We changed an attribution window as a housekeeping exercise and the reported performance of three channels changed materially. No campaign changed. No budget moved. The same events produced a different picture.
That was the moment I stopped treating attribution reports as measurements and started treating them as models with settings.
What a window actually is
The period after an interaction during which a subsequent conversion will be credited to it.
A seven day click window means a purchase eight days after a click is not attributed to that click. A thirty day window means it is.
There are usually separate windows for clicks and for views, and view windows are typically much shorter, for reasons that are contested.
None of these numbers is derived from anything about your business. They are defaults, chosen by platforms, and they are frequently different between platforms measuring the same activity.
Why the choice changes the answer
Different channels operate at different distances from the decision.
Display and video generally influence people earlier, and any conversion that results comes later. Short windows systematically under-credit them.
Search, particularly branded search, sits at the moment of intent. Conversions follow quickly, so it is fully credited under any window.
Which means the window choice determines the apparent ranking of channels, and shortening a window will always appear to favour the ones closest to the conversion.
If your product has a long consideration period, a short window is measuring the wrong thing. If purchases are impulsive, a long window is crediting interactions that were irrelevant.
The mismatch between platforms
The practical mess that follows.
Each platform reports on its own data with its own windows and its own model, and each claims conversions it can see.
Sum the platform-reported conversions across three channels and the total frequently exceeds the actual number of conversions, sometimes substantially. Everybody has counted the same purchase.
This is not fraud; it is each system honestly reporting what it observed under its own rules. It does mean that platform-reported numbers cannot be added together, which people do constantly.
The only reconciliation is an independent source of truth — your own conversion data — and comparing the sum of claims against it.
What signal loss changed
The context that has made all of this harder over the last few years.
Restrictions on third-party cookies, mobile identifier changes, and privacy regulation have reduced the ability to observe individual journeys across sites.
The response has been a shift toward modelled conversions, where platforms estimate the conversions they cannot observe.
Modelled numbers are not measurements. They are estimates produced by a system with an interest in the outcome, using a methodology that is generally not disclosed in detail.
That is not an accusation of dishonesty. It is a statement about what the number is, and it should change how much weight it carries.
What we do now
We set windows deliberately, based on our own observed lag between first interaction and purchase, which we measured from our own data rather than accepting a default.
We report platform numbers as directional and reconcile everything against our own conversion records.
We hold the window constant when comparing periods, because changing it makes any time comparison meaningless, which is exactly the mistake we made.
And for anything important, we run a holdout test, because that is the only method that answers the causal question rather than the correlational one.
The framing that helped
Attribution answers the question of which observed interactions preceded a conversion.
It does not answer the question of which interactions caused it, and people use it as though it does.
Once the team internalised that distinction, the arguments about which model was correct largely stopped, because everybody understood we were choosing a lens rather than discovering a fact.
The productive question became what decision we were trying to make, and which view of the data was least misleading for that specific decision. Different decisions genuinely warrant different settings, which sounds like fudging and is the honest position.
Comparing across platforms
A practical note for anybody assembling a cross-channel view.
Before comparing platform-reported numbers, align the windows and the models as far as each platform allows. Many permit configuration, and comparing a seven day view against a twenty-eight day view is meaningless.
Where alignment is impossible, note it explicitly in the report rather than presenting the numbers side by side as though they are comparable.
The habit that has saved us most argument is putting the window and model in the column header of every table. It is ugly and it prevents an entire category of misunderstanding.
View-through, specifically
The most contested setting and worth a position.
A view-through conversion credits an impression that was served and not clicked. Whether that represents influence or coincidence depends entirely on whether the impression was actually seen, which viewability measurement addresses only partially.
Our practice is to report view-through separately rather than folded into a total, so that anybody reading the number knows what it contains.
We also weight it down heavily in any budget decision, on the basis that an unclicked impression on a page somebody scrolled past is weak evidence of anything.