Notes from Inc42Plus Maker’s Summit 2021–Fix The Leaky Bucket: Taming The Retention Beast
Published in
5 min readMar 24, 2021
You can find the original video on Inc42’s website.
Outline
- How to focus on and identify right metrics when working on user retention
- How to run quick a/b tests on multiple user groups to drive user retention
- Importance of UX in User retention
Why should I care about retention?
- Affects everything growth related
A Higher retention means:
- Faster user acquisition: More new users retained, faster.
- More monetization: The ability to spend more on paid acquisition.
- Faster recovery of marketing spend.
A higher retention results in faster user acquistion, more monetization and a faster recovery of marketing spending.
Defining your retention metric
- What is the frequency of the problem/need your service solves? = Problem frequency
- What is the event in your service that indicates that your service solved that problem? = Value event
Retention metric = Problem Frequency + Value Event
Problem frequency and value event
- Ask your users
- Look at your data
- Use your intuition/judgment
- Each use case has a different value event and problem frequency
E.g. 1:
- Service: Dunzo
- Problem: Ran out of vegetables and fruits.
- Problem Frequency: Once/twice a week
- Value event: Completed a vegetables and fruits order (At what point in the service do you feel like we’ve delivered the value to you?)
E.g. 2:
- Service: Instagram
- Problem: I want to know what my friends are up to
- Problem Frequency: Multiple times a day
- Value event: Viewed a story/post
Validating value event
- If you are confused about what is your value event, you can check data to see the cohort of new users who did that event and see which cohort has a higher retention
- This will be a balancing act of ensuring a large enough cohort of new users do this and the step jump in retention you get as you go deeper in the funnel.
- Kindle App: Is the value a user buying a book, or is it a user completed a book or user completing a book and rating it 4+?
- If the delta is large, pick the event that has a higher retention.
- If the delta is small, pick the event that is done by more users.
Validating Problem Frequency
E.g.
- The frequency with which Dunzo users are coming back to the platform
- Users seem to be coming back after a week — this seems to be the problem frequency as according to Fig 1.
Defining your retention metric
- Guided by your problem frequency and value event
- Daily (problem frequency guided) active (users who perform the value event) users
Analyzing Retention
What is good retention?
- Some portions of the users are staying on the app “forever”
- Too many users shouldn’t be dropping off
- The number depends on the business model
- What is the retention I need to be able to make my business work?
- E.g. Adding 1000 users per month with 45% M12 retention and 8% M24 retention
- Check for the deviation from the average retention — can give you outliers
- Very high positive deviations are good, very high negative deviations are poor = analyze and understand why either of these happened
- Is a full refund a sustainable user retention strategy? → Depends on the userbase. Can A/B test this. Need to test the impact on retention and whether it’s financially feasible over the long run.
Improving retention
- Use user attributes to identify where the skew in retention is happening from
- User demographics like age, gender, have kids/no kids, have pets/no pets, etc.
- Look at the existing competition and look at unsolved problems — ask yourself: why aren’t people solving these problems? Can I solve these differently? (Require a differentiation factor)
Ask yourself, does this meaningfully change the way the user uses my product?
Some questions to ask while considering user attributes:
- All things equal, can a user’s income affect the way someone uses a delivery app?
- Does this person value convenience or money more?
- Does the person’s attributes impact the frequency they order?
- Device (iOS/Android) — a proxy for income or a specific device version can be performing poorly
- Geography — a proxy for income/service quality/user attributes
- Product category — healthy food orders vs pizza food orders vs Chinese food orders
- Product feature usage — first order through search vs through categories; first order placed via COD vs UPI
- Once you have an understanding of what is causing better retention, you can run experiments to skew towards attributes that are known to give you more retention
- E.g. if you have a better % of M12/signups with users who are living in apartments with rent > 25k/month. Then running branding activities in similar apartments is a valid experiment.
- E.g. if say having a pay later option results in a higher % of M12/Activations then increasing the number of users who have a pay later option is a valid experiment.
Difficulties in measuring retention
- Running a long-term experiment on retention is tough — have to wait for a long time and jumps in retention are very small.
- Require a proxy for retention — if this happens, then retention is also going to happen.
- E.g. of a retention proxy: Users have a higher likelihood of retaining when they use the product a certain number of times (X) within a certain period (Y) of signing up.
- Can discover X and Y using a correlation analysis + can take an accurate guess based on a problem frequency
- Fig 2 shows hypotheses with the correlation. Can also use other metrics such as Positive Predictive Value (PPV) and Negative Predictive Value (NPV).
- PPV = (This event happened AND user retained)/(This event happened)
- NPV = (This event did not happen AND user did not retain)/(This event did not happen)