A dashboard reports that paid search deserves 40% credit for last quarter's conversions, email 25%, and organic social 15%. These numbers look precise, feed directly into budget allocation decisions, and are, in a meaningful sense, made up — not because the math is wrong, but because the question they're answering doesn't have the clean, single answer the model implies it does.
Attribution rules are allocation conventions, not causal measurements
Last-click, linear, and time-decay attribution models each allocate credit across touchpoints using a fixed rule — all credit to the final touch, equal credit to every touch, more credit to touches closer to conversion. None of these rules is derived from evidence about which touch actually caused the customer to convert; they're conventions chosen for simplicity and consistency. A customer who saw a paid search ad, then an email, then converted after a direct visit could have converted from the direct visit alone — the paid search ad might have contributed nothing, or everything, or something in between, and the attribution model has no way to actually know which, because it never observes the counterfactual (what would have happened without that touch).
The core problem: attribution can't distinguish causation from a touch that was along for the ride
A customer already planning to purchase, who happens to also see a retargeting ad on the way to checkout, gets that ad credited with contributing to the conversion under most attribution models — even if the purchase would have happened at the identical rate with no ad at all. This is the central flaw: attribution measures correlation between touchpoint exposure and conversion, not whether the touchpoint caused an incremental conversion that wouldn't have otherwise occurred. Channels that reach people already likely to convert (branded search is the classic example) tend to look highly effective under standard attribution precisely because they're correlated with intent that already existed, not because they're causing new conversions.
What actually measures causation: incrementality testing
Incrementality testing — holding out a randomized group from a given channel's exposure and comparing their conversion rate to an exposed group, or running geo-level experiments where a channel is turned off in some regions and not others — measures the actual counterfactual: what happens with this channel versus without it. This is slower, coarser-grained, and can't attribute individual conversions to individual touchpoints the way a dashboard can. What it produces instead is a real, causally grounded answer to the only question that actually matters for a budget decision: does spending more on this channel produce conversions that wouldn't have happened otherwise.
Media mix modeling as a middle ground
For channels where clean holdout experiments are impractical (broad brand advertising, for instance), statistical media mix modeling — regressing aggregate outcomes against spend across channels over time, with proper controls for seasonality and other confounds — provides a coarser but still causally-motivated estimate, in contrast to touch-level attribution's allocation-by-convention approach. It requires real statistical rigor to do correctly and is still an estimate, not a certainty, but it's answering the same causal question incrementality testing asks, just with a different method suited to channels that can't be cleanly held out.
What to actually do with attribution dashboards
- Use attribution data for directional pattern-spotting (which touchpoints commonly appear in converting paths), not as a precise input to budget allocation
- Before increasing spend meaningfully on any channel based on its attributed performance, validate with an incrementality test if the channel and budget size justify it
- Be specifically skeptical of channels that look highly effective under last-click or linear attribution but reach audiences with obviously pre-existing intent (branded search, retargeting)
- Treat media mix modeling as the right tool for channels too broad or brand-oriented for a clean holdout experiment
The fix for attribution's false precision isn't a more sophisticated allocation rule — every allocation rule has the same underlying flaw. It's accepting that the precise, per-touch number is directional at best, and reserving actual budget conviction for whatever channels have been validated against a real counterfactual.