Incrementality over attribution

Attribution is a story about credit. It takes the sales you already made and decides which channels deserve the applause. That is a reasonable thing to want, and it is the wrong question to run a budget on. The budget question is not who gets the credit for the sales that happened. It is which spend caused sales that would not have happened otherwise. Those are different questions, and confusing them is how efficient-looking accounts quietly waste large amounts of money.

The shift from attribution thinking to incrementality thinking is the single most valuable change most advertisers can make. It is also uncomfortable, because it requires admitting that some of the spend currently sitting in the win column may not be earning its place there.

Credit is not cause

Picture a customer who was already going to buy from you this week. They have decided. On the way to checkout they click a retargeting ad, because retargeting ads are placed precisely in front of people who are already close. The attribution system records a conversion for retargeting. Credit assigned, dashboard happy.

But nothing was caused. That sale was coming with or without the ad. The retargeting did not create demand, it intercepted it and took the receipt. Multiply that across thousands of conversions and you have a channel that looks like a star performer while contributing far less than its numbers suggest. Attribution cannot tell the difference between causing a sale and standing next to one. Incrementality is built specifically to tell them apart.

Average return hides the number that matters

Most people judge a channel by its average return: total revenue credited, divided by total spend. The problem is that the average flattens a curve. The first dollars into a channel usually work hard, reaching people who are persuadable. As you spend more, you reach further into audiences that were already going to convert or were never going to, and each additional dollar does less. This is diminishing returns, and it is the central fact of media buying.

What you actually need to know is the marginal return: what the next dollar will do, not what the average dollar did. A channel can show a strong average ROAS while its marginal ROAS has already fallen below break-even, which means the next dollar loses money even though the blended number still looks fine. Scaling a channel on its average return is one of the most common and expensive mistakes in paid media, and it is invisible to attribution.

How to actually test for incrementality

You measure cause the way every other field that takes evidence seriously does it: with a control group. You deliberately withhold the marketing from a comparable slice of your audience or market, then compare. The lift between the group that got the ads and the group that did not is the incremental effect. Everything else is correlation.

Holdout audiences

Withhold a channel or campaign from a random share of your addressable audience and serve it to the rest. Because the split is random, the two groups are alike in every way except exposure, so any difference in outcomes is the effect of the marketing. This is the cleanest design where the platform supports it.

Geo experiments

When you cannot split at the person level, split by geography. Turn a channel off in a set of matched regions, keep it on in comparable ones, and measure the difference in sales between them. Geo tests are powerful precisely because they capture the full effect including the offline and untrackable touches that person-level tracking misses entirely.

Enough scale to read the result

An incrementality test only works if the effect is large enough to see above normal noise. That requires enough conversions, a long enough window, and a big enough difference to be confident the result is real rather than chance. A test run too small or too short produces a number you cannot trust, which is worse than no number, because it carries false authority. Designing the test so it can detect the effect is half the work.

Reading the result and acting on it

The output of a good test is incremental return on ad spend: the revenue that would not have existed without the spend, against the cost of that spend. It is almost always lower than the platform-reported figure, sometimes dramatically, and that is the point. Now you know the real number.

From there the moves are concrete. Where marginal return is still strong, there is room to scale and you can do it with confidence. Where incremental return has collapsed below the line, you can cut or cap that spend and lose little or no real revenue, freeing budget for channels and audiences that are actually creating demand. Decisions stop being arguments about whose attribution model is right and become reallocations backed by evidence.

Guardrails so the discipline survives contact with reality

Incrementality is powerful and easy to misuse. A few principles keep it honest. Test the channels carrying the most budget first, because that is where a wrong belief costs the most. Re-test periodically rather than once, because audiences saturate and effects decay over time. Resist the urge to over-interpret a single result, and triangulate it against your mix modeling and your attribution view rather than treating any one lens as the final word. And keep the tests simple enough that the people making budget decisions actually understand and trust them, because a result nobody believes changes nothing.

Attribution will always have a place for understanding paths and sequence. But when the question is where to put the next dollar, stop asking who deserves credit and start measuring what your spend causes. The number will be humbler than the dashboard's. It will also be true, and you can build a budget on it.