Crystal Widjaja
Chief Product Officer at Kumu, former head of growth at Gojek
Where to hear and read about Crystal Widjaja1 episode or test
Crystal Widjaja is Chief Product Officer at Kumu and the former head of growth at Gojek, where she built the data team and then ran growth. The most useful thing to learn from her is to fix one real constraint in the model you already have, using cheap manual tests and data that explains why users act.
measurements do not equate to insights
How to scrappily hire for, measure, and unlock growth | Crystal Widjaja, Gojek and Kumu
Crystal Widjaja: the story and the lessons
1Crystal Widjaja's path from bank research to Gojek
Widjaja reached tech through self-taught data work, not a technical degree. She majored in political science, and her first job was investment banking research. She owned an Excel database of 130,000 rows and over 60 columns and taught herself to build a MySQL replacement.
The bank asked what open source was and whether MySQL was expensive. She left, convinced tech would not happen there. As a first-generation American with no one pointing her to consulting, she had found that first job on Craigslist.
A cold email to Gojek's HR started five fast years. Her family is from Indonesia, so she applied her banking pattern-matching to Southeast Asian companies. She built the data team from scratch, then took on fraud and risk, then performance marketing, then growth.
Her pitch was "I'm willing to move to Indonesia, take a bet on me." Having all the data showed how much fraud was in the system, which is why fraud and risk came next.
She joined a risky startup because the product and market were obvious. Jakarta traffic can mean two hours for 20 kilometers, so a motorcycle taxi was an obvious fix. Drivers already existed, but connecting them to riders and haggling over price was painful.
The app channel was shaky because most drivers had no smartphone, but she judged it solvable. She calls it a Warren Buffett approach: understand the product and market first. Gojek then rented a stadium and signed up about 60,000 drivers in a couple of weeks. The company was working out of a house at 4,000 orders a day when she joined, and she briefly thought she had made a huge mistake.
She now leads product at Kumu and runs Generation Girl. Kumu is a social super app, mainly in the Philippines, with feeds, audio, video and multi-seat rooms. Gojek had about 170 million users and over 20 services.
She co-founded Generation Girl in Indonesia with women from Gojek. It offers free STEM classes for girls 12 to 17 and trains teachers. Several thousand students have joined, and a girl deciding engineering isn't for her counts as success because the choice is hers.
2How Widjaja tests growth ideas cheaply before building
Widjaja validates ideas by hand before engineering builds anything. Gojek added 100 random drivers to a WhatsApp group to pitch a $10 subscription to riders. Interns found the customer in the backend, issued vouchers and deducted $10 from the driver's balance. It validated the value prop and expected conversion rates with no build.
For an onboarding redesign, a designer overlaid the new screens on a screenshot of the app and sent it as an in-app message. Typeform quizzes and in-app web pages served as fake features without waiting on a mobile release.
She says 30 data points beat zero, and an untestable idea is useless. With a sample of 30, the trend generally stays the same as the sample grows to 100. Only the precision changes.
If you cannot think of a way to run an experiment on an idea, she treats it as close to useless. At 30 people every idea is cheap, so unscalable things work better than at 100.
She fixes the step right before conversion, not retention directly. For GoFood the barrier was trust in a new restaurant. Gojek used Facebook Connect to show which merchants a user's friends had bought from, and users who saw that were twice as likely to buy from a brand new restaurant.
She calls retention too vague a goal, like asking for everything orange in the world instead of everything orange on a construction site.
She shortens the longest, most painful path that enough users attempt. At Kumu, users searched for friends on signup, but search was a slow, underused API. Speeding it up and adding more content lifted conversion from 60 to 90 percent over a few weeks.
She layers these sure wins with a few bets on new features, and runs an experiment on the bets if she can.
When visitors land and do nothing, she rewrites the copy first. Those users haven't experienced the product, so the product isn't the problem. She writes to the pain point rather than the solution so users can see themselves in the use case.
She sees founders who chase a brand new feature with no Wizard of Oz test or user data as grasping at straws. The question to ask is why this instead of literally anything else.
