Funnel analysis has become the cornerstone of event-based analytics. Most funnel analysis is done by setting up a predetermined sequence of events and then tracking conversion rates at each step. On Amplitude, by using our Pathfinder and Behavioral Cohorts features, you can easily identify and set up insightful funnels. This article teaches you how to set up a funnel, analyze data, and take action on the results to improve your funnel conversion rates.
Table of Contents
- Step 1: How to Identify Funnel Events
- Step 2: Set up the Funnel
- Step 3: Acting upon the Funnel Insights
- What We Learned
Step 1: How to Identify Funnel Events
Funnels should track the flow of users along a critical path on your platform. Each funnel step is a distinct action that a user can take on your platform along a flow that you want to analyze.
For example, if you were measuring an onboarding funnel, you may track the following events:
Create Account → SignUpPage1 → SignupPage2 → Registration Complete
If you were an e-commerce platform and wanted to track a purchase funnel, you may track the following sequence of events:
View Item → Add to Cart → Checkout → Purchase Confirmation
These examples work great if you know the paths users take, but it is impossible to know every possible navigation route within your platform. What if you are missing other navigation routes other than the predefined paths you have created? You may miss some really important flow of events as they may be counterintuitive at first.
Our Pathfinder feature helps you discover alternate navigation routes by showing you the most common paths of events after a chosen start action or before a chosen end action. The former is called a "starting with" flow and the latter is called an "ending with" flow.
Let’s take the example of the Amplitude demo app, a music streaming app with a social component. We will walk through an example of how to put your product hat on and analyze funnels to discover insights that may not have been intuitive at first.
For example, let’s say that after using Compass, you find that ‘AddFriend’ is the event most correlated with retention, your most important KPI, or let’s say that a key focus area is to identify how you can get users to add more friends. Given either of these two situations, the next step is to understand different paths users take to add friends on your platform so that you can encourage that behavior. In order to understand what actions users are doing immediately before they did ‘AddFriend’, we analyze the following "ending with" flow:
The graph shows us that 'PlaySong' is the most common event performed immediately before doing the event "AddFriend." Let’s assume that the other events are also pretty much in line with what we expected as the top actions except the last one; we were not expecting 'AddToList' to be one of the top events immediately preceding 'AddFriend.' We should investigate this further!
Next, we flip the analysis. What are some events users are doing immediately after they add a friend? To answer this question, we build a "starting with" flow:
It is intriguing that 'AddFriend,' 'PlaySong,' and 'AddToList' seem to be happening so close to each other since they are parts of different overall path flows for our app. Can we get more users to perform "AddFriend" if we can get them to play or favorite more songs?
Now we know what events to analyze for our funnel.
Step 2: Set up the Funnel
While setting up the funnel, there are two important questions we need to ask ourselves.
Question 1: "This Order" Mode or "Any Order" Mode?
Do we care about the sequence of events for our funnel or just the fact that someone did all the events within a specified time frame? In "Any Order" Mode, the user can perform the funnel steps in any order and so is still counted as converted even if they did not follow the exact sequence of events in the funnel. We can set our conversion window of the funnel in "Any Order" Mode to be anywhere from 1 day to 90 days.
In "This Order" Mode, users have to follow the exact sequence of events to be counted as converted. We can also go down to a funnel conversion window of one second!
Read more about "This Order" and "Any Order" here.
Question 2: What is the conversion window for the funnel?
This is where we need a good understanding of your platform. Based on our knowledge of the platform and the flow of events we want to analyze, we can select a conversion window for the funnel to be anywhere from 1 second to 90 days.
Let’s say we have been sending push notifications to certain users to play songs. However, now that we know that a lot of people do 'AddFriend,' 'PlaySong,' and ‘AddToList’ close to each other, we want to understand whether sending push notifications that encourage users to play songs makes those users more likely to also add more friends. If yes, then that would be a great insight and a validation that push notifications are having a complementary effect on other major KPIs in addition to more songs being played.
