Attribution is not credit
Attribution divides the credit for a sale. It does not prove what caused it. On the difference between the two, and the money lost between them.
What is the difference between attribution and credit? Attribution is a rule for dividing credit for a sale among the touchpoints that preceded it. Credit is a claim that something caused the sale. The first is bookkeeping. The second is evidence. Confusing the two is how marketing budgets get spent on the wrong thing, month after month, with a clean report to prove it.
Consider the relay. Four runners carry one baton around the track, and the crowd remembers the one who crossed the line. Now give the photographer the job of deciding who won the race, and you have the modern attribution report. The picture is accurate. The conclusion drawn from it is not.
What does attribution actually measure?
An attribution model is a set of rules for dividing credit among touchpoints. Last click gives everything to the final ad before the sale. First click gives everything to the first. Linear divides it evenly. Time decay leans towards the end. Position-based rewards the beginning and the end. Data-driven lets an algorithm decide.
Same journey, same money, six different winners. Note this carefully, because it is the whole point. When the raw data is fixed and only the rule changes, the model is not measuring cause. It is applying a convention. The models are not competing theories of what won the customer. They are competing conventions for dividing the prize.
That does not make them worthless. A convention can be useful. A ledger that ignores the difference between conventions and causes, however, will eventually fund the wrong department.
Why does last click keep winning?
Because it is simple, it arrives first, and it flatters the channel nearest the sale. Those are three powerful advantages, and not one of them is truth.
The bias is measurable. Analyses of common attribution breakdowns find the same distortions recurring. Email frequently carries around 35 per cent of the credit in reports, when what it is mostly measuring is reactivation, not generation. The email arrived to a customer who was already convinced, and it gets the medal for the persuasion. Direct traffic, which is to say untracked traffic, often carries around 25 per cent, and much of it is dark social and broken tracking wearing a respectable name.
A coincidence dressed as a cause is still a coincidence. The last click was usually the last in a queue, and queues are not proof of who did the work.
What happens when attribution is checked against experiments?
Here is where the story gets interesting, because someone finally ran the check.
Incrementality testing compares two groups. One sees the marketing, one does not, and the difference between the outcomes is the lift that the marketing actually caused. It answers a different question to attribution, and the question it answers is the one that moves money. Not which touchpoint appeared before the sale, but which touchpoint changed the odds.
Measured’s work on Meta campaigns is the clearest public example. For remarketing, the incremental lift was roughly 39 per cent, with an 80 per cent confidence interval of 33.3 to 44.3 per cent. Put plainly, about 6 in 10 conversions that attribution credited to remarketing would have happened anyway. The tactic was over-credited by roughly 2.6 times. For prospecting, the lift was roughly 93 per cent, meaning attribution was close to accurate there.
Read that again and notice the shape of it. The error is not uniform. Attribution flatters the channel closest to the sale and undervalues the ones that create demand. The budget then follows the flattery, and the business wonders why the numbers improve while the revenue does not.
What is incrementality, and what is it not?
Incrementality testing is a controlled experiment. A holdout group sees no marketing, the rest sees the campaign, and the gap between them is the lift the campaign caused. eMarketer’s April 2026 definition is as good as any. Incrementality measures whether a campaign caused outcomes that would not have occurred without the exposure.
It is not a better attribution model. It is a different instrument. Attribution is a logbook, quick and granular and always on. Incrementality is the referee, slow and honest and occasionally unwelcome. The smartest operators use both, at different altitudes. The logbook for direction, the referee for the budget.
How do AI answers make the arithmetic worse?
Everything above describes a world where the journey can be tracked. That world is shrinking.
A customer asks an AI assistant for a recommendation, reads the answer, types the name of a business into a browser, and arrives as direct traffic. The recommendation that won the sale is invisible in every report that follows. The channel that gets the credit is whichever link happens to be tracked. Journeys fragment further every quarter, and the distance between reported numbers and true causes grows with them.
This is the honest position, and it cuts at agencies as much as anyone. Attribution was an approximation before AI answers. It is now an approximation of a journey nobody can see. Which is precisely why the discipline of holdout tests and incremental lift matters more than ever, not less.
What should a business actually do about it?
Three things, none of them expensive.
Ask any agency or platform for incremental evidence before renewing a tactic, not after. Remarketing reports in particular deserve the question, given the size of the over-crediting found when anyone checks.
Treat platform-reported numbers as directional signals rather than accounts of cause. They are useful. They are not evidence.
Watch the share of traffic reported as direct. When it climbs, that is not a surge of brand love. It is usually a tracking problem or a fragmented journey, and every credit decision built on top of it is quietly wrong.
The difference that costs money
We built the measurement side of this business for exactly this reason. Reporting that separates activity from outcomes, and touchpoint logs from claims about cause. It is the least glamorous work we do, and it is the work everything else stands on. If this is the conversation worth having first, our analytics and GA4 and GTM pages set out how we think about it. The argument between the two columns was made at greater length in What we refuse to call a result.
Credit is what the report says. Cause is what the bank balance says. The distance between the two is where budgets go to die, and that distance has never been wider or more politely ignored.
The choice, as ever, is yours.