What attribution is for
Attribution assigns credit for an outcome to the things that may have caused it. The mistake that organises most attribution debates is treating it as one question with competing answers. It is three questions: each method answers one of them well and the others badly.
Before any of that, the point most attribution work skips: a tracked click proves nothing about causation. It tells you a send was opened and acted on, not that the send changed what the user would have done. Only a controlled holdout establishes that. No method means much until you accept that it only approximates a causal answer you can get cleanly from a holdout.
Last-click and multi-touch
With that caveat in place, last-click gives all credit to the final touch. It is simple, available and systematically wrong in one direction: it over credits whatever sat closest to a conversion that was often already going to happen, while crediting nothing to the demand built earlier. Multi-touch attribution (MTA) spreads credit across the observed journey and is sharper for near real time, within channel optimisation, but it only sees the trackable slice and that slice has been eroded by cookie deprecation, Apple’s App Tracking Transparency and the walled gardens. Both are bottom up and correlational. Neither establishes cause.
Marketing mix modelling
MMM works top down, regressing outcomes on aggregate, time-series spend and external factors. It values offline and brand alongside digital. Needing no user-level tracking, it also survives the privacy decay that breaks MTA. A credible model captures what a naive regression misses. Adstock is the carryover by which a campaign’s effect decays over later weeks rather than landing all at once. Saturation is the diminishing return as spend in a channel rises. The model fits a curve, not a straight line. The costs follow. Estimating those shapes needs long histories. The model needs real expertise to build and validate. It is vulnerable to multicollinearity: channels whose budgets move together are hard to tell apart. A clean incrementality test is the standard way to calibrate and sanity-check the coefficients. MMM smooths over short-term and tactical effects; its answer is how to allocate budget across the portfolio, not which message to send next.
Incrementality as the referee
Only a controlled experiment, treatment against a randomised holdout or geo control, measures causation directly. Incrementality does not replace attribution and MMM; it calibrates them, acting as the ground truth a disagreement is resolved against. When last touch says one number and MMM says another, a clean lift test is the tiebreaker. See holdouts and control groups.
What this means for a lifecycle programme
A tracked click is not proof of cause
Owned channel attribution looks easy: the click is yours to track. That is exactly the trap. A tracked click is not proof the send caused the outcome.
Build the read on the authenticated cohort, where identity is deterministic and you can follow channel to destination and destination to outcome. Anchor it on a holdout rather than on the click. The destination conversion frame is also the one that survives platform intermediation and the agentic shock. See measuring intermediation.