Read more about the semantics of our funnel conversion window here.
Building the Funnel
For our first funnel, we make an "Any Order" Funnel, since we are only concerned with whether the user did all of the steps and not the sequence in which the user did the steps. We are also unsure if more people perform "AddFriend" or "AddToList" first, and we want to capture both sets of users to get a high-level understanding of whether we have a cluster of users who do all four of these events together. In other words, for the first funnel, we just want to get a sense of whether this hypothesis is worth looking into further or not.
For our conversion window, we set it to one day because that is an appropriate amount of time to give someone to do the following four actions:
Receive Push Notifications → Play Song→ Add To List → Add a Friend
The funnel set up looks like:
The below image shows our funnel results.
We see a 78.6% conversion rate for the funnel, which means that 78.8% of all users who entered the funnel did all four actions within one day. At a high level, this seems like a strong conversion rate and interesting enough to further analyze, but it does not tell us anything actionable on its own.
Next, we compare two different of users for this funnel. We compare users who got a push notification and then played a song to users who did not play a song. Our goal is to see if we can get users to increase their network of friends quickly if we can get them to play more songs from push notifications.
To get that list of users, we build a second funnel with the following setup.
With this second funnel, we want a list of users who received a push notification and then played a song within 5 minutes of receiving the push notification. Since we care about the order of events in this case, we will compute this funnel in "This Order" Mode.
The below image shows our funnel results.
To create the two cohorts described previously (users who played a song and users who did not play a song after receiving a push notification), we use Microscope to create a cohort of users who converted from this funnel and name the cohort “Played song from Push”.
Now, let’s revert back to our initial funnel and segment the funnel to compare the two different cohorts of users; those in the “PlaySong from push” cohort (blue segment) and those not in the cohort (green segment).
Voilà! The results are indeed very interesting.
75.5% of the users who played a song within five minutes of receiving a push notification also did both 'AddToList' and 'AddFriend' within one day. This is 119% more than users who did not play songs from push notifications.
In other words, a user is 2.19x more likely to do 'AddFriend' if the user receives a push notification and plays at least one song after receiving the notification.
Funnel Details
Next, we will check to see if there are particular trends on conversion over time and see if there are some critical insights in the property distribution of the 'PlaySong' event. Looking at the 'SongSource' event property for 'PlaySong,' it seems that the contribution of 'OnDemand' and 'Radio' is very similar.
We also see trends in conversion over time seem to stay consistent for both cohorts (the graph below is following the same color pattern as above, so blue color segment represents users in our cohort and green represents those not in the cohort).
Overall, we know that users who receive push notifications and play a song within five minutes are three times more likely to add friends, but there are no critical insights from Property Source distribution (for the type of song they play) or conversion over time.
Step 3: Acting upon the Funnel Insights
Our next goal is to create a retention loop where we can get more users to come back and do the desired action -- add more friends -- which is correlated with retention, our main KPI.
We go back to our funnel and use the Microscope feature to make cohorts of users who dropped off from our funnel at critical points like 'AddToList' and 'AddFriend' and message them to take the corresponding actions.
You can use an in-house messaging platform or one of our push notification partners like Urban Airship or Kahuna to message these cohorts.
What We Learned
To summarize, we landed on a very critical insight to possibly boost our most important KPI by using the following framework:
- Exploring Pathfinder to analyze patterns around how users reach the critical event ('AddFriend') and what they do immediately before and after the critical event.
- Testing our hypothesis in funnel analyses by building funnels with the appropriate conversion mode and conversion window.
- Acting upon the funnel results to encourage the right actions using Microscope cohorts.
We hope this article helps you understand how to get the most out of funnel analysis on Amplitude. Funnels can provide some incredible insights if you use them in conjunction with other features like Pathfinder, Behavioral Cohorts, and the Compass to get the maximum value.