She ties unfamiliar products to something users already know. Many Gojek users didn't know what a digital wallet was. The team put each user's virtual account number on a picture of a credit card, and users typed it into an ATM like a debit account. Top-ups rose just from sending the picture.
She reads cancellation reasons to find the small fix that stops churn. For ABN InBev's beer subscriptions, the top cancellation reason was having too much beer, but users could only cancel or resume. A pause button fixed a temporary problem without a permanent cancellation.
No new products or reactivation emails were needed. "canceling the subscription is a permanent solution to having too much beer."
She adds deliberate friction to spot users with real intent. At Kumu a real friendship is hard to see, so the app asks users to fill out a questionnaire to join a community. Those who complete it raise a hand for community, and those who skip it probably want short-form entertainment.
The app leaves breadcrumbs to trace these paths.
3How Widjaja found Gojek's growth lever in its drivers
She maps the business's physics before choosing any growth lever. Step one is to define market, product, model and channels. For Gojek that was Indonesian consumers and drivers, a mobile app, a charge per order, and channels like push notifications, Facebook ads and drivers in jackets.
Drivers on the street were marketing, and seeing them helped word of mouth as Gojek entered new cities. Step two is checking that loops and funnels fit those physics. If a loop needs four or five changes at once, it doesn't fit.
She turned drivers into a sales channel for GoPay. When a driver was allocated, a service checked whether the customer had ever topped up GoPay. If not, the driver got a message offering extra money for turning the rider's cash into wallet balance. This drove about 60% of acquisition once released.
Nothing in the product changed, only the use of an underused lever. "you wouldn't believe how great of a salesperson someone can be when you were literally trapped in a car with them."
She starts from what works and fixes the biggest constraint. She finds the constraint first, then optimizes the funnel or loop, changing one small thing at a time. In her view it is very hard to move the physics of a whole model.
She warns that launch spikes can be pulled-forward demand. When Netflix and Spotify launch in places like Southeast Asia, few people hold accepted credit cards. The week-one surge is everyone who could subscribe, and growth gets harder afterward.
She checks whether early adopters are exhausting a limited pool before assuming growth will continue.
4Widjaja's retention benchmarks and analytics setup
She expects free products to flatten near 60% weekly retention. A paid product should flatten at 20 to 30%, at scale. Friends and family retention should be near 80%, because if people who care about you won't use it, strangers won't either.
She reads retention by cohort and gives teams two to three frequency periods to see the curve level off. Gojek's early retention was 60 to 70%, which told her the product just needs to work.
She puts properties on every event so data answers why. Gojek's map-loaded event carried drivers on screen, pickup location, city, surge pricing, minimum fare and voucher code. It showed users who saw two drivers booked far less than those who saw five, which pointed to areas with only two drivers.
She can judge a team's instrumentation just by reading its tracking spec. A spec where each event has one property or none is the main warning sign.
She separates observations from insights that change what you do. "Power users book four times more" is only an observation. Learning that power users use free shipping on high-value baskets is an insight, and it changed marketing spend for power versus non-power users.
Her method is to segment users by event properties, form a hypothesis, then run a test to get a causal answer.
She starts with free analytics tools and adds heavier ones as the company grows. With one data warehouse, Google Data Studio is free and enough, and Metabase is open source for SQL-skilled teams. For in-app mobile events she picks CleverTap, since Mixpanel failed her and Amplitude lacked CRM features.
As the company scales she adds Amplitude for analytics power, Segment to pipe data, and Eppo for experimentation, which replaced dashboards she once built by hand.
5How Widjaja builds and hires a growth team
Growth at Gojek began as a crew filling gaps the core team left. The core team was busy building features like phone number masking. Growth took onboarding gaps, such as choosing which SMS provider sends the OTP for each telco.
Drivers with thousands of rides assumed every customer knew how Gojek worked, so the fix was telling the driver a customer was new. Growth PMs were later embedded in teams and became owners of product areas.
She says a separate growth team works only with insane product-market fit. When demand outruns the core team, a side team can act as the cleanup crew and plug in what was forgotten. Her team was heavy on statistics graduates who cared about numbers, odds and probabilities.
