By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
CHAPTERS
- 0:00 – 3:30
Why Retention Matters and the Seven Mechanisms
Bryan Kim asks what retention lessons from the first consumer wave can apply to AI products. The guest explains why retention is hard to move, recalling a one-point D30 gain at Snap that earned a party, and estimates that combining these methods can lift D30 by ten to fifteen points.
- •Retention is the bread and butter of consumer products; without it you get a 'burned forest' with no new acquisition left
- •At Snap, a one percent gain in D30 retention was cause for celebration
- •The guest estimates these methods can move D30 up ten to fifteen percent, from a good 20% to an exceptional 35%
- •The guest has seen seven mechanisms among founders over the past two to two and a half years
- •An AI-native founder reported weekly retention shot up after applying them
- 3:30 – 6:15
Speed to Core Product Value
The guest says onboarding matters because a user who doesn't find value on day zero is very unlikely to be recovered. Delivering core value quickly cuts friction, as Google's focus on search speed showed, and in AI better models return faster, more relevant results that improve retention.
- •It is very hard to recover from low D0 and D1 retention
- •Google, under Larry Page, treated search speed as critical and built load speed into search ranking
- •A companionship product tracked cohorts on GPT-1, GPT-2 and GPT-3, and retention out to day 90 improved with model quality
- •Aravind at Perplexity incorporates each state-of-the-art model on the day it launches
- 6:15 – 10:10
Feature-Gated Onboarding
Gating product usage behind an unlock step lowers conversion into the product but raises retention among those who do activate. The guest cites Lapse, which required inviting five friends, and Viggle, which required completing steps in Discord before creating.
- •Gating trades lower initial conversion for higher retention, since users unwilling to do the work probably wouldn't stay anyway
- •Lapse required a lengthy onboarding and five invited friends, force-building a friend graph
- •Viggle.ai makes users read rules and learn the Discord sequence before creating
- •Having made a first creation, users are more likely to return
- •Bryan Kim frames it as building a product good enough that people will sit through the friction
- 10:10 – 12:13
Designing Reciprocity
The guest calls reciprocity, or 'give to get', probably the highest retention gain, though he has no AI example yet. BeReal shows how requiring contribution before consumption works, and he suggests AI products could gate content consumption behind creation.
- •BeReal doesn't let you see friends' moments until you share your own
- •Contributing earns reactions that give a jolt of joy or dopamine, encouraging more contribution
- •This produces richer content on the platform, feeding a loop
- •AI products with rich creation flows could gate the feed behind creation, though it won't suit every product
- 12:13 – 14:58
Building Smart Notifications
Notifications can bring users back but are double-edged, since users rarely turn them back on once off. The guest breaks them into perceived sender, payload and frequency, with examples from Snapchat, Partiful and Perplexity.
- •Three factors: who the perceived sender is, what the payload is, and the frequency
- •A Snap notification feels like a friend wanting to talk, while a happy Earth Day message feels like a company
- •Partiful messages about venue or time changes feel like they come from the host
- •Perplexity notifications deliver relevant news, such as during the OpenAI saga
- 14:58 – 17:14
Customizing Notifications and Character AI's Companion Toggle
Asked about letting consumers customize notifications, the guest says most products, especially early AI ones, needn't offer granular controls and should focus on delivering value. He points to Character AI's toggle for being contacted first by a companion as a nuanced emotional choice.
- •Commerce and social products may offer granular choices, such as only order updates
- •Granular settings are probably less relevant for AI products right now
- •Character AI let users choose whether the companion reaches out first
- •The toggle asks whether you see the being as a friend you'd want to hear from proactively
- •Bryan Kim describes this as coding boundaries
- 17:14 – 19:56
Keeping Streaks Alive
The guest argues that well-done streaks are not mere gamification but effortful, meaningful signals, using Snapchat streaks as a costly friendship badge. He sees them fitting AI products where habit helps users, such as journaling and language learning.
- •A Snapchat streak required sending a video or image each day, and the streak appears only after three days
- •A 1,000-day streak is an expensive friendship badge, unlike a like or retweet
- •AI journaling apps like Rosebud and Jot, and Speak AI for language learning, use streaks
- •Streaks suit products where habituation benefits the user
- •Bryan Kim adds that streaks are identity-based, as with long-running Duolingo streaks
- 19:56 – 22:17
Wraps and AI Summaries
The guest says wraps, popularized by Spotify, fit LLMs because they are good at summarizing behavior. Users who anticipate a summary keep engaging, and AI makes insights possible without a giant dataset.
- •LLMs can sift through behavior and engagement to produce summaries
- •Anticipating a wrap motivates continued use
- •Oops Finance can surface spending themes, unlike an old 20-page American Express statement
- •Bryan Kim suggests finance wraps such as being in the top five percent of savers
- •Airbuds shows top-one-percent listener status, and AI removes the need for the Spotify dataset
- 22:17 – 24:49
Status for Power Users
The guest says retention should be looked at granularly because power users are far more valuable than once-a-month users, though both count as one MAU. Granting status, as Reddit, Strava, Civit AI and Readwise do, encourages them to keep contributing.
- •A user active every day and one active once are both one MAU, though one is far more valuable
- •Reddit moderators and Strava's Local Legend are status honors people want to keep
- •Civit AI uses status and leaderboards to encourage model creators to return
- •Readwise shows top-percent reader rankings, which Bryan Kim relates to identity
- •The guest cites Eugene Wei's article 'Status as a Service'
- 24:49 – 26:27
Closing Takeaway: Low-Cost Mechanisms
Bryan Kim notes that all seven mechanisms are free to implement beyond engineering time and can be baked into almost any product. The guest agrees and says founders should pick what best strengthens their core product value, and both hope to see a reciprocity example.
- •None of the seven mechanisms requires a large budget
- •Strava Local Legends would rather hold status than be paid
- •Both look forward to an AI example of designing reciprocity
- •Founders should pick the mechanisms that boost their core product value
