euro returns less than the one before. The first impressions capture live demand, the last ones chase people already convinced or indifferent. Incrementality measured at different spend levels shows where the curve flattens.
You start from a testable claim: this campaign generates sales that would not exist without it. You pick a single outcome metric and decide in advance what will count as proof, before seeing any data.
You divide the audience, or the territory, into two comparable groups: one sees the campaign, one does not. Assignment should be random or, where that is not possible, based on areas as similar as possible in historical behaviour.
During the test you avoid changing prices, offers and other channels on the groups, otherwise the result measures noise instead of effect. The discipline of touching nothing is worth more than any statistical refinement.
When the period ends, you compare the results of the two groups. The gap, not the total of the exposed, is the lift. You also assess how solid that gap is, or whether it is down to chance.
The number alone is not enough: it becomes a cost per incremental conversion and is compared with other lines of spend. Only at this point do you decide whether to increase, cut or move the budget.
A clean cycle takes weeks, not days: the difference needs time to emerge above natural variability.
To have a control you must not speak to part of your potential customers. It is a real, immediate cost, traded for knowledge that lasts. Anyone unwilling to make that trade is not measuring incrementality: they are only decorating attribution.
A serious test cannot be read the next morning. Organisations used to optimising every day find it frustrating to wait for a signal to settle. But a decision taken on daily noise costs more than a few weeks of patience.
Not everything can be tested with the same cleanliness. Some channels lend themselves to sharp experiments, others must be estimated with more indirect methods and wider margins. Recognising where the measure is solid and where it is only indicative is part of the method's honesty.
Often the real lift is lower than attribution reports promised. It is uncomfortable but useful news: it shows where you were paying for sales already secured. A method that only returns good news is measuring nothing.
We start from the question what would have happened anyway, and only then look at the numbers. A dashboard not grounded in a measurement design is a well-formatted opinion. We prefer a few defensible metrics to many suggestive ones.
Incrementality is not there to prove right whoever already set the budget, but to put it into question. We use it to move spend from where it intercepts existing demand to where it truly creates it, even when the result is awkward for internal habits.
Better one clean experiment on a choice that matters than ten rough measures on marginal details. We focus the effort where the budget stakes are high and where the result will genuinely change an allocation.
Every estimate has a margin, and we say so. Treating a noisy number as a certainty leads to confident decisions on fragile ground. Communicating honestly what we know, and what we do not, is what makes the measure usable.
Is last-click attribution useless?
No, but it answers a different question. It is fine for understanding how customers move along the path, not for deciding how much a channel adds. Using it to allocate budget is like measuring temperature with a ruler: the tool works, just not for that task.
Do you need a huge budget to test incrementality?
You need discipline more than size. A simple but rigorous design, with comparable groups and rules decided in advance, says more than a complex apparatus built on top of confused data. The cleanliness of the method matters more than scale.
How often should tests be repeated?
Incrementality is not a fixed truth: it changes with the season, the competition, the saturation of the audience. It should be re-checked when something substantial shifts in the market or the strategy, not at every weekly meeting.
What if the test says a channel adds nothing?
That is exactly the kind of answer a test exists for. It means that spend was collecting sales already bound to arrive. The right reaction is not to defend the channel, but to cut or reallocate and check again.