At that scale almost anything does something, so the question becomes what has the most impact and unlocks the future.
She hires the first growth person for statistics, not flash. The hire must be intuitively good at sampling, selection bias and not misreading ratios from small bases. She would hire anyone willing to take an intro statistics course, young or experienced.
She gives them a known gap to own rather than asking them to model all of growth. "the worst possible thing is to have a growth person who thinks they are doing the right thing and is measuring things wrong."
She screens growth candidates with a take-home experiment design. She presents something observed, asks how the candidate knows it is true, then has them design an experiment. Candidates get five days for about four hours of work, because live doesn't work.
She looks for random sampling and a measured approach, not build and hope. She likes hearing that candidates had to google and read white papers.
She favors quick, hacky experiments over heavy tool rollouts. Gojek's first growth experiment was a Python script on the Twilio API that texted drivers from a CSV, telling them to accept more rides. It raised acceptance by 2%, and more for brand new drivers, which led to work with driver onboarding.
She warns that onboarding many tools that don't integrate can eat six months with nothing shipped. "growth is an opportunity cost of time trade off with what you could have been doing to the product in that time."
How to scrappily hire for, measure, and unlock growth | Crystal Widjaja, Gojek and Kumu
Crystal Widjaja explains how she grew Gojek and Kumu with scrappy tests, targeted levers, and analytics built to produce insights instead of dashboards. She shares retention benchmarks, an instrumentation approach, and a hiring process for the first growth person, and notes that Gojek's driver-led GoPay cross-sell drove about 60% of acquisition.
Widjaja reached tech through self-taught data work, not a technical degree. She majored in political science, and her first job was investment banking research. She owned an Excel database of 130,000 rows and over 60 columns and taught herself to build a MySQL replacement.
The bank asked what open source was and whether MySQL was expensive. She left, convinced tech would not happen there. As a first-generation American with no one pointing her to consulting, she had found that first job on Craigslist.
A cold email to Gojek's HR started five fast years. Her family is from Indonesia, so she applied her banking pattern-matching to Southeast Asian companies. She built the data team from scratch, then took on fraud and risk, then performance marketing, then growth.
Her pitch was "I'm willing to move to Indonesia, take a bet on me." Having all the data showed how much fraud was in the system, which is why fraud and risk came next.
She joined a risky startup because the product and market were obvious. Jakarta traffic can mean two hours for 20 kilometers, so a motorcycle taxi was an obvious fix. Drivers already existed, but connecting them to riders and haggling over price was painful.
The app channel was shaky because most drivers had no smartphone, but she judged it solvable. She calls it a Warren Buffett approach: understand the product and market first. Gojek then rented a stadium and signed up about 60,000 drivers in a couple of weeks. The company was working out of a house at 4,000 orders a day when she joined, and she briefly thought she had made a huge mistake.
She now leads product at Kumu and runs Generation Girl. Kumu is a social super app, mainly in the Philippines, with feeds, audio, video and multi-seat rooms. Gojek had about 170 million users and over 20 services.
She co-founded Generation Girl in Indonesia with women from Gojek. It offers free STEM classes for girls 12 to 17 and trains teachers. Several thousand students have joined, and a girl deciding engineering isn't for her counts as success because the choice is hers.
Widjaja validates ideas by hand before engineering builds anything. Gojek added 100 random drivers to a WhatsApp group to pitch a $10 subscription to riders. Interns found the customer in the backend, issued vouchers and deducted $10 from the driver's balance. It validated the value prop and expected conversion rates with no build.
For an onboarding redesign, a designer overlaid the new screens on a screenshot of the app and sent it as an in-app message. Typeform quizzes and in-app web pages served as fake features without waiting on a mobile release.
She says 30 data points beat zero, and an untestable idea is useless. With a sample of 30, the trend generally stays the same as the sample grows to 100. Only the precision changes.
If you cannot think of a way to run an experiment on an idea, she treats it as close to useless. At 30 people every idea is cheap, so unscalable things work better than at 100.
She fixes the step right before conversion, not retention directly. For GoFood the barrier was trust in a new restaurant. Gojek used Facebook Connect to show which merchants a user's friends had bought from, and users who saw that were twice as likely to buy from a brand new restaurant.
