Rethinking Bumble's subscription churn through a safe peer-feedback system
Problem + AI
Only 5% of Bumble's 50M users pay, and I questioned the brief's proximity fix from the start.
The original design brief framed this as a timing and proximity problem: if users knew they had a match physically nearby, they'd be more motivated to convert. The proposed solution was a precise location notification feature across mobile, desktop, and watch.
I was given a design brief at the beginning of this project that tried to solve this problem. It assumed users would convert if they knew a match was physically nearby, and proposed a precise location notification to act on that. A clear hypothesis, but an untested one.
Churn data showed proximity had nothing to do with it, so the focus moved to value.
When I analyzed subscription churn data, there was no evidence that nearby matches led to stronger conversion or retention. Instead, a prevailing trend was that cancellations consistently followed the same pattern: users paid, saw little change in their outcomes, and left. That shifted the focus from proximity to value.
I ran synthetic user interviews to pressure-test why people were really leaving.
I created synthetic users representing key Bumble personas and ran structured interview simulations to surface patterns. This allowed me to rapidly pressure-test alternative explanations, and I never treated it as a source of truth. Patterns only counted when real behavior confirmed them.
Across every segment, users believed premium buys visibility while real results stay flat.
They'd pay once, feel nothing changed, and cancel. Fixing churn meant changing what users believed a subscription could do, not adding another feature.
Problem + AI
Only 5% of Bumble's 50M users pay, and I questioned the brief's proximity fix from the start.
The original design brief framed this as a timing and proximity problem: if users knew they had a match physically nearby, they'd be more motivated to convert. The proposed solution was a precise location notification feature across mobile, desktop, and watch.
I was given a design brief at the beginning of this project that tried to solve this problem. It assumed users would convert if they knew a match was physically nearby, and proposed a precise location notification to act on that. A clear hypothesis, but an untested one.
Churn data showed proximity had nothing to do with it, so the focus moved to value.
When I analyzed subscription churn data, there was no evidence that nearby matches led to stronger conversion or retention. Instead, a prevailing trend was that cancellations consistently followed the same pattern: users paid, saw little change in their outcomes, and left. That shifted the focus from proximity to value.
I ran synthetic user interviews to pressure-test why people were really leaving.
I created synthetic users representing key Bumble personas and ran structured interview simulations to surface patterns. This allowed me to rapidly pressure-test alternative explanations, and I never treated it as a source of truth. Patterns only counted when real behavior confirmed them.
I didn't believe a subscription would change my results.
Across every segment, users believed premium buys visibility while real results stay flat.
They'd pay once, feel nothing changed, and cancel. Fixing churn meant changing what users believed a subscription could do, not adding another feature.
Concept + AI
My answer was Peer Profile Reviews: honest feedback on the profiles people swipe through.
After identifying the root cause of Bumble's churn problem, I proposed a feature that let users give feedback on the profiles they come across, so people learn what's actually landing and what isn't.
Open feedback invited abuse, so I used AI to challenge my thinking instead of solving it for me.
During critique, a major risk surfaced.
People can say anything they want online. Without guardrails, an open text field could quickly become a channel for insults, harassment, or personal attacks. What was intended as helpful feedback could easily leave users feeling judged instead.
My first instinct was preset tags. Safe and easy to moderate, but felt too limiting. Users confirmed it: when they want to say what they actually mean, tags get in the way.
That left me stuck between flexible-but-unsafe and safe-but-rigid. To break the deadlock, I used AI as a meta-cognitive thinking partner. Instead of asking it for solutions, I asked it to challenge me: “What's the hardest part here?” “Does this really stand out?”
The safety-versus-flexibility trade-off was never real, so I kept open text and built safety around it.
That process surfaced the real assumption I was making. I kept the open text users wanted and layered safety in around it, some of it powered by AI capability matchmaking.
Concept + AI
My answer was Peer Profile Reviews: honest feedback on the profiles people swipe through.
After identifying the root cause of Bumble's churn problem, I proposed a feature that let users give feedback on the profiles they come across, so people learn what's actually landing and what isn't.
Honest feedback
without the abuse.
Open feedback invited abuse, so I used AI to challenge my thinking instead of solving it for me.
During critique, a major risk surfaced.
People can say anything they want online. Without guardrails, an open text field could quickly become a channel for insults, harassment, or personal attacks. What was intended as helpful feedback could easily leave users feeling judged instead.
My first instinct was preset tags. Safe and easy to moderate, but felt too limiting. Users confirmed it: when they want to say what they actually mean, tags get in the way.
That left me stuck between flexible-but-unsafe and safe-but-rigid. To break the deadlock, I used AI as a meta-cognitive thinking partner. Instead of asking it for solutions, I asked it to challenge me: “What's the hardest part here?” “Does this really stand out?”
The safety-versus-flexibility trade-off was never real, so I kept open text and built safety around it.
That process surfaced the real assumption I was making. I kept the open text users wanted and layered safety in around it, some of it powered by AI capability matchmaking.
Research + AI
Talking to dating app users revealed their real fear was language: what gets written and how it reads.
What kept coming up wasn't the feedback itself — it was who was behind it and how it was worded. Being judged, getting a cruel comment, having no way to push back.
Every fear users named was essentially a language problem: what gets written and how it reads. Fixing that meant acting on the words themselves, in the moment, for every comment. AI was the only guardrail that could sit between the comment and the reader without removing the open text users wanted.
I ran a matchmaking activity to pair each fear with something AI could realistically do.
I listed the moments I got from them that would make them feel unsafe, and together we matched it to something AI could realistically do about it.
A few patterns came out, each pointing to a guardrail:
- Users want to choose what they're judged on, not be evaluated wholesale
- Due to the nature of the platform, they trust the space more when there's a social agreement about behavior
- They want raw, blunt comments softened before those comments ever reach them
- They want the final say over what lands on their profile
Users pointed out that opting in, picking categories, and reading a contract was a lot of work before they got anything back. So I built a reason to engage into every step.
See all these features in the next section!
Research + AI
Talking to dating app users revealed their real fear was language: what gets written and how it reads.
What kept coming up wasn't the feedback itself — it was who was behind it and how it was worded. Being judged, getting a cruel comment, having no way to push back.
Every fear users named was essentially a language problem: what gets written and how it reads. Fixing that meant acting on the words themselves, in the moment, for every comment. AI was the only guardrail that could sit between the comment and the reader without removing the open text users wanted.
I ran a matchmaking activity to pair each fear with something AI could realistically do.
I listed the moments I got from them that would make them feel unsafe, and together we matched it to something AI could realistically do about it.
A few patterns came out, each pointing to a guardrail:
- Users want to choose what they're judged on, not be evaluated wholesale
- Due to the nature of the platform, they trust the space more when there's a social agreement about behavior
- They want raw, blunt comments softened before those comments ever reach them
- They want the final say over what lands on their profile
Users pointed out that opting in, picking categories, and reading a contract was a lot of work before they got anything back. So I built a reason to engage into every step.
See all these features in the next section!
Solution + AI
Hive Insights turns honest peer feedback into safe, actionable profile improvements across three moments.
The final feature lets users give and receive real feedback on the profiles they come across, with AI working as a guardrail rather than a filter on top. It follows the journey users actually take: opting in on their own terms, giving feedback that stays constructive, and receiving insights reframed to be useful and fully in their control. Every moment maps back to a fear users named in research.