Is incrementality just another attribution model?

No. Attribution models divide credit among touchpoints for conversions that are already recorded, so they start from the assumption that the conversion belongs to your marketing. Incrementality asks whether the conversion would have existed without the ad, which needs a comparison group rather than a rule for splitting credit.

Can I measure incrementality without running a test?

Not reliably. Without a group deliberately withheld from the ads there is nothing to compare against, and before and after comparisons absorb seasonality, promotions and every other thing that moved at the same time. Modelled approaches such as media mix modelling give a directional read at portfolio level, but they estimate the counterfactual rather than observe it.

Does a weak incrementality result mean I should switch the campaign off?

Not automatically. It means the campaign is capturing demand rather than creating it, which can still be worth paying for if it protects a checkout journey or defends a search result. The practical response is to shift part of the budget, measure again, and check whether the incremental cost per acquisition still clears your margin.

Which campaigns should be tested first?

Start with the largest spend line where you suspect the reported numbers are flattering, which is usually a small warm retargeting audience that overlaps heavily with your email list or existing customers. A test only pays for itself when the budget behind it is big enough that a different answer would change what you do next.

What is incrementality in ad measurement

What is incrementality? It is the share of your results that would not have happened if the ad had never run. Attribution asks which touchpoint deserves credit for a conversion. Incrementality asks a harder question first: would that conversion have happened anyway? On the same campaign the two answers can point in opposite directions, and only one of them tells you what your next lira of spend is really buying.

What is incrementality when you write it as a comparison

Every incrementality claim is a comparison between two worlds: the one where the ad ran and the one where it did not. You only ever observe the first. The second has to be estimated from a group of people, or a set of regions, that did not receive the ads but is otherwise comparable. Incremental conversions are what the exposed group did, minus what the comparison group implies would have happened regardless.

That is why incrementality is a causal measure and attribution is a bookkeeping one. Credit models decide how a conversion that already exists gets divided between channels; they take the conversion itself as given. If you want the reasoning behind those models, multi touch attribution covers it properly. Nothing in that machinery can tell you whether the sale needed the ad.

Attributed and caused are not the same thing

Picture a customer who has already decided to buy. She searches your brand, gets distracted, and comes back two days later through a retargeting ad. The platform records a conversion, and it is not lying: the click happened, the purchase happened. What the report cannot show you is that the purchase was coming with or without the ad. Attributed, yes. Caused, no.

The uncomfortable part is that this is not a tracking bug. Modern delivery is very good at finding people who are likely to convert. The better the system gets at identifying near certain buyers, the more attributed conversions it collects, and the harder it becomes to separate persuasion from selection. A campaign can look outstanding in the dashboard precisely because it is standing in front of people who were already walking to the checkout.

Where the gap between reported and incremental tends to open up

You cannot know your own gap without measuring it, but you can decide where to measure first. These are the places worth putting at the top of the list.

Small, high intent retargeting pools. A short window cart abandoner audience is full of people who were coming back anyway. The warmer and tighter the pool, the more of its conversions are borrowed rather than created.
Heavy overlap with owned channels. When the same people get your email, follow the account and see the ads, several channels each claim the same purchase and every one of them looks efficient.
Discount ads served to existing customers. Here the contribution can be negative in margin terms: you paid for the impression and then paid again with the coupon, on a sale you already had.
Brand search. Whether your own brand terms are additive is a separate decision with its own defensive logic, so keep it out of a general lift discussion and treat it on its own.
Upper funnel video and reach campaigns. The bias here usually runs the other way. Click led reporting undercounts them, so the reported number can be harsher than the real contribution.

Notice that the first four are cases where reported results flatter the campaign and the last one is a case where they punish it. Incrementality is not a budget cutting tool. It is a way of finding out which direction your reporting is wrong in.

How incrementality is measured, and what does not count as evidence

There is only one honest source of a baseline: a group that could have received the ads and deliberately did not. That means a randomised holdout at the user level, which the platforms package as conversion lift studies, or a geographic split where matched regions are switched on and off. Modelled approaches such as media mix modelling estimate contribution from spend and outcome history instead, which is useful at portfolio level but much coarser than an experiment. Designing any of these, including how long to run and how to read a result that comes back inconclusive, is a separate exercise from understanding what the number means.

What does not count: pausing a campaign for a week and comparing revenue with the week before. Seasonality, promotions, stock levels, competitor activity and the rest of your media all moved during that week too. A before and after comparison with no control group is a story, not a measurement, and it is the most common way teams talk themselves into believing a channel does or does not work.

The arithmetic that changes the decision

The concept stays abstract until it lands in the numbers you manage by. Take a four week retargeting campaign. The figures below are invented to keep the arithmetic legible; they are an illustration, not a benchmark of any kind.

Spend across the four weeks: 60,000 TRY.
Attributed purchases: 300. Average order value: 800 TRY.
Reported CPA: 60,000 divided by 300 = 200 TRY.
Reported revenue: 300 x 800 = 240,000 TRY, so reported ROAS is 240,000 divided by 60,000 = 4.0.

Now suppose a holdout suggests that, at the exposed group's size, 180 of those purchases would have happened without the ads. Incremental purchases are 300 - 180 = 120, which is 120 divided by 300 = 40% of what the platform reported. Everything downstream moves with that one subtraction.

Incremental CPA: 60,000 divided by 120 = 500 TRY, which is 500 divided by 200 = 2.5 times the reported CPA.
Incremental revenue: 120 x 800 = 96,000 TRY.
Incremental ROAS: 96,000 divided by 60,000 = 1.6.

Whether 1.6 is acceptable depends on your margin, not on your taste. At a 40% contribution margin, break even sits at 1 divided by 0.40 = 2.5, and you can run your own figure through the break even ROAS calculator. Reported ROAS of 4.0 clears that bar comfortably. Incremental ROAS of 1.6 does not. Same campaign, same four weeks, opposite verdicts, and the acquisition cost you should be steering by is the 500 TRY figure rather than the 200 TRY one. Carry that into any customer acquisition cost work as well, or you end up optimising a number that partly measures demand you already had.

What to do once you have an incremental number

The first instinct is to cut whatever scored low. Resist it for one cycle. A weak lift result tells you the spend is not creating demand; it does not tell you what happens when the ad disappears from a journey where it was doing part of the persuading. Move budget at the margin, a quarter of the line rather than all of it, and measure again instead of deleting the campaign and losing your ability to compare.

Second, treat overlap as a targeting problem rather than a measurement one. If retargeting keeps buying conversions that prospecting already earned, the fix lives in exclusions, audience definitions and frequency, which is the layer covered on the AI targeting page. Measurement told you where to look; it does not do the reorganising for you.

Third, hold one definition and one window across the whole account. If you report an incremental CPA for one campaign and a platform attributed CPA for everything else, the comparison table stops meaning anything, so label every figure with how it was produced. And accept that lift testing costs money, because a holdout is revenue you chose not to chase. Reserve it for lines big enough that the answer changes a decision, and rerun it when the audience, the offer or the creative rotation changes materially rather than on a fixed calendar.