An app retargeting campaign, also called an “app re-engagement campaign”, is targeted toward users of an app who previously installed the app, were active for a period of time, and have churned. Because the audience is accurate and relevant, these campaigns can be highly effective and profitable – if done correctly. Furthermore, advertising efficiency can be improved based on historical data known about the user activity in the advertised app. Using data on user interests and behavior outside of the customer app can improve results even further.
Measuring the performance of a retargeting campaign in order to price its value and improve it – as opposed to measuring a user acquisition (UA) campaign – can be difficult. This is mainly because of two challenging characteristics of measuring retargeting campaigns that aren’t especially problematic for UA campaigns:
- Users can return to the app regardless of the campaign (this is known as the “natural return rate,” or NRR), and we’re aiming to measure only the number of users returning as a result of the campaign.
- An ad may remind a user to return to the app, and the user does come back – but without clicking on the ad. This happens more in a retargeting campaign than a UA campaign because the users already know the advertised app and may already have it installed on their device.
In order to properly measure a retargeting campaign, we first need to randomly split all the relevant users into two groups: a treatment group and a control group. The users of the treatment group are displayed ads of the advertised app, and are sometimes offered bonuses like free coins for returning to the app through the ad. The users of the control group are not shown ads of the app at all. Thus, it’s possible to compare the two groups and understand the extent to which the retargeting campaign added to the NRR of users to the app, and how much it increased sales within the app. The NRR is measured by the control group.
How should experiments using treatment group and control groups be performed?
First, the audience needs to be divided into two groups randomly. The groups don’t have to be the same size, but we recommend keeping the sizes as close as possible to avoid there being a different noise level in the measurements of the two groups. Next, two parallel campaigns should be performed, each targeting one of the groups.
There are two ways to perform the experiment:
- Wins vs. Wins. In this method, the campaign targeting the treatment group displays ads for the advertised app, while the campaign targeting the control group displays other, placeholder ads that are unrelated to the advertised app and don’t compete with it. In this experiment, it's important that all campaign parameters (price, settings, etc.) are the same for both campaigns. It's important to make sure that the treatment campaign and the control campaign don’t perform different campaign optimizations, although it is not the audience and the parameters that are best for the product advertised in the control ads. This experiment cannot be performed on a platform that only optimizes the campaign automatically, as this will lead to a difference between the campaigns, making it impossible to compare the campaigns. For this experiment, automatic optimizations must be turned off. Alternatively, if possible, define the two campaigns (of both groups) to perform the same optimizations according to the treatment group.
- Wins vs. Bids. In this method, the campaign targeting the treatment group displays ads for the advertised app, while the campaign targeting the control group doesn’t display ads at all, and only keeps the information about the users it might display ads to (for example, they were available to target in the exchange’s stream). With this, we cannot be sure whether we would have won and presented these users the ads, which causes a difference between the two groups, and limits our ability to compare them.
The table below shows the differences between the two methods:
In the experiments we performed, we saw no significant difference between the performance measured in the two different methods. This may indicate that the bias generated by the Wins vs. Bids method is small, and can be significantly reduced by statistical tools, thus reducing the experiment costs.
Once we've implemented the campaigns using one of the methods described above, we should measure campaign performance. The main measurements are Return Rate Uplift, ROI and Engagement Lift.
Return Rate Uplift
This measures how much the campaign increased the percentage of users who returned to using the app over the natural return rate.
ROI (Return on Investment)
This measures the ratio between the additional revenue from the users that returned due to the campaign, and the cost of the campaign. If we conducted the experiment using Wins vs. Wins, the cost of the ads presented to the control group is not included in the calculation, because this cost was required only for the experiment; it is not considered an investment. This measurement is performed in a cohort way, so the revenue is calculated on the first X days of each user in the app after returning.
The table below shows how to calculate the ROI:
This measurement examines the user’s additional active days in the app. An active day is defined by a unique day in which a user was active in the app. This measurement is also performed in a cohort way, so the count of active days is calculated on the first X days of each user in the app. If we want to normalize this metric to audience size, we can divide the extra active days by the number of users reached in the treatment group.
The table below shows how to calculate the extra active days:
ROI can be calculated separately for different time frames (for example, per week) and can be displayed over time for different time periods, as can be seen in the graph below:
There is much to be gained from retargeting campaigns. While they can be complicated to execute, following the tips provided here will help ensure that your retargeting campaign is done right. And remember, always split your audience randomly, and perform both campaigns using the same parameters (price, settings, etc.) and campaign optimizations.