Attribution tells you which touchpoint was nearest the sale. Incrementality tells you which spend actually caused it. This is how to build the second one without waiting for a perfect data warehouse.
Why attribution flatters everything it touches
Every attribution model answers the same question: given that a conversion happened, which touchpoints should get the credit? Notice what that question assumes — that the conversion happened because of the touchpoints in the path. It cannot tell you about the customer who would have bought anyway, and that customer is the single largest line item in most paid budgets.
This is why branded search looks miraculous. Someone who already decided to buy types your name, clicks the ad sitting above the organic result, and converts. Attribution records a win. The business paid for a click it would have received free. The same logic inflates retargeting, which mostly reaches people already deep in consideration, and it deflates upper-funnel work whose effect shows up weeks later under a different last click.
None of this means attribution is useless. It is a good navigation instrument: it shows you the shape of the journey, which creatives get engagement, where drop-off clusters. It is a bad accounting instrument, and most companies use it as one.
Start with the counterfactual, not the model
Incrementality asks a different question: what would have happened if this spend had not existed? That question has one honest answer format — a comparison between a group that received the advertising and a comparable group that did not.
You do not need a data science team to start. The cheapest useful experiment is a clean on/off test in a geography you can afford to go dark in. Pick a region with enough weekly conversions that a change would be visible, hold it dark for a period long enough to cover your purchase cycle, and compare against matched regions that kept running. The comparison is the measurement. The platform dashboard is not.
The most common mistake is running this for two weeks on a product with a six-week consideration cycle, seeing no drop, and concluding the channel does nothing. You measured a window shorter than the effect you were testing for.
The experiment ladder
Not every question deserves the same rigour, and treating them equally is how measurement programmes stall. Order your questions by how much money rides on the answer, then match each to the cheapest method that can answer it credibly.
At the bottom of the ladder sit platform-native lift tests — conversion lift on Meta, geo experiments in Google. They are cheap, fast, and marked by the same party selling you the media, so treat them as directional rather than final. In the middle sit geo holdouts you design and analyse yourself: more work, no conflict of interest, and strong enough for budget decisions. At the top sit switchback tests and matched-market designs for questions where the whole channel budget is at stake.
A useful discipline: before running anything, write down what result would make you move money, and how much. If no result would change a decision, the test is entertainment.
Turning lift into a budget decision
A lift result is a percentage. A budget decision needs money. The translation is where most experiments die quietly in a slide deck.
Take the incremental conversions the test produced, multiply by the contribution margin per conversion — not revenue, margin — and compare against the spend that produced them. That ratio is the only ROAS that means anything, and it will be substantially lower than the platform figure. Expect that. The gap between the two numbers is precisely the volume the platform was claiming and not causing.
Do this for two or three channels and you get something more valuable than a report: a rough marginal-return curve. It tells you not whether a channel works, but where the next thousand stops working — which is the actual budgeting question.
Making it a cadence, not a project
A single incrementality study has a shelf life of roughly a quarter. Creative rotates, competitors change bids, seasonality moves, the platform ships a new optimisation. A number measured in March is a historical fact by September.
The teams that get value from this run a standing calendar: one meaningful test in flight at all times, a fixed monthly slot to review results, and a written record of what each test changed. The record matters more than it sounds — without it, every new hire re-litigates decisions the team already paid to settle.
Build in a holdout you never advertise to, sized at a few percent of addressable audience. It costs a little revenue and gives you a permanent baseline, which is worth more than any single study.
Where this fails in practice
Low volume is the real constraint. Below roughly a few hundred conversions per period per cell, the confidence intervals are wide enough to be useless, and running the test anyway produces a number that feels rigorous and is not. In that situation, spend the effort on qualitative signals and on fixing the obvious instead.
The second failure is organisational. Incrementality almost always shows that some channel a team is judged on is worth less than they believed. If the person who ran the experiment is also the person whose budget it threatens, the result gets explained away. Decide who owns the measurement before you run anything.
The third is impatience. Causal measurement is slow by construction. Anyone who promises it in a fortnight is selling you attribution with a different label.