The lookback window: the assumption inside every conversion figure

A lookback window is a counting rule: how far back from a conversion the system will look for an ad interaction, and which interactions are allowed to claim it. Both parts are set independently, and neither is a property of the customer's behaviour. They are decisions about bookkeeping that determine the total on the report.

This is why 2 platforms report different conversion counts for the same week without either being broken. They are answering different questions. Until both windows are stated, the numbers cannot be compared, added, or used to rank channels.

Three settings hiding inside one number

Length is the obvious one, and the one most people mean by attribution window. The other 2 do more damage precisely because they are rarely discussed.

• Length. How many days before the conversion an interaction may sit and still claim it. Longer windows always credit more conversions, never fewer.
• Interaction class. A click is a deliberate act by the person. An impression is something that happened near them. Some platforms add an intermediate class for a video watched past a threshold. These are different events and a window admits them separately.
• Date basis. Whether the conversion is reported against the day it happened or the day of the interaction that earned it. This changes nothing about the total for an all-time count and everything about a weekly series.
• Precedence. When both a click and an impression qualify, which one takes the credit, and whether the other is discarded or counted separately.

No setting here is stable across vendors, and none is safe to assume from memory: defaults change, and the platform's own documentation is the only place worth reading for current behaviour. What survives that churn is the mechanism, so this page describes the classes rather than naming anyone's default.

The same 120 sales, counted five ways

Take one illustrative month: 120 sales, one channel, one fixed budget. The interactions preceding those sales do not change. Only the rule for counting them does.

From the top row to the bottom, credited conversions rise by 74% and cost per conversion falls by more than 40%, with identical spend and identical sales. Nothing happened in the market. If those 2 rows sat in 2 different platform reports, a media team would rationally move budget toward the channel that is merely counting more generously.

Two channels reported under different windows are not 2 measurements of performance. They are 2 different definitions of the word conversion, presented in the same column.

What has to be true before 2 channels can be compared

The only durable fix is to stop comparing platform-reported figures and recount from your own event data under a single rule you control. That means keeping the raw touch history rather than only the credited totals, because a window can be re-applied to raw interactions and cannot be recovered from a summarised count.

1. Record every qualifying interaction with a timestamp, a channel, an identifier and its class. Without the class, a click and an impression become indistinguishable a month later.
2. Define one house rule and version it. One length, one set of admitted classes, one precedence order, written down with the date it took effect.
3. Recount conversions from the raw table under the house rule, and keep the platform figure beside it as a second column rather than replacing it.
4. Reconcile the difference once, then monitor it. A stable gap is a definitional difference to document; a moving one is a defect to chase.
5. Re-verify after every connector change, because the window is a parameter of the request and defaults drift — the risk described in a connector is a contract with someone else's API.

Writing the window into the report, not the appendix

Put the rule on the page: window length, admitted classes and date basis, per channel, in the same view as the numbers. A report whose definition lives in someone's memory produces an argument every time a platform changes a default. The reporting layer that carries those definitions with the figures is ordinary internal tooling, and it is cheaper than the meeting it prevents.

Treat a change to the window as a restatement: recount history under the new rule, publish both series for a period, and say plainly that the movement is definitional. It matters beyond the report, because an outcome column assembled under an inconsistent window feeds every downstream model, and it is one of the quieter reasons a mix model returns something nobody believes, traced in the model that hands a channel a negative coefficient. The reason recent weeks look weak is conversion lag rather than the window, covered in how conversion lag bends a weekly read. All of it sits inside the attribution, incrementality and mix work we do for marketing and advertising teams.

• attribution
• measurement
• reporting
• conversion tracking