Short vs. long conversion windows: the single setting that rewrites every model

Short vs. long conversion windows: the single setting that rewrites every model

Before arguing about Markov versus Shapley, settle the window — the lookback period during which a touch can claim a conversion. It silently dictates which channels win.

The mechanical effect
A 1-day window credits only touches just before conversion: bottom-funnel, retargeting, branded search. A 90-day window lets upper-funnel awareness touches claim conversions months later. The same data, re-windowed, can flip a channel from "top performer" to "waste" with no change in reality.

Why longer isn't more accurate
Intuition says a long window is more complete. But a long window is also where coincidence accumulates: the longer you look back, the more likely a channel appears in the path by chance rather than influence. Long windows inflate view-through and upper-funnel credit precisely because they harvest unrelated conversions. Short windows do the opposite — they amputate genuine long-cycle influence and over-reward closers.

Match the window to the cycle, not to a default
The right window is empirical: look at your actual time-to-conversion distribution. If 80% of conversions happen within 7 days, a 30-day window is mostly adding noise. If your B2B cycle runs 60 days, a 7-day window is structurally blind to demand generation.

Bottom line for practitioners: Never inherit your tool's default window. Derive it from your time-to-conversion distribution, and report results at two windows (short and long) to see which channels are window-dependent — those are the ones whose credit is least robust. The window is not a setting; it's an assumption about how long influence lasts, and it deserves the same scrutiny as the model itself.