Attribution windows: how to choose one and stick to it

Attribution is not a measurement problem, it is an accounting convention. There is no true answer to which touch caused an order, because causation is not recorded anywhere. What you choose is a consistent rule for assigning credit, and the value comes almost entirely from applying it the same way every time.

Shops lose more analytical value to changed windows than to badly chosen ones.

Choosing the window

Match the window to how long your customers actually take to decide. That is a number you can compute: the median gap between first session and first order.

• Consumables and repeat purchases — short. Seven days post-click is usually generous.
• Considered purchases — furniture, electronics, anything over a few hundred euro — genuinely need 30 days, sometimes more.
• Email specifically is well served by 7 days post-click. Emails are opened and acted on quickly or not at all.
• Never use a window longer than your data retention. A 90-day window over 30 days of events silently truncates.

View-through is where the argument starts

Click-through attribution requires an action. View-through requires only an impression, and it is why ad platform numbers are so much larger than yours. It is not fraud — it is a different and much looser definition, and each platform applies it to its own inventory.

• Report click-through internally. It is tighter, comparable across channels, and harder to inflate.
• Look at view-through for upper-funnel work where clicks are genuinely rare, but never add it to click-through totals.
• Never sum across platforms. Meta and Google will both claim the same order, and the total will exceed your actual revenue.

Last touch, first touch, and why last touch usually wins

Last touch credits the final interaction before the order. First touch credits the first. Linear and time-decay models spread credit across the path.

For most shops, last touch is the right default — not because it is accurate, but because it is stable, comprehensible and cheap. Multi-touch modelling requires complete path data that consent rules increasingly prevent you from having, and an incomplete multi-touch model is more misleading than an honest simple one.

The exception worth making: report first touch separately for acquisition channels, so paid search does not look worthless because email closed the sale it started.

Write it down, in one place

The document is four lines and it prevents most reporting arguments:

1. Window: 7 days post-click, for all channels.
2. Model: last non-direct touch.
3. Revenue basis: goods total, net of refunds and returns, excluding shipping.
4. Changed on: never, without a note here and a restated comparison.

When you do change it — and one day you will — restate at least the previous year on the new basis before reporting a single number on it. Otherwise the change looks like performance.

Living with the gap to the platforms

Your numbers will always be lower than Meta's and Google's. The gap is structural: they count view-through, they attribute on their own window, they model conversions they cannot observe, and they have every incentive to be generous.

• Use platform numbers to compare campaigns within that platform. That is what they are good for.
• Use your own numbers to compare channels and to decide budget.
• Track the ratio between the two per channel. A stable ratio is fine. A ratio that suddenly moves means something changed in tracking, consent or the platform's modelling, and that is worth finding.