To set up distribution modeling, follow these steps in the Adjust dashboard.
- Find your app and select your app options caret (^)
- Select All Settings > Fraud Prevention
- Switch the DISTRIBUTION MODELING toggle ON
The LEVEL option lets you set the distribution modeling threshold. Change this as you move between testing and production environments.
Advanced level (recommended): Allows the filter to work at full efficiency and should be used for full production mode
Standard level: Intended for trials and testing purposes only
- Navigate to your app and select your app options caret (^)
- Select Statistics
- Select the Fraud Prevention tab
Installs rejected for click spamming will appear in one of the following columns:
- Rejected Installs: Too Many Engagements (RI TME)
- Rejected Installs: Distribution Outlier (RI DO)
Reattributions rejected for click spamming will appear in one of the following columns:
- Rejected Reattributions: Too Many Engagements (RR TME)
- Rejected Reattributions: Distribution Outlier (RR DO)
Note: Installs that are rejected for click spamming will either be attributed to an authentic source found through Adjust’s attribution methodology, or, if no other source is found, to your Organic tracker.
For information on the fraud prevention KPIs and how to read your Statistics, see our fraud prevention reporting article.
At Adjust, we define click spam as all illegitimate click activity. For fraudsters, the goal of click spamming is to poach attributions from your organic users - that is, to have a certain number of your organic installs falsely attributed to a network. This way, their campaigns appear to generate high volumes of valuable users.
Ultimately, not all click spam is premeditated fraud. It can range from networks that send views as clicks to servers that send catalogs of artificial clicks. Another common example is when an app invisibly loads and clicks ads in the background.
Adjust’s method for rejecting attribution from click spam is based on the click-to-install-time distribution. The first step is to disqualify high-frequency clicks that try to manipulate the click-to-install distribution. The second step is to reject attribution using distribution modeling.
In order to mimic realistic click-to-install-time distributions, fraudsters will repeatedly send the same click in recurring intervals. In doing so, they produce a “last click” that is always relatively close to the install.
When an install occurs, Adjust checks all eligible clicks within the relevant attribution window. If we recognize any high-volume click patterns, we remove the clicks from consideration. This allows us to correctly attribute the install to the next legitimate click or as an organic user.
Once we have eliminated all attempts to manipulate the click-to-install-time distribution we can apply distribution modeling to detect the remaining click spam.
We developed our method of distribution modeling by reviewing statistical data and analyzing actual fraudulent activity in real-time. Based on this research, we determined that more than 85% of installs were recorded within the first hour of click time. This behavior indicates a strong correlation between the click and the install time.
Fraud, however, shows no such correlation between click and install. Since the user never actually clicked, and was never redirected to the store, their install will be independent of the click time. Therefore, when organic users are randomly poached by click spam, the click-to-install-time distribution will be evenly spread out across the entire attribution window.
Knowing this, we defined a lower threshold for installs recorded within the first hour of clicking. If the number of installs made after the first hour of click time exceeds a certain percentage of installs made within the first hour of clicking, Adjust will begin to disqualify clicks from attribution. Therefore, the install will be attributed to the next eligible tracked source or as organic traffic.