Recording is not explaining

A tracking system produces a record of touchpoints. An attribution claim is a causal statement about which of those touchpoints mattered. The gap between the two is not an engineering problem you can close with more instrumentation.

Every attribution model is a set of assumptions wearing a number as a disguise. Last-touch assumes the final interaction did the work. Linear assumes every touch contributed equally. Neither is true, and neither presents itself as an assumption.

Why more data makes it worse

More touchpoints mean more ways to divide the same fixed credit, which makes the output more precise-looking and no more correct. Precision and accuracy diverge, and the interface usually reports the former.

It also creates a false sense that the remaining uncertainty is a measurement gap rather than a structural one. It is structural. You cannot observe the counterfactual.

What we report instead

Our attribution view reports a range, the model that produced it, and the assumptions that model makes. Changing the model changes the number in front of you, visibly, which is the honest presentation.

• A range, not a point estimate
• The named model, switchable, with its assumptions stated in plain language
• An explicit confidence signal that drops when the data is thin
• A list of what the view cannot tell you

What it refuses to claim

It will not tell you that a channel caused a number of conversions. It will tell you that under a stated model, a channel is associated with a range, and that a different reasonable model gives a different range.

This is less satisfying than a dashboard with one big number. It is also the only version that survives being checked.