Biz Tips: Reducing Churn With Machine Learning Powered Push Messages
GROWTH:
Reducing Churn With Machine Learning Powered Push Messages
According to research conducted by UrbanAirship, 95% of opt-in users who don’t receive a push notification in the first 90 days will churn.
And users who received more than one push message a day had 820% higher retention rates than users who received zero notifications.
But just blasting users with frequent messages isn’t a guaranteed way to reduce churn, and the wrong message at the wrong time could make churn even worse.
Research done by Localytics showed that push messages personalized based on a user’s profile and behavior increased CTR by nearly 3X, and lead to 4X as many post-message app sessions compared to non-personalized broadcast style push notifications.
Machine Learning Takes Personalization To The Next Level, and Helps You Anticipate When Users Are At Risk of Churning
Now, machine learning and predictive analytics are taking personalization of push messages to the next level. Allowing you to predict which segment of users is likely to churn before it happens. And giving you the power to steer users who veer off course back toward the optimal user experience, so they get the most value from your product and become longtime users.
And the best part is, you no longer need to have a deep understanding of statistics or programming to leverage the power of machine learning to grow your user base. In this article, we’ll look at three SaaS solutions that will allow you to set-up your own predictive push messaging campaigns right out of the box.
We’ll also look at some of the limitations of these SaaS solutions, and explore some cases where you might want to invest in the development of a custom churn prediction solution.
Segmenting Users By Churn Risk and Sending Perfectly Timed Push Messages With UrbanAirship
The first SaaS tool we’ll look at is UrbanAirship, which offers a suite of messaging features including both web and mobile push notifications, SMS, and email.
These messaging features all integrate with UrbanAirship’s churn prediction tool. This machine learning model looks at two key sets of data to make a prediction on how likely a user is to churn:
- How recently and frequently a user opens your app.
- How recently and frequently they are receiving push messages from you.
The first one is kind of a no-brainer — if a user is using your app less, it makes sense that they’ll be more likely to churn.
But UrbanAirship also found a strong correlation between the frequency of push messages and the likelihood of users to churn. They analyzed 90 days of app usage for 63 million new app users, and compared users who received zero notifications with users who received more than one notification a day. The users who received more than one notification per day showed 3X retention rates on iOS and 10X retention on Android, compared to the users who received no notifications at all.
Combining this app open and push message frequency data, UrbanAirship’s machine learning model will rate a user’s likelihood of churning on a three category risk scale; high risk of churning, medium risk of churning, and low risk of churning.
You can then segment push messages based on a user’s churn risk category. You can also add other non-predictive variables to build more narrowly defined segments around the churn prediction. UrbanAirship lets you create segments based on things like location data (if users have been in a certain geographic location within a set timeframe), actions a user performs like browsing a piece of content on the web or in your app, or device characteristics.
UrbanAirship also allows you to set-up triggered messages when a user moves from one churn risk category to another, for example when a user moves from a low to high churn risk. They also leverage machine learning to figure out the optimal time to send push notifications to each individual user, so you can automatically time messages to arrive when users are most receptive and engaged.
Crafting Hyper Personalized Push Messages, and Relevant Churn Interventions
Layering churn risk on top of other segment data allows you to craft even more personalized push messages. But most importantly, it allows you to identify users most at risk of churning and intervene before it’s too late.
That intervention might involve something like…
- Making some kind of special offer to draw users back into the app.
- Increasing push message frequency to re-gain some mental real estate.
But push message retention campaigns will work best if they’re driven by a deeper understanding of why users are at risk of churning in the first place. This is where combining a machine learning derived churn risk score with your own research and pre-defined segments can be really powerful.
For example, you might identify a correlation in your own data analysis that engagement with a specific feature is key to long term retention. You could create a segment for users who haven’t engaged with that feature and combine it with the high churn risk segment, then create a push campaign educating users on the feature along with an extended trial to draw them back into the app.
