Explore, Exploit, Retain: How Albert Cheng Grows Duolingo, Grammarly and Chess.com-Style Subscription Products

Albert Cheng grew up as a serious pianist, the eldest child of Taiwanese immigrants, practicing every day under a 90-minute countdown clock on top of the family's upright piano. Perfect pitch let him tell instantly whether he was playing correctly, and he seriously considered a music conservatory before choosing engineering school instead.

Looking back, he sees two parallels between music and growth. Both rely on consistent repetition with a very tight feedback loop, where you make mistakes constantly and learn through them. Both also have a structural layer (scales and theory in music; growth models, metrics, channels and experiments in growth) and a creative layer that needs passion and flow to produce something beautiful.

Cheng picked up the explore-and-exploit framework from his engineering partner at Grammarly and notes its roots in Brian Balfour's Reforge material. Exploration is finding the right mountain to climb; exploitation is concentrating resources on climbing it. Spend too long exploring and the team feels scattershot, with no through line; spend too long exploiting and you saturate, locally maximizing the same thing.

He applies it at the level of individual insights. At Chess.com, a PM working on game review, the post-game coach that critiques your moves, found that 80% of reviews happen after a win, contrary to the team's assumption that people would use it to study losses. After a loss, the product now shows your best moves with an encouraging coach message, which grew game reviews by 25%, subscriptions by 20% and retention substantially.

The insight then gets shared widely so that, for example, the puzzles PM can audit cold patterns in their own area and make success ratings, copy and button colors more positive. Cheng says typical experiment win rates run 30-50%, and even strongly losing experiments are valuable to surface. He also notes that wins in one area tend to be reusable, because human psychology often resonates across the whole product even when companies split it among many teams.

Cheng's team runs roughly 250 experiments a year and invests in experiment explorer tools to spot patterns across hypotheses and learnings. A recent effort is a text-to-SQL Slack bot that answers ad hoc questions like how many subscribers there are in South Africa, doing the actual analysis rather than just returning a query. He also observes that tools like this unlock questions people would otherwise never ask because they feel embarrassed or don't want to bother an analyst.

On the product side, PMs use V0, designers use Figma Make, engineers use Cursor, Claude Code and GitHub Copilot, and Chess.com has built AI prototypes of core screens such as onboarding, home and the chessboard so anyone can layer ideas on top. His complaint is that the hand-off between tinkering and shipping is not yet seamless at their scale, so function hand-offs persist while they invest in design-system components and MCPs.

Grammarly is freemium, with over 90% of users on the free plan. Cheng's team first had to instrument which suggestions users saw and how often they hit paywalls, and found that only a small share of free users accepted all their suggestions; most picked and chose. At the same time, the generative AI shift meant the free product still looked like a spelling-and-grammar fixer, since correctness suggestions were all free users saw.

The team therefore sampled various paid suggestions, such as tone and clarity rewrites, and interspersed them among free suggestions as a limited taste. The worry was that giving too much away would cut subscriptions; instead people saw Grammarly as a far more powerful tool and upgrade rates nearly doubled. Cheng calls it "a reverse free trial" that works in real time while writing rather than over time.

On freemium more broadly, he leans toward a permanently free core value proposition with a sampling of premium features layered on top. That fits mission-driven companies, supports word of mouth and network effects, and even helps Grammarly's team adoption. Reverse trials suit B2B products with lock-in, he says, while ordinary free trials tend to remain the norm for consumer products.

Asked what indie developers miss about consumer subscriptions, Cheng answers that user retention is gold. Without it you must get people to pay on day one and pay to acquire them, whereas Duolingo and Chess.com grew as organic, word-of-mouth businesses in markets they expanded. He considers day-one retention of around 30-40% solid for a consumer app, though he says a lower figure makes it hard to acquire enough users mathematically to build a big daily active base.

As companies mature, he argues, the retention of the existing, habitual user base becomes the biggest lever, with the exception of products like Grammarly that you install and do not open daily, where activation and the aha moment matter most. At Chess.com, he says roughly 80% of active users are current users, with new and resurrected users making up the rest in roughly similar size.

Because mature products accumulate hundreds of millions of dormant users, the resurrection experience deserves real design attention. He cites Duolingo's social notifications when a friend starts the app and re-placement tests for returning learners who have forgotten what they knew. In his view such mechanics can deliver good ROI for a more mature company.

Cheng says Duolingo stood out for a product-development ethos spread through the whole company, documented in a playbook called the Green Machine. It hires energetic people out of college, gives them excellent experimentation tooling and obsesses over clock speed; each user's experience changes multiple times a day, and product reviews take 10 to 15 minutes. He stresses that the process is rigid but the ideas, from memes to Super Bowl ads, are anything but.

Grammarly started as a paid product for students, became freemium for everyone and then layered on a managed enterprise motion, so Cheng's job included finding the teams and companies that sales and demand gen should approach. The core product experience, meaning the frequency and quality of suggestions, drove ongoing usage, so he eventually realized his growth team was getting in the way and handed that metric to the core product leader.

