Targeting and segmentation are powerful tools
...but the way they're practically executed makes a huge difference.
For example, if we create segments around the wrong user demographics or feature, our segmentation exercise won't deliver on its promises.
Typical marketing segmentation and targeting exercises are based on a relatively small number customer attributes, events, or actions, but more often than not, the events or attributes that we perceive to be the most important are much less significant than we think.
So, how do we know which events are *actually* most significant to achieving our user-focused goals and KPIs, and how do we use them to create more efficient triggers and segments?
This is where propensity modeling and targeting comes in.
What is propensity modeling?
“Propensity” is just another word for prediction, but it's a little more sophisticated than that. Propensities come from statistical models that predict whether or not a user will do something in the future (e.g., purchase, subscribe, churn, etc.) based on what they did in the past.
For propensity modeling to work effectively, you need to have a good idea of which actions tend to come after specific sequences of events. Therefore, you need a large set of "features" or events that you can observe many customers or users moving through (if you have a mobile app, it's your lucky day, because this is exactly what your event stream is).
Some of the events propensity modeling takes into account are those typically chosen by marketing teams (e.g. clicks, visits, subscriptions, purchases, etc.) but other events which can be considered by a propensity model are "sleepers" — events we typically gloss over which may actually provide a huge indicator of future customer actions — things like enabling GPS permissions, applying a gift card, or copying a wish list.
In some ways, propensity modeling can be thought of as very complex behavioral segmentation.
How propensity modeling provides accessible and advanced behavioral segmentation
Again, at its most base level, propensity modeling analyzes sequences of events (or behaviors) to form predictions about future user actions.
As opposed to the more common forms of Segmentation like demographic segmentation or geographic segmentation, which either group individuals together by shared demographic attributes (e.g. age, gender, etc.) or by location and proximity, behavioral segmentation draws connections between users who perform similar actions, which is a much stronger predictor of future activity.
*To be clear, demographic and geographic segmentation are more common because they're typically easier to do, not because they're more effective. (You can usually pull location and demographic information directly from your first party user data. On the other hand, mapping interactions from hundreds or thousands of users across complex webs of hundreds of events hasn't really been possible in mainstream marketing tools until recently.)
Propensity modeling for marketing segmentation: An example
For this example, let's look at an eCommerce app with almost 13 million monthly active users (MAUs).
Instead of setting up campaigns and triggers around a handful of static events, this app provides us with second-by-second measurements of all their users doing one or more of 363 different app events (basically their event stream).
This may sound intimidating, but it's actually pretty simple to set up this data feed. Most customers are up and running in less than an hour without having to involve engineering resources. If you're interested in having us run this for you, please reach out.
We trained a propensity model to predict next-month retention influence of each of those 363 events. (Next-month retention means we pick a day and user and look at what events that user did on that day. We then skip 30 days later, and start monitoring what the user did after that. The question we’re asking the model is, “how do the things a user does on a particular day influence the chances we’ll see that user still on the app 30-60 days later?”)
Note: We typically ingest historical data, so you can start utilizing 90 days worth of insight on day 1.
The model then predicts the probability that doing an event will influence next-month retention.
If the probability is 50% that means the event didn’t really have an influence - it’s a coin toss. If it’s below 50%, it means doing that thing reduces retention (increases churn). If the probability is above 50%, it means that users who do those things stick around longer — those events correspond to reduced churn.
Here's what actual data from this exercise looks like:
This is the top 20 and bottom 20 app events with their retention propensities. Darker blue means the event corresponds to increased retention, while darker red means it corresponds to increased churn.
We also calculated propensities for four different kinds of users —
- Those who are on their very first day on the app
- Those who have been users for a week
- Those who have been users for a month
- Those who have been users for more than a month.
Notice a few things:
- Many events that have a high impact on retention for brand new users have a very different impact for returning users (in fact, many of the strongest retention propensities for new users were actually moderately strong churn propensities for users who’d been on the app more than a month).
- The 8th event from the bottom (right-hand list) is actually a placeholder for the user simply not showing up at all. That’s only 8th from the bottom. There are things users can do on the app that actually have a higher impact on churn than not showing up at all.
- Some of the churn propensities suggest ways you might want to segment your users from the very start. For example, using a gift card on your first day on the app has a strong relationship with churn. That suggests there are some users who come to the app purely because they’ve been given a financial incentive to do so. Counting those users in your overall churn numbers might not make sense - you might want to track those users separately and message them differently.
- Submitting a rating on the first day - and actually, submitting a rating no matter how long you’ve been on the app, indicates a strong propensity for retention. When was the last time you sent a messaging push to encourage users to submit ratings?
- Only one of the events in the top 20 has to do with making a purchase and it was number 20, and it wasn’t the actual completing of the purchase, but a particular optional step users could take in the step out process. It often doesn’t make sense to push-push-push users to buy early and often. That can push them away. There are other things you can encourage them to do.
The most influential event for first-day users was to click into a loyalty savings program that incentivizes users to stick around for a month. For first-week users, it was reviewing an order (so, yes, actually buying things does impact retention). For first-month users, it was reviewing a schedule page that listed upcoming events. And for users more than a month on the app, is was starting a review - not completing a review, just starting it.
Propensities help you focus on the most significant events
If you work in CRM, then you’re likely already familiar with the concept of triggers, and you probably already have several user journeys set up which trigger on different events (An example being a journey that triggers on an abandoned cart event which sends a user a message (or 12) after they leave a product in their cart without making a purchase.)
Propensities tied to concrete app events, as pictured above, give you much more trigger options. For example, what messaging could you send those first-day gift card users that might encourage them to stick around? You can also trigger messages based on the lack of a high-propensity event within a certain timeframe. For example, you can trigger messages for new users who haven’t clicked on any of your loyalty program materials.
Triggering on propensities leads to significantly higher CRM effectiveness
Most segmentation and triggering activities are based on assumptions:
- If a customer sees all the features of my app, they'll stick around.
- If a customer does event X, they're most likely to convert.
- If a user hasn't done event X in Y timeframe, they're likely to churn.
...but propensity modeling shows us that these are all overly simplistic assumptions.
In reality, there are hundreds of thousands of different patterns of interactions users perform with all of your different events that are indicators of many different outcomes. The beautiful thing is that we don't have to gloss over this complexity anymore. We can embrace it for what it is.
What propensity triggering looks like in practice:
"Retention" or "user churn" is a spectrum
Here's what I mean:
If a user is 99% likely to churn, is it really worth trying to retain them?
No, right? What's the point?
They're about to leave, so our attention would be better spent investing in another user...say, one that's only 65% likely to churn. We have a much better chance of saving that user because they only have one foot out of the door.
Similarly, if a user is 99% likely to make a purchase, does it really make sense to send them a big discount?
Absolutely not! They're about to make a purchase. We'd just be throwing money away!
It makes much more sense to use our discounts strategically on users who are 60-80% likely to complete a purchase.
(It also doesn't make sense to send a discount to a user who's only 10-20% likely to make a purchase. These users would benefit from more educational information....or you could hold off on any messaging until their propensity for completing a purchase improves, to reduce the chances of driving them away from over-messaging.)
How to set up a propensity trigger in Aampe:
You can create a propensity trigger in Aampe in just a couple of clicks:
Then you create your messages as you normally would and add this trigger to your message:
Now, when a user enters this category based on their calculated propensity, they become eligible for this messaging (and, once they exit this probability, they are no longer eligible).
So, are your triggers based on static events or on your individual user's propensities to complete them?
Check out Aampe and get started with propensity-powered triggers today!
Cover image credit: rawpixel on Freepik
.png)