Unlocking The Power of Predictions For Individual Events, and Mapping A User’s Journey Between Churn Risk Segments With CleverTap
CleverTap’s machine learning model takes a similar approach to UrbanAirship, using recency and frequency as the main data points to score a user’s churn risk. However, it gives you more flexibility in how to apply it — letting you run a churn risk calculation on any event you choose, instead of just using app opens.
So, if you’d like to see the churn risk for a specific product, feature, or action you can do that. It’s worth clarifying that these deeper churn drill downs aren’t what would classically be considered as churn. If you ran a calculation using engagement with a specific app feature for example, you’d get a score of how likely users would be to continue using that feature. Not whether or not they would churn, or stop using your app completely.
That said, applying machine learning to specific events like this could be very useful if you’ve already conducted analysis and found correlations between specific features or products and long term customer retention. If you’ve found the use of a specific feature to be highly correlated to long term retention, you could set-up personalized push campaigns around this specific feature and a user’s likelihood to continue using it or not.
CleverTap also breaks down churn risk in a more granular way, creating 10 separate segments with varying levels of churn risk instead of UrbanAirship’s three.
Not only does CleverTap automatically build segments around its churn predictions, it also gives you a big picture overview of how users are moving in between those segments. For example, you might find new users are moving into a high churn risk segment shortly after signing up. This will give you some insight into where your problem areas may be. In that case, you may want to focus a push message campaign on onboarding those new users.
It also assigns a revenue estimate to each segment to help you allocate resources more effectively. And it gives you a breakdown of what percentage of each segment is reachable via different channels like push, sms, and email. If a big chunk of a segment has opted out of push notifications but have given you permission to email them, you’ll know which channel to focus your resources on.
Building a More Accurate Churn Prediction Model With Firebase
Where Firebase really shines over solutions like CleverTap and UrbanAirship is in the data it uses to model its predictions.
While CleverTap and UrbanAirship focus on recency and frequency of a specific event to build their prediction, Firebase Predictions integrate with all the data points available in Google Analytics to build a much more complete model of user behavior.
In other words, instead of only looking at a user’s history of app opens to predict future app opens, Firebase looks at all of a user’s behavior data collected in Google Analytics to predict if they will open the app in the future. Essentially bringing dozens of additional data points into consideration in formulating the prediction.
The ultimate benefit to you is greater accuracy in the predictions, which Firebase puts a lot of importance on. In the control panel, you can set the risk tolerance level of each prediction — high, medium, or low. With high risk allowing for a greater number of potential false positives (users who are predicted to churn, but ultimately don’t). And a low risk offering the most precise prediction, but providing a smaller segment for targeting.
The biggest drawback of Firebase is that it only provides predictions on behavior seven days into the future. For example, when predicting churn you’ll only be able to identify a segment of users likely to churn in the next seven days. For B2B SaaS companies with long customer lifecycles, seven days notice is likely too little too late to salvage a broken relationship.
However, Firebase offers a high level of customization when it comes to building segments based on predictions. Not only can you predict the likelihood of specific events across the customer lifecycle taking place, you can combine event predictions with other standard non-predictive variables to more tightly target a segment.
So you could target users who are predicted to not trigger key events occurring earlier in the customer lifecycle that you’ve identified as being strongly correlated with long term customer retention. Like completing key onboarding steps, or completing important milestone events deeper into the user journey.
As Predictive Analytics Goes Mainstream, Be Aware of These Limitations
As predictive analytics is still it its infancy, all of these out-of-the-box SaaS machine learning solutions have pretty significant limitations.
UrbanAirship’s prediction model only focuses on the recency and frequency of app opens, ignoring loads of critical user behavior that could be influencing churn.
CleverTap lets you drill into events beyond just app opens, but still only draws on data tied to the recency and frequency of the event it’s trying to predict.
Firebase builds a much more robust model using all the user behavior data from Google Analytics, but limits its predictions to a seven day window.
And the biggest limitation of all these solutions is that their churn predictions are essentially a black box. You get a prediction on which users are likely to churn, but you don’t know why. That’s a super critical piece of the puzzle if you’re going to make smart decisions about how to intervene to save the user relationship, or how you can improve your product and marketing experience to prevent users from getting bucketed into a high churn risk segment in the first place.