Attribution windows in physical retail
An attribution window is the period after exposure in which a purchase is counted as potentially influenced. In physical retail it has to be chosen rather than observed, because the exposure and the purchase are rarely linked at individual level.
The choice affects the result substantially, which is why it belongs in the plan rather than in the analysis.
What does the window have to reflect?
The plausible decision horizon for the category. An immediate-consumption item bought in the same visit has a window of minutes; a considered purchase made on a later trip has a window of days or weeks.
Using one window across categories with different horizons produces incomparable figures.
Why must it be set in advance?
Because a window chosen after the data is seen selects the result. Trying several windows and reporting the most favourable is a well-known route to a finding that does not replicate.
Pre-registering the window, even informally in the plan, is the discipline that prevents it.
What is the same-visit case?
The strongest one in this channel: a shopper sees a screen at the register and the product is within reach. The window is the visit, the measurement is the basket, and the link is as close as this medium gets.
It also limits the claim, because only some categories can be bought in that moment.
What about later visits?
In a daily-visit channel, a later visit arrives quickly, which makes multi-visit windows more meaningful here than in channels where the next opportunity is distant. A window of several days covers several visits for a regular shopper.
That is an advantage and it widens the window within which other explanations also operate.
How does the window interact with the analysis?
A longer window includes more sales and more confounding. The measured effect may rise and the confidence in attributing it falls, which is a genuine trade rather than a free choice.
Reporting a short-window and a long-window result, both pre-specified, is more informative than either alone.
What cannot be claimed?
Individual attribution. Without linking an exposed person to a purchase, the analysis is a store-level comparison, and describing it as attribution in the digital sense overstates what happened.
Aggregate attribution at store level is a legitimate and different claim.
How should the window appear in reporting?
Stated, with the rationale. A lift figure without its window is not interpretable, and two figures from different windows are not comparable.
What is the practical default?
The visit for immediate-consumption categories and a short multi-day window for others, both agreed before the flight. Anything longer needs a specific argument.
Frequently asked questions
Is there a standard window for in-store media?
No, and any figure offered as standard should be questioned, because the right window depends on the category's purchase cycle. Consistency within a brand's own measurement matters more than matching somebody else's convention.
Can I test several windows?
You can compute several and you should pre-specify which is primary. Reporting all of them with the primary identified is honest; reporting only the best one is not.
Does a longer window always show a larger effect?
Often, because it includes more sales, and the additional sales include more that the campaign did not cause. Larger and less attributable move together.
How does this differ from digital attribution?
Digital links an identified exposure to an identified conversion; this compares exposed and unexposed stores. The word attribution is used for both and the mechanisms are not comparable.
What about a campaign aimed at awareness?
Sales windows may be the wrong measure entirely, and a brand measure over a longer horizon may fit the objective better. Matching the measure to the objective precedes choosing the window.
Should the window include a pre-period?
The analysis should, as a baseline. The attribution window itself is the post-exposure period.
Does the window affect the control group?
It applies identically to both groups, which is the point of a controlled comparison. Applying different windows to test and control would be a serious error.
How do I handle a promotion in the window?
Either exclude the period or model it, and say which. A promotion inside an attribution window is a large confound.
What is the single rule worth holding?
Set it before you look. Industry practice on measurement standards is discussed by bodies such as the IAB.
### How should two windows be reported together?
As primary and secondary, with the primary named in advance. Reporting both is informative; reporting whichever is larger without saying which was primary is selection.
The difference between them is also interesting in itself, because it indicates how much of the effect was immediate rather than accumulated.
Does the window need to match the category's purchase cycle exactly?
It needs to be defensible rather than exact. A window materially shorter than the cycle misses real effects and one materially longer absorbs unrelated variation, and the middle ground is wide.
### How does the window interact with the flight?
A window extending past the flight captures carryover and picks up more unrelated variation. A window contained within the flight measures only concurrent effect.
Both are defensible and the choice should follow the objective. An activation campaign is usually measured concurrently; a campaign expected to build is measured over a longer horizon.
Should the window differ by store?
No, it should be uniform for comparability. Varying it by store introduces a difference between test and comparison groups that is not the campaign.
What if the category has no obvious cycle?
Use the store's own repeat interval for the category as an estimate, which is derivable from transaction data. It is better than a convention borrowed from another channel.