Chess.com is staffed by people fanatical about chess, globally remote, with Slack full of games and staff who dogfood constantly. Cheng argues that Duolingo's real origin was motivation and habit-building, with language only the first vehicle, and that Grammarly's superpower is being integrated across countless applications. His point is that there is no single right way, because all three companies are succeeding.

Cheng used to think brand marketing and growth experimentation were opposed, but now sees them combining as rocket fuel. At Duolingo he watched the owl gain a personality through push notifications and product experience, which the marketing team then amplified on TikTok and YouTube; on some days social channels drove 20-30% of new users in the self-reported "how did you hear about us" data.

Chess.com spent about 15 years under the radar until the pandemic, The Queen's Gambit, YouTube and Twitch streamers and kids playing in school changed everything about five years ago. Cheng describes experimentation as the slow-and-steady, continual-improvement engine, while occasional big waves can quadruple registrations overnight, and you would be a fool not to be ready to capture them.

Cheng points out that chess and AI have been intertwined for nearly a century, from early computing pioneers through IBM's Deep Blue beating Garry Kasparov in 1997. He says more people play chess now than ever, and that engines like Stockfish, rated around 3,600 on the Elo scale versus roughly 2,800 for Magnus Carlsen, are dramatically stronger than any human.

Chess.com's game review runs engines behind the scenes to evaluate every move, and then LLMs translate those evaluations into approachable language and speech with a personality. Cheng says the key is to keep the customer as North Star and apply the right technology to each feature rather than using LLMs just because they are the hot new thing, noting that LLMs themselves hallucinate chess moves and are poor at deep calculation.

For growth work specifically, Cheng says AI speeds up the experiment cycle: ChatGPT can summarize another person's analysis doc and suggest ideas to try, and visual prototyping makes bold exploratory ideas far cheaper to explore. PMs are not yet shipping code to production themselves, but the time from concept to something clickable has dropped dramatically.

Cheng defines growth's job as connecting users to the value of the product, pushing back on the reputation that growth is cold metrics hacking with paywalls and friction. Value changes over time, he says: what someone who is not yet a user needs to understand differs greatly from what a three-year habitual user needs, and teams should be organized around those journey stages.

His first experimentation tip is simply to start: he cites an Atlassian report saying 40% of product teams do not experiment, and says he is wrong often enough that skipping tests leaves opportunities on the table. Use a third-party tool or have engineers spin something up, and crawl, walk, run. He used Statsig at Grammarly, while Duolingo and Chess.com built in-house tools, but he does not recommend building in-house from day one.

Before 2023 Chess.com barely experimented; last year it ran about 50, this year it is on pace for about 250, and next year's target is 1,000. Cheng admits he made up the number, and says a thousand experiments without learning would be a waste. The value is in discussing what would need to be true to hit it.

Those conversations lead to experiments beyond product and engineering: lifecycle marketing copy for pushes and emails, app store screenshots and keywords, content marketing, and no-code configurable screens like the home and pricing pages. He also wants proper observability on progress. The culture shift, he says, depended on a CEO and co-founders, Erik and Danny, who were mentally flexible and preached product-led growth themselves.

The system matters as much as any single test: start from a growth model, instrument the product thoroughly, and trust your tooling carefully. At one company with an in-house tool, retention was configured backwards, so every positive result was actually negative. He adds that wins need to be celebrated, because people only buy in when they see the metrics move.

He also describes how Duolingo's virality work took a lean-in approach: the team temporarily tracked screenshots to find where users already shared, such as streak milestones, funny challenges and leaderboard promotions. It then staffed those moments with illustrators and animators to make them more delightful, which he says can be 5X or 10X growth drivers.

On habit formation, Cheng describes a gamification model with three pillars: the core loop (a lesson, a reward, a streak and next-day notification), the metagame (Duolingo's path, leaderboards and achievements) and the profile as a reflection of your investment. At Chess.com, more than 75% of new users are beginners, fewer than a third win their first game, and losing makes retention 10% worse. The team is testing a gentler learn-to-play path for self-declared newcomers and hiding ratings for the first five games.

On teams, his counterintuitive lesson is that the best performers often have high agency and clock speed rather than deep domain experience, which can even become a crutch while AI forces people to discard learned habits. He reads those traits in the soft signals outside the interview itself: the questions candidates ask, whether they dug into the product, their references and energy. He also finds that he shines at medium-sized companies of roughly 500 to 1,000 people, 10 to 20 years old and still at inflection points.

For his failure story, Cheng recounts leading product at Chariot, the San Francisco commuter shuttle startup, where dynamic routing was a solution in search of a problem. The team also underweighted drivers and operations as users in a marketplace, and ran PR before validating demand, creating sunk-cost pressure. He closes by saying his own lessons mostly come from absorbing what others tried, like a mental sponge.