What is Attribution?

Attribution is the rule an account uses to decide which advertising gets credit for a conversion.

Somebody saw a video ad on a social feed, searched the brand name a week later, clicked a search ad, left, came back through an email link and bought. One purchase, four touchpoints, and a question with no observational answer: which of them caused it.

Attribution does not answer that question. It applies a convention. Last-click gives everything to the search ad. First-click gives everything to the video. A linear model splits it evenly. A data-driven model distributes it according to patterns learned from journeys that did and did not convert.

All four are defensible and all four are conventions. The data records what happened; it does not record what would have happened otherwise, and no amount of modelling recovers information that was never observed.

Every platform reports the conversions it believes it influenced, using its own window, its own view-through rules and its own modelling.

Those claims overlap. The customer above is claimed by the social platform, by the search platform and by the email system, and each is reporting honestly under its own convention.

Summed, they routinely exceed the revenue the business actually took. This is not a reporting error to be reconciled; it is what happens when several parties each answer "did we influence this" and nobody answers "who caused it".

The blended view — total revenue divided by total advertising spend — starts from the figure in the bank and cannot be double-counted. It cannot allocate, which is why it does not replace attribution. Read together, with the blended figure as the control, the pair is hard to fool. The gap between them is itself a number worth tracking: a widening gap means the platforms are claiming more of the same conversions, which is a measurement change rather than a performance improvement.

Attribution has a consistent bias and knowing its direction is most of what makes the numbers usable.

Lower-funnel activity is over-credited. Branded search and retargeting advertise to people already close to buying, and they are present at the conversion by construction. Every model rewards presence.

Upper-funnel activity is under-credited. The video that created the demand is furthest from the purchase, frequently outside the window, and often produces no click at all.

Channels that generate clicks beat channels that generate memory. Attribution can see a click and cannot see a recollection.

An account that allocates strictly on attributed performance therefore drifts steadily toward harvesting existing demand and away from creating it. The drift is gradual, every individual decision is justified by the numbers, and the effect only becomes visible when the demand being harvested stops arriving.

Coverage has fallen and the reasons are structural rather than temporary.

Consent requirements mean some visitors are never observed. Browser tracking prevention shortens the life of identifiers and removes cross-site ones. In-app browsing contexts break continuity between an ad and a site. Mobile platforms restrict what an advertiser can join together.

Platforms responded by modelling the gaps — estimating conversions they could not observe from the patterns of the ones they could. The estimates are reasonable and they are estimates: they cannot be reconciled line by line against an order table, and they move when the modelling changes rather than when the advertising does.

The practical consequence is not that attribution is useless. It is that attributed figures are noisier and less auditable than they were, which raises the value of the two views that do not depend on observation at all: the blended ratio and the controlled test.

Pick one model and stay on it. Switching models produces a step change in every report that looks like a performance change and is a definitional one.

Read the window with the model. A seven-day window and a thirty-day window on the same model are different measurements, and long windows flatter channels that reach people early.

Compare like with like. Two campaigns under the same model and window are comparable. Two platforms' own figures are not.

Keep the blended control visible. When every channel reports an improvement and the blended figure is flat, the attribution changed rather than the business.

Escalate to a test when the stakes justify it. Whether a channel deserves to exist, whether retargeting is doing anything, whether a brand campaign pays — these are causal questions and attribution is the wrong instrument for all three.

For all the caveats, attribution does real work that nothing else can.

It is available daily, at campaign, ad and keyword level, without a design or a holdout or a waiting period. It catches breakage — a tracking fault, a page change, a campaign that stopped converting — within a day. It supports the hundreds of small allocation decisions an account makes between broadly comparable things, where the bias is roughly constant and therefore cancels.

Used for that, and checked periodically against the two views that measure rather than claim, it is a perfectly good instrument. Used as evidence about what advertising caused, it is answering a question it was never able to ask.

Most disagreements about attribution dissolve once the three available views are held apart.

Attributed figures come from the platforms. Daily, granular, biased toward the last touch, and collectively optimistic. Use them to steer between comparable things.

The blended ratio comes from the accounts. Total revenue over total spend, immune to double-counting, blind to which channel did anything. Use it as the control that catches attribution drift.

Controlled tests come from withholding advertising from a comparable group. Slow, coarse, expensive and the only one that measures cause. Use them on the questions where being wrong is costly.

An account running all three rarely gets badly lost. An account running only the first will, eventually, in a direction that takes a year to become visible.

A model and a window define a measurement jointly, and quoting one without the other is incomplete.

Last-click on a seven-day window and last-click on a thirty-day window are different instruments producing different numbers from identical delivery. So are data-driven on thirty days and last-click on thirty days.

The practical rule for any figure leaving the advertising team: name the model, name the window, and never compare two numbers that differ in either.

Attribution assigns credit for a conversion that happened; incrementality asks whether it would have happened anyway. That is a difference of question rather than of method, and it is why no attribution model answers the second one however sophisticated it becomes. What attribution can do is describe an observed sequence and distribute credit across it by a rule somebody chose — which is fast, granular, available every day, and the only instrument that can guide the hundreds of small allocation decisions an account makes between broadly comparable things. Its weakness is that the rule is a convention: the same week produces different numbers under last-click, data-driven and multi-touch models, and platform figures summed together routinely account for more revenue than the business took. Read that way — as steering rather than as evidence of cause — it does real work. The moment a figure from it is used to argue that a channel earned its budget, it is answering a question it was never able to ask.