Attribution waterfall

The attribution waterfall determines how Adjust selects the best engagement to award for attribution. The waterfall methodology is designed to reflect the accuracy of the two attribution methods that Adjust uses: deterministic attribution and probabilistic modeling.

Adjust follows this order when performing attribution:

  1. Click-based determinsitic attribution OR click-based referrer matching (Android only)
  2. Click-based probabilistic modeling
  3. Impression-based deterministic attribution
  4. Impression-based probabilistic modeling
  5. Organic

How it works

Clicks vs. impressions

Adjust always prioritizes clicks over impressions, even if the impression happened minutes before install. This is because we consider click engagements to be more deliberate and active.

Deterministic vs Probabilistic modeling

Deterministic attribution is preferred over probabilistic modeling because it is the more reliable and accurate method. Probabilistic modeling uses a variety of different datapoints, some of which can change - such as the user's IP address.

Waterfall in action

From the moment of install Adjust works backwards looking for data points within the attribution window. Then, the data points are judged to determine which holds the most robust information and can be awarded the attribution.

  1. Click with determinsitic matching: First, we look to see if there are any clicks with the same device ID or advertising ID as the device the app was installed on.
    • Click with Android referrer: For Android devices, we also check for a match with the Play Store referrer. If there are two engagements that happen at the same time, and one has the referrer and the other has a device identifier, we give preference to the click with the referrer.
  2. Click with probabilistic modeling: If there's no click data with matching IDs, we look for click engagements that came through with other matching datapoints, such as IP, type of device, device name, and operating system. We create a scorecard for each click, and we award attribution to whichever has the most commonalities with the install information.
  3. Impression with deterministic matching: If there are no eligible clicks, we look for impressions that carry the same device/advertising ID as the install device.
  4. Impression with probabilistic modeling: If there are no impressions carrying device/advertising IDs, we check for engagements with other matching datapoints. We create a scorecard, and award attribution to the engagement with the most commonalities with the install information.
  5. Organic: If we go through all check and find no matching engagement, the user is attributed as organic.