Incrementality

Why attributed results overstate what marketing caused

Every attribution model hands out 100% of the credit for every conversion, because that is what a model does. None of them has any way to represent the buyer who would have arrived regardless. That gap is not a rounding error. The channels that look best in an attribution report are frequently the ones positioned closest to people who had already decided, which means the report is partly measuring your own targeting rules rather than your marketing.

• Branded search. Someone typing your company name has already been convinced by something else. The ad takes the click and the credit, and in many accounts a large share of those visits would have arrived organically.
• Retargeting. The audience is defined as people who already visited, so retargeting wins on almost every intermediate metric by construction. That is precisely why those metrics cannot be the ones that judge it.
• Discounts and coupon placements. Some redemptions come from buyers who were at the checkout anyway. The measured revenue is real; the incremental revenue is smaller than the report shows.
• Anything running during a launch or a seasonal peak. A campaign that coincides with demand it did not create will inherit the credit for it.

The practical consequence is that a channel can be genuinely worth running and still be worth much less than its dashboard claims. The only way to tell those two apart is a comparison with a world in which it did not run.

How it is actually measured

Every real method is a comparison against something that did not receive the marketing. A holdout keeps a random slice of the audience out of the campaign and compares the two groups. A geo test runs the campaign in some regions and not others. An on-off test switches a channel off for long enough that the total has a chance to move. The rigour comes from the comparison group, never from the sophistication of the reporting layered on top of it.

For most early-stage companies the statistically clean versions are out of reach, because they need conversion volume a startup does not have — a holdout on forty signups a month cannot separate a real effect from an ordinary quiet fortnight. The crude version still works and is badly under-used: turn one thing off, leave it off long enough to cover your sales cycle, and watch the total rather than the channel. If nothing moves at the bottom of the funnel, you have learned something an attribution report is structurally incapable of telling you.

What we found measuring our own best-performing section

We can put concrete numbers on this from our own site. AgentCeres — the AI Growth Officer, agentceres.com — publishes a large library of pages, and one section of about seventy of them was, by every intermediate measure available, the best thing we had. Across 28 days it was the single largest content section by clicks: 39 clicks, roughly 35% of everything the whole site earned from search, at a per-page click rate of 0.56 that comfortably beat our average.

Over two consecutive 28-day windows that same section produced 86 organic landings and zero signups. The cause was not ranking and not the writing — it was intent. The searches bringing people in were the names of the open-source projects those pages discussed, so the visitors wanted the project itself and we are not it. A better click-through rate would have made the failure bigger, not smaller. We stopped writing new pages for that section and left the existing ones up, because keeping them costs nothing while the incremental value of adding more was measurably zero. Those are two separate decisions, and a single traffic number would have answered both of them wrongly.

Where the idea gets misused

Incrementality is a reason to distrust your best-looking numbers, not a licence to cut anything that cannot prove itself inside a fortnight. Brand, content and community work mostly surface as demand somewhere else later, and a short holdout on a slow channel reliably reports nothing — which is a statement about the length of the test rather than about the channel.

The second misuse is treating it as a permanent verdict. A channel's incrementality moves with how much you spend and who that spend reaches, so the useful output is a decision about the next increment of budget, not a grade. And when a result is written down for other people to read — in a board update, or in a case study — say plainly whether you are claiming a cause or an association. That distinction is the whole of the claim, and it is the first thing a careful reader tests.