Building revenue attribution that survives scrutiny
First touch, last touch and multi-touch all produce different numbers from the same data. Here is how to pick a model you can defend and what to do about the gap.
Attribution arguments are rarely about attribution. They are about budget, and the model is chosen after the conclusion. Being explicit about that up front makes the conversation shorter and the reporting better.
What each model actually answers
There is no correct model. There are models that answer different questions, and using one to answer another question is where the arguments come from.
First touch answers: what creates awareness? It over-credits the top of the funnel and it is the right model for evaluating demand generation in isolation.
Last touch answers: what closes? It over-credits the bottom — usually a demo request form that would have been filled in regardless — and it is the right model for evaluating conversion mechanics.
Linear multi-touch answers: what was involved? It credits everything equally, which is defensible precisely because it makes no claim about importance. It is the most honest model and the least useful for making a decision.
W-shaped (first touch, lead creation, opportunity creation) answers: what moved the deal between its stages? It is the most useful for B2B with long cycles, and the most work to maintain.
The rule that prevents most arguments
Pick the model before you run the analysis, write down why, and do not change it for a quarter.
This sounds procedural. It is the single highest-value governance rule in reporting, because attribution models are trivially easy to select post-hoc to support a conclusion, and everyone in the room knows it.
Where every model breaks
Three problems no model solves. Say them out loud in the meeting rather than being caught by them.
The dark funnel
A prospect hears about you on a podcast, reads three articles over six months without ever identifying themselves, asks a peer in a private Slack, and then types your domain directly. Attribution records: direct traffic.
Everything that mattered is invisible. This is not a tracking problem to be solved with more tracking; it is a structural limit. The mitigation is a self-reported attribution field on the demo form — one open question, "how did you hear about us?" — which is imprecise and is the only instrument that sees the dark funnel at all.
Sales-influenced marketing and vice versa
An SDR books a meeting with an account that has read nine of your articles. Marketing claims content. Sales claims outbound. Both are right, and the model you chose decides who gets the number.
The useful framing is not "who gets credit" but "what would have happened without each?" You cannot answer that from attribution data. You can answer it from a holdout test, which is the only method here with real inferential power and the one nobody runs because it means deliberately not marketing to a segment.
Time lag
A B2B cycle of nine months means the pipeline you closed this quarter was created by spend from three quarters ago. Reporting spend and revenue in the same period compares two unrelated things.
Report against cohort by creation date, not by close date, and the picture becomes interpretable — at the cost of the most recent two quarters being incomplete, which is a real cost and must be stated on the chart rather than in a footnote.
What to actually build
- One primary model, chosen deliberately, applied consistently. W-shaped if cycles are long, last touch if they are short.
- A secondary model shown alongside it. The gap between two models is information: a channel that looks strong in first touch and weak in last touch is doing awareness work, and that is a finding, not a contradiction.
- Self-reported attribution on every inbound form. Imprecise, and the only view into the dark funnel.
- Cohort reporting by creation date. With the incomplete recent periods marked.
- One holdout test a year. Uncomfortable, and the only thing on this list that establishes causation rather than correlation.
Attribution tells you what happened. Only a holdout tells you what would not have happened otherwise. Most reporting confuses the two, and most budget decisions are made on the confusion.