She calls retention too vague a goal, like asking for everything orange in the world instead of everything orange on a construction site.
She shortens the longest, most painful path that enough users attempt. At Kumu, users searched for friends on signup, but search was a slow, underused API. Speeding it up and adding more content lifted conversion from 60 to 90 percent over a few weeks.
She layers these sure wins with a few bets on new features, and runs an experiment on the bets if she can.
When visitors land and do nothing, she rewrites the copy first. Those users haven't experienced the product, so the product isn't the problem. She writes to the pain point rather than the solution so users can see themselves in the use case.
She sees founders who chase a brand new feature with no Wizard of Oz test or user data as grasping at straws. The question to ask is why this instead of literally anything else.
She ties unfamiliar products to something users already know. Many Gojek users didn't know what a digital wallet was. The team put each user's virtual account number on a picture of a credit card, and users typed it into an ATM like a debit account. Top-ups rose just from sending the picture.
She reads cancellation reasons to find the small fix that stops churn. For ABN InBev's beer subscriptions, the top cancellation reason was having too much beer, but users could only cancel or resume. A pause button fixed a temporary problem without a permanent cancellation.
No new products or reactivation emails were needed. "canceling the subscription is a permanent solution to having too much beer."
She adds deliberate friction to spot users with real intent. At Kumu a real friendship is hard to see, so the app asks users to fill out a questionnaire to join a community. Those who complete it raise a hand for community, and those who skip it probably want short-form entertainment.
The app leaves breadcrumbs to trace these paths.
She maps the business's physics before choosing any growth lever. Step one is to define market, product, model and channels. For Gojek that was Indonesian consumers and drivers, a mobile app, a charge per order, and channels like push notifications, Facebook ads and drivers in jackets.
Drivers on the street were marketing, and seeing them helped word of mouth as Gojek entered new cities. Step two is checking that loops and funnels fit those physics. If a loop needs four or five changes at once, it doesn't fit.
She turned drivers into a sales channel for GoPay. When a driver was allocated, a service checked whether the customer had ever topped up GoPay. If not, the driver got a message offering extra money for turning the rider's cash into wallet balance. This drove about 60% of acquisition once released.
Nothing in the product changed, only the use of an underused lever. "you wouldn't believe how great of a salesperson someone can be when you were literally trapped in a car with them."
She starts from what works and fixes the biggest constraint. She finds the constraint first, then optimizes the funnel or loop, changing one small thing at a time. In her view it is very hard to move the physics of a whole model.
She warns that launch spikes can be pulled-forward demand. When Netflix and Spotify launch in places like Southeast Asia, few people hold accepted credit cards. The week-one surge is everyone who could subscribe, and growth gets harder afterward.
She checks whether early adopters are exhausting a limited pool before assuming growth will continue.
She expects free products to flatten near 60% weekly retention. A paid product should flatten at 20 to 30%, at scale. Friends and family retention should be near 80%, because if people who care about you won't use it, strangers won't either.
She reads retention by cohort and gives teams two to three frequency periods to see the curve level off. Gojek's early retention was 60 to 70%, which told her the product just needs to work.
She puts properties on every event so data answers why. Gojek's map-loaded event carried drivers on screen, pickup location, city, surge pricing, minimum fare and voucher code. It showed users who saw two drivers booked far less than those who saw five, which pointed to areas with only two drivers.
She can judge a team's instrumentation just by reading its tracking spec. A spec where each event has one property or none is the main warning sign.
She separates observations from insights that change what you do. "Power users book four times more" is only an observation. Learning that power users use free shipping on high-value baskets is an insight, and it changed marketing spend for power versus non-power users.
Her method is to segment users by event properties, form a hypothesis, then run a test to get a causal answer.
She starts with free analytics tools and adds heavier ones as the company grows. With one data warehouse, Google Data Studio is free and enough, and Metabase is open source for SQL-skilled teams. For in-app mobile events she picks CleverTap, since Mixpanel failed her and Amplitude lacked CRM features.
As the company scales she adds Amplitude for analytics power, Segment to pipe