Skip to main content
Cover/05 · Window sensitivity
M Measurement Field Guide All topics
05
Figure 05 · Sensitivity

Stretch the lookback window — attribution moves, incrementality doesn't.

Same outcome. Three windows. Three answers.

The slider moves the attribution window — a backward-looking accounting rule that decides which prior touches get credit for a conversion. This is distinct from the measurement window (forward-looking, how long the test runs) — both are sometimes called "lookback" in different communities. Total observed conversions (110) and the true incremental lift (10) do not change when the attribution rule changes. Only the partition between "touched" and "untouched" inside that fixed total moves.

Drag · one outcome, three windows
1-day click 7-day click 30-day view + click
Baseline
Counterfactual, no marketing.
100
100
Attribution · all three windows
Window-sensitive · the bar moves with the lookback rule.
90
20
1-dayclick only
80
30
7-dayclick
70
40
30-dayview + click
current 7-day · impact = 30
Incrementality
Counterfactual-defined · window-independent.
100
+10
impact = 10
Attribution 20 → 30 → 40
Incrementality 10 · 10 · 10
Tighter ≠ honest

Even the 1-day click window credits 20 when the true causal lift is 10. Shorter windows shrink the gap; they don't close it. Clicks correlate with high-intent users who would have converted anyway — a 1-day-click attribution still double-counts. The fix isn't a better window; it's a holdout test alongside the attribution scorecard.

About view-through

Adding view-through (impression seen, no click — the 30-day tier above) often doubles or triples attributed credit, not the modest +10 shown here. Many marketers reject view-through entirely because "saw an ad" rarely causes anything. Apple's SKAdNetwork supports view-through only under tight constraints — a registered impression, a few-seconds minimum on-screen time, and a short view-through window — and iOS 14.5+ broke a lot of the looser view-through measurement that came before; some teams treat it as informational only and don't include it in iROAS calculations.

Per-channel variance

Channels respond to window changes very differently. Branded paid search: paths are short, attribution barely moves with window. Display retargeting: paths are long and impression-saturated, attribution scales rapidly. Email: usually last-touch only, window irrelevant. A 1d → 30d switch can leave search untouched while doubling display credit — silently rebalancing your channel rankings without anything changing in the world.

Takeaway Lengthening the window pulls more journeys into the "touched" bucket. The reported attribution number inflates without anything actually changing in the world. Incrementality is defined against a no-marketing counterfactual — not by which touches you choose to count. The business consequence: 30–50% budget reallocations across channels can stem from a window-rule change alone, even when underlying performance is identical.

What's the right window? There's no theoretical optimum without ground truth — and ground truth only comes from incrementality testing. The practical answer: pick the window your platform / team uses as a convention, hold it stable across reporting periods so trends are comparable, and calibrate periodically with a holdout to learn the ratio between your attribution scorecard and true incrementality for each channel. Also remember: same conversion can be attributed across platforms — display + paid search + email can each claim full credit independently. Within-platform window choice is only half the story; cross-platform double-counting is the other half. And the "10·10·10" incrementality line holds only for a fixed test measurement period; longer measurement windows can return different incremental estimates because of carryover and adstock.

Methods note

Numbers throughout are illustrative. The attribution figures that swing with the lookback (20 → 30 → 40) and the flat incrementality line they're measured against (10 · 10 · 10) are chosen to make the lookback's leverage visible; real curves vary by channel, funnel stage, and touch definition.

Further reading
  • Localized Shift vs Overall Causal Impact
  • Adstock & attribution window considerations
  • Test Design · Power, α, p-value, tails
  • Superiority vs Non-inferiority