If you spend any time reading about marketing attribution, you’ll find a lot of posts that look like decision trees. First-touch if you’re optimizing for awareness. Last-touch if you’re optimizing for conversion. Multi-touch if you’ve graduated to a more sophisticated view of the funnel. Linear, U-shaped, W-shaped, time-decay, data-driven, position-based. The posts treat the choice like a software selection: figure out your requirements, find the model that matches, implement.
That framing is wrong. Or at least, it skips the question that actually matters.
The honest starting point is that every attribution model is wrong. Not in a pedantic philosophical sense. Wrong in a specific, mechanical sense. First-touch ignores everything that happened after the first interaction, which is most of the journey. Last-touch ignores everything that happened before the final interaction, which is also most of the journey. Multi-touch models distribute credit across touches using rules that are themselves assumptions, not measurements. None of these models is measuring causation. They’re all measuring correlation with rules attached.
This isn’t a reason to give up on attribution. It’s a reason to stop asking “which model is right” and start asking “which kind of wrong is most useful to me.”
That question has an answer, and the answer depends entirely on what decision you’re trying to make.
If you’re trying to decide where to invest at the top of the funnel, first-touch is the right kind of wrong. It overweights the channel that introduced people to your brand. That’s exactly the bias you want when you’re allocating budget for awareness. You don’t care that the channel didn’t close the deal. You care that it created the opportunity for everything downstream to happen. The model’s blindness to mid-funnel touches is fine, because that’s not the decision you’re using it for.
If you’re trying to decide which late-stage programs are converting, last-touch is the right kind of wrong. It overweights the touch closest to the conversion event. That’s the bias you want when you’re evaluating bottom-funnel content, sales-assist programs, or retargeting. The model’s blindness to early touches doesn’t matter, because the decision is about what closes deals, not what starts them.
If you’re trying to decide how to rebalance investment across the full funnel, multi-touch is the right kind of wrong, with caveats. Multi-touch models distribute credit across touches, which sounds more sophisticated but introduces a new problem: the rules for how credit gets distributed are themselves assumptions. A U-shaped model assumes first and last touches matter most. A linear model assumes every touch matters equally. A time-decay model assumes recency matters most. None of those is true in any deep sense. They’re conventions. The value of multi-touch isn’t accuracy. It’s that it forces you to acknowledge that the journey has more than two points.
The actual practitioner move, the one most posts don’t tell you, is that mature teams run multiple models simultaneously and use them for different decisions. The first-touch report tells the demand-gen team which channels are working. The last-touch report tells the field marketing team which programs are converting. The multi-touch report tells leadership how the funnel is balanced overall. Nobody asks any of these models to be right in some absolute sense. They ask each model to be useful for the specific decision in front of them.
There’s a precondition for any of this working, and it’s the part that gets skipped most often. Your data has to actually reflect the touches it claims to reflect. If your last-touch field is silently being overwritten by a sync rule, your last-touch model isn’t measuring what you think. If your first-touch field never gets updated when a contact is re-imported, your first-touch model is corrupted. Multi-touch models compound this problem, because they rely on every touch being captured correctly. A multi-touch model on top of broken touch data is more confidently wrong than a single-touch model on the same data.
So the real sequence is: fix the touch data first, pick the kind of wrong that matches your decision second, and run multiple models in parallel third. Anyone who tells you to pick a model before doing the first step is selling you a tool, not a method.
The question to answer before you pick an attribution model isn’t which model is best. It’s which decision you’re trying to make, and whether your underlying data can support any model at all.