Finding hidden growth opportunities: lessons from Duolingo, Grammarly, and Chess.com

Executive overview

Most growth teams either scatter across too many ideas or grind the same metric until they hit a wall. The explore-exploit framework fixes this by separating the search for the right opportunity from the work of maximising it.

Albert Cheng has led growth and monetisation at Duolingo, Grammarly, and Chess.com. His core principle: growth's job is to connect users to the value of your product — not to hack metrics. Every tactic, experiment, and team structure should serve that goal.

Surfacing a single human-psychology insight and spreading it across the whole product can drive more growth than running dozens of unrelated experiments.

The explore-exploit framework

• Explore = finding the right mountain to climb; divergent, creative, hypothesis-generating.
• Exploit = climbing that mountain efficiently; focused, iterative, scaling what works.
• Too much exploration leaves teams scattered with no through-line.
• Too much exploitation leads to local maxima and stagnation.
• Operate at the insight level, not just the strategic level: one experiment win should immediately prompt adjacent teams to test variants of the same underlying hypothesis.
• Signal to re-enter explore mode: experiments stop reaching statistical significance — the area has been squeezed.
• Chess.com example: discovering that 80% of game reviews happen after a win (not a loss) led to showing players their best moves after losses instead of their blunders. Result: game reviews +25%, subscriptions +20%, retention meaningfully up.

Consumer subscription fundamentals

• User retention is the foundation; without it, you must convert users to paid on day one — an extremely hard model.
• D1 retention of 30–40% is a reasonable baseline for a consumer app.
• At maturity, roughly 80% of daily or weekly actives are existing or resurrected users — new users are a minority of your active base.
• Investing in the resurrected-user experience compounds over time as dormant user pools grow.
• Word-of-mouth and organic growth are almost prerequisite for consumer subscription at scale; paid acquisition without retention is a treadmill.
• Freemium works best when the free tier reflects the full power of the product — a limited taste of paid features converts better than keeping them invisible.

The Grammarly monetisation win

• Most free users experienced Grammarly as a spelling-and-grammar tool only, because those were the free suggestions shown.
• Instrumentation revealed that very few free users accepted all suggestions; most picked and chose, rarely hitting the paywall.
• Intervention: intersperse a limited sample of paid-tier suggestions into the free experience.
• Concern that giving away more would reduce upgrade intent proved wrong — upgrade rates nearly doubled.
• Mental model: treat free tier as a real-time, in-context reverse trial rather than a time-limited one.
• Applies broadly: if your free product doesn't show users what the paid product can do, they have no reason to pay.

Experimentation at scale

• Start somewhere — 40% of product teams run no experiments at all; even a basic A/B test in a third-party tool beats nothing.
• Consumer behaviour is unpredictable; domain expertise can be a crutch; running experiments surfaces what actually happens versus what you'd predict.
• Typical win rate for experiments is 30–50%; losing experiments are equally valuable inputs.
• The system matters more than any single experiment: growth model, instrumentation, tooling, and cross-team visibility are prerequisites.
• Third-party tools (StatSig, etc.) are the right starting point; building in-house only makes sense at significant scale.
• Chess.com went from zero experiments pre-2023 to ~50 in one year to ~250 the next, targeting 1,000.
• The 1,000-experiment goal is a forcing function: it surfaces what would need to change — no-code configuration, lifecycle marketing experiments, App Store tests, engineering-enabled templates — not just a vanity metric.
• Culture shifts through visible wins, not mandates; teams need to see the learning compound before they believe in the system.

AI in growth work

• Text-to-SQL Slack bots that answer ad hoc data questions democratise data access and dramatically increase the volume of questions asked — removing friction also removes the social cost of "dumb" questions.
• AI-powered prototyping (V0, Lovable, Figma Make) shortens the explore phase: visualising a bold idea before writing a full spec makes it discussable and testable faster.
• The bottleneck is bridging tinkering to production workflow; interoperability between tools across functions is still an unsolved problem at mid-size companies.
• Using LLMs to summarise experiment analyses and generate next-step hypotheses accelerates the exploit cycle.
• Chess engines (Stockfish) and LLMs serve different roles in chess products — engines for deep move evaluation, LLMs for natural-language explanation and personality.

Brand, community, and product growth

• Brand and growth are not in competition; brand can be rocket fuel.
• Tracking "how did you hear about us?" at Duolingo revealed that viral TikTok and social moments could drive 20–30% of new users on a given day.
• Growth experimentation is the steady compounding; brand moments are the waves — you must be positioned to capitalise on both.
• Chess.com growth accelerated dramatically from external cultural events (pandemic, Queen's Gambit, streaming) combined with product readiness — neither alone would have had the same effect.
• Virality is hard to manufacture; find where users are already organically sharing (screenshot tracking at Duolingo), then staff those moments with illustrators and animators to make them 5–10x more shareable.

Habit formation and new-user experience

• Gamification has three pillars: core loop (lesson → reward → streak → notification), metagame (leaderboard, achievements, long-term goals), and profile (visible investment that creates switching cost).
• Getting the core loop tight is the prerequisite; metagame and profile layer on top.
• At Chess.com, fewer than one-third of new beginners win their first game; losing reduces retention by ~10% — small at the individual level, significant at scale.
• Interventions for new users: dedicated learn-to-play flows, hiding ratings for the first few games, playing against bots or coaches before live opponents.
• The instinct to protect new users from early failure is counterintuitive but data-supported across multiple products.

How Duolingo, Grammarly, and Chess.com differ

• Duolingo: extremely high clock speed, rigid process (10–15 minute product reviews, templates for every step), creative ideas inside a structured system. Superpower is motivation and habit mechanics, with language learning as the vehicle.
• Grammarly: freemium to enterprise motion, B2C to B2B conversion embedded in the product, deep integration across applications as the core moat. Current-user retention is most influenced by core product suggestion quality, not growth team tactics.
• Chess.com: mission-first culture where everyone uses the product daily, globally remote, chess-obsessed. Growth culture had to be introduced deliberately; the foundation was community and passion, not process.

Building high-performance teams

• High agency matters more than domain experience, especially as AI shifts the underlying skill set.
• Look for clock speed, energy, and the ability to discard learned habits — in a fast-moving environment, experience can be a crutch.
• Signals appear outside the interview: how candidates set up the meeting, whether they've deeply used the product, the quality of their questions.
• Work trials and reference checks surface agency better than structured rubric interviews.
• Company stage fit is real: big companies offer scale and best practices but move slowly; tiny startups move fast but growth is gruelling; mid-size (500–1,000 people, 10–20 years old, profitable) offers both contribution at scale and daily execution pace.

Key failure: Chariot Direct

• Dynamic routing add-on for a commuter shuttle service — a solution searching for a problem, not rooted in a specific user need.
• Focused entirely on the rider app without adequately considering driver experience and operations; when drivers were confused, the whole product suffered.
• Launched PR before validating customer demand, creating sunk cost pressure to continue even without evidence it worked.
• Lessons: start from the problem not the solution; consider all sides of a marketplace; validate before you announce.