Choose an attribution window

A window says how long after seeing or clicking an ad a conversion may still be credited to it. Platforms usually distinguish the two — a click is stronger evidence than a view — and they carry separate lengths.

The windows are the platforms' own settings, configured on their side. Tempio reads what each platform reports under whatever window it is configured with. Nothing here re-attributes conversions, holds a window of its own, or reconciles the platforms against each other, because deciding which platform deserves credit for a shared sale is a judgement about your business rather than an arithmetic operation.

What Tempio does hold is the counts, one row per entity per day, with every ratio computed at the moment you look. That matters for this subject in one specific way: because nothing is stored as a rate, a window change does not corrupt history — it changes what the platform reports from that day forward, and the counts before and after are simply counts of different things.

The number comes from your order data, not from a benchmark. A figure about the world stated with no source behind it is exactly what this subject is already drowning in.

Take your own distribution of time between first touch and purchase. Where does most of it sit — a day, a week, three weeks? The window that captures most of that distribution is the honest one. A window materially shorter than it under-counts by construction. A window much longer than it is mostly adding purchases the advertising had little to do with.

Two adjustments.

Considered purchases stretch. Something researched over a fortnight needs a window that covers the fortnight or it will look worse than it is, and that is a business fact rather than a preference.

Repeat purchases contaminate. If existing customers buy again inside the window and the advertising is aimed at new ones, a long window quietly credits the ad for retention. Separate the audiences before you lengthen the window.

The arithmetic that makes any of this actionable runs on your margin rather than on an industry figure: break-even CPA and break-even ROAS take your own numbers, and what they produce is the line the attributed figures have to clear.

This is the whole discipline and it is harder than choosing.

Changing a window changes what counts, so every comparison across the change date is a comparison of two different measurements wearing the same label. It will look like a performance change and it is not one. If a window genuinely has to move, treat the change date as a break in the series and say so wherever the numbers are quoted afterwards.

The pressure to move it always arrives at the same moment: a bad period, where lengthening the window would make the figures acceptable. That is the moment a window stops being a measurement decision and becomes a narrative one, and it is worth deciding in advance that you will not.

The useful stance is that the window is a constant you judge changes against, not a variable you tune until the answer is tolerable.

Three habits that stop an attributed number from being over-read.

Never read a blank as a zero. A ratio whose denominator is zero is empty rather than nought, and the result says so. No conversions means no cost per conversion; it does not mean acquisitions were free.

Do not sum across platforms. Two platforms can each honestly claim the same sale, so the sum of what they claim can exceed what happened. A blended figure is useful as a trend and dangerous as a total, which is why Read a blended report honestly exists.

Watch for restatement. A long window means more of the recent past is still moving, and for a mechanical reason: a conversion credited three weeks after the ad lands on the day of the ad rather than on today, so a long-window account's last fortnight keeps being rewritten under you. The connection carries a restated mark when a sync rewrote a lot of stored history, and on a long-window account that mark is routine rather than alarming.

The arrangement the reading side sits inside, stated once: your connections read your accounts, a change is written as a draft on the record with the state it would change from and the checks it passed, applying it is gated on approval — by you, or by a rule you wrote — and no platform write reaches a customer account until that platform has granted write access. Applying a cleared change to a platform is in private beta.

Whether the advertising caused the purchase. No window can, because every window is a rule for assigning credit rather than a measurement of cause.

The only evidence that answers it is a test that withholds the advertising from a comparable group and measures the difference. Running one is a decision about time and volume rather than a reporting setting, and most accounts are too small for it on most questions.

Which leaves the practical position: pick a window from your own time-to-purchase, hold it constant, judge changes against it, and treat the absolute figures as a currency you use consistently rather than as truth. That is a worse answer than a real test and a considerably better one than moving the window until the number is acceptable.

Why platform numbers disagree is the wider version of this problem, and Build a weekly performance review is the rhythm that keeps a constant window honest.