Attribution window

Definition

An attribution window is the length of time after an ad click or impression during which a resulting install or conversion is still credited to that ad interaction.

It is the single setting that most often explains why your ad network reports more installs than your own analytics does. Windows differ by network, differ between clicks and impressions, and on iOS are not yours to set at all — SKAdNetwork's are fixed by Apple. Comparing two reports without first comparing their windows is comparing two different questions.

Click windows and view windows are not the same setting

Every network keeps two separate clocks. A click-through window starts when someone taps the ad; a view-through window starts when the ad was merely rendered on screen. View windows are always shorter, because the causal claim is weaker — see view-through attribution for why that gap exists and how it inflates reporting when the two are summed into one number.

Why the same campaign reports different install counts

Three mechanisms produce almost every discrepancy, and they compound. Once you can name which one you are looking at, the numbers usually stop being mysterious.

The time-zone one is the most embarrassing and the most common. A network reporting in America/Los_Angeles and a warehouse reporting in UTC will never agree on a Tuesday, and both are right.

-- Last-touch attribution with an explicit 7-day click window.
-- The window is a predicate, not a property of the data: change the
-- interval and the same rows produce a different answer.
WITH candidate AS (
  SELECT
    i.install_id,
    i.installed_at,
    c.click_id,
    c.campaign,
    c.clicked_at,
    ROW_NUMBER() OVER (
      PARTITION BY i.install_id
      ORDER BY c.clicked_at DESC        -- last touch wins
    ) AS rn
  FROM installs i
  JOIN clicks c
    ON c.device_match = i.device_match
   AND c.clicked_at <= i.installed_at
   AND c.clicked_at >  i.installed_at - INTERVAL '7 days'
)
SELECT campaign, COUNT(*) AS attributed_installs
FROM candidate
WHERE rn = 1
GROUP BY campaign
ORDER BY attributed_installs DESC;

Two things are worth noticing in that query. The window is a WHERE clause, which means re-running last month's report with a 30-day window instead of 7 legitimately changes history. And ORDER BY clicked_at DESC encodes last-touch; swapping it to ASC gives first-touch, and neither is more correct than the other — see multi-touch attribution.

On iOS, the window is Apple's

SKAdNetwork does not have a lookback window you configure. It has fixed activity windows measured from install, and postbacks that arrive after each one closes with a deliberate random delay. The consequence is that a SKAN report and a click-window report are not two views of the same measurement — they answer different questions on different clocks.

The Android side stays under your control. The Play install referrer carries referrerClickTimestampSeconds and installBeginTimestampSeconds, which means you can compute the actual elapsed time for every install and choose a window from the observed distribution rather than from a default. If most of your genuine installs land inside four hours, a 30-day window is not measuring patience — it is collecting coincidences.

That distribution is also a fraud signal in its own right. A window wide enough to hide a click-spamming campaign is a window wide enough to pay for one, which is why click-to-install time belongs on the same dashboard as the window setting.

Choosing a window you can defend

1. Plot your click-to-install time distribution first. The window should cover the bulk of real conversions and stop, not extend to a round number someone liked.
2. Keep click and view windows separate in every report, always, even when a network hands you one merged figure.
3. Write the window into the report itself. A number labelled "installs" with no window attached is not reproducible.
4. Change windows rarely, and never mid-quarter. Every change rewrites the comparison you were about to make.
5. Accept that SKAN sits outside all of this and report it in its own column.

The honest position is that a window is a business decision wearing a technical costume. Widening it does not discover installs that were always there; it changes the standard of evidence you accept for a causal claim. Networks default to generous windows for a reason, and the reason is not accuracy.