marketing-attribution

Marketing Attribution (models, windows, incrementality)

Attribution answers "who gets credit", never "what did spend cause"

Every rule-based model is an accounting convention chosen before the data arrives: it redistributes credit among observed touchpoints and cannot say what would have happened without the spend. That question is causal and needs a holdout. Attribution is for budgeting hygiene and daily routing; incrementality decides whether a channel deserves funding at all.

The model menu and what each one is biased toward

last-click / last-touch  → over-credits closing, high-intent, retargeting, branded search
first-click / first-touch→ over-credits discovery, upper funnel, prospecting
linear                   → equal split; ignores that touches differ in influence
time-decay               → weights recency via a half-life; still last-click-leaning
data-driven / algorithmic→ fits observed paths (Shapley/Markov-style); sees only tracked touches
MMM                      → aggregate time-series on spend/seasonality/price/promo; no user IDs
incrementality / holdout → randomized exposed vs withheld; the only design measuring causation

Rule-based models differ only in the weighting rule, so "which is right" is unanswerable in their own terms — all are wrong about causation and differ in which channel they flatter. Data-driven models inherit every observability gap: untracked, offline and privacy-suppressed touches are absent from the path, so credit flows to whatever is measurable. MMM needs long history and real spend variance and reaches untrackable channels, but stays correlational until calibrated against experiments.

Windows, lookback, and view-through

State the window on every figure: click-through lookback (commonly 7 days), view-through lookback (commonly 1 day), and the post-install conversion window. Changing a window changes the reported result with no change in reality, so rule that out before believing a step-change. View-through conversions (impression, no click) are the softest currency here: a 1-day-view and a 7-day-click conversion are different objects, never summable, and view-through-heavy "performance" is unproven without a holdout.