Attribution vs Incrementality: What a Metric Can and Cannot Prove

Two questions sound almost identical and are completely different. Which ad got the credit for this conversion? And: Did the ad cause the conversion at all? The first is attribution. The second is incrementality. Most of advertising's measurement confusion — and a good deal of its wasted spend — comes from answering the first while believing you've answered the second.

Attribution assigns credit among touchpoints. Incrementality measures causation. A metric can be excellent at one and silent on the other, and the gap between them is where advertisers quietly fund conversions that would have happened anyway. Knowing which question a number actually answers is the beginning of honest measurement.

What attribution actually does

Attribution is a credit-assignment rule. Given a conversion and the touchpoints that preceded it, attribution decides who gets the points. Last-touch gives everything to the final click. First-touch credits the introduction. Multi-touch spreads credit across the path by some formula.

Every one of these is a convention, not a measurement. Attribution does not ask whether the touchpoints mattered — it assumes they did and divides the spoils. That makes it useful for operational questions ("which channels are involved in conversions?") and treacherous for causal ones ("which channels are causing conversions?"). The most popular convention, last-touch, is also the most misleading: it lavishes credit on whatever ad happened to be last, including retargeting that reached people already about to buy.

What incrementality actually does

Incrementality asks the only question that matters for budget: what would have happened anyway? It compares a group exposed to advertising against a comparable group that was not (a holdout), and measures the difference in outcomes. That difference — the lift — is the causal contribution of the advertising. Everything else is conversions you would have gotten for free.

This is a fundamentally harder question, because it requires a counterfactual: a credible picture of the world without the ad. You cannot observe it directly; you have to construct it, through experiments (randomized holdouts), or models that approximate it (marketing-mix modeling, geo experiments). But it is the right question, because it separates advertising that changed behavior from advertising that merely witnessed it.

Why the distinction is expensive to ignore

Consider retargeting. By last-touch attribution it looks spectacular: it reaches people near the moment of purchase, so it is "present" at conversion and collects the credit. By incrementality it often looks far weaker, because many of those people would have converted without being retargeted. An advertiser optimizing to attribution will increase retargeting spend; an advertiser optimizing to incrementality might cut it. Same data, opposite decisions — and only one of them is grounded in causation.

Attribution can make an ad that changed nothing look like your best performer. Incrementality is the discipline of refusing to pay for outcomes you would have gotten for free.

When each is valid

This is not an argument that attribution is worthless. Each tool has a valid domain:

• Attribution is valid for operational visibility, pacing and allocation heuristics, and understanding the shape of conversion paths — as long as nobody mistakes credit for causation.
• Incrementality is valid for budget decisions, channel justification, and any claim that advertising worked — because only it speaks to causation.
• MMM, MTA, and experiments sit on a spectrum: experiments (holdouts) give the cleanest causal read but are operationally demanding; MMM estimates causation from aggregate patterns; MTA refines credit assignment but remains, at heart, attribution.

The mature posture is to use attribution for steering and incrementality for judging — and never to let an attribution number masquerade as proof of effect.

Measurement needs the right raw material

Here is where measurement meets infrastructure. Incrementality is only possible if you can construct comparable exposed and unexposed groups and measure their outcomes — which requires event-level data: who was exposed, who converted, when, under what conditions. Vendor aggregates cannot support it, because they have already collapsed the events into totals. This is the quiet link between honest measurement and data ownership: you cannot run a credible incrementality analysis on a dashboard. You need the underlying conversion and audience events, joinable and yours. The evidentiary discipline — proof over assertion — applies to measurement as much as to reporting.

Common misconceptions

• "Attribution measures effectiveness." It assigns credit; effectiveness is a causal question only incrementality answers.
• "Last-touch shows what's working." It shows what was last, which flatters lower-funnel tactics that reach the already-convinced.
• "Multi-touch attribution solves it." MTA is more sophisticated credit-splitting; it is still attribution, not causation.
• "Incrementality is too hard to bother with." Holdouts and geo tests are practical; the cost of skipping them is funding free conversions forever.
• "A high ROAS proves the ads worked." ROAS built on attributed conversions can be high while incremental ROAS is near zero.

What good operation looks like

• Use attribution to steer day-to-day and incrementality to judge whether spend works.
• Treat last-touch with suspicion, especially for retargeting and lower-funnel tactics.
• Run holdouts / geo experiments where the budget justifies a causal read.
• Build measurement on event-level data you own, so causal analysis is even possible.
• Be explicit, in every report, about which question a number answers.

Open questions

• What is the minimum holdout discipline that makes incrementality credible without crippling delivery?
• How should attribution and incrementality be reconciled when they disagree (as they often will)?
• Can incremental measurement be made routine and cheap enough to run continuously, not just in occasional studies?

The comfortable number is the attributed one: it is always available, always flattering, and almost always beside the point. The honest number is the incremental one: harder to get, sometimes disappointing, and the only one that tells you whether your advertising changed anything. Attribution answers "who gets the credit." Incrementality answers "did it matter." Spend your budget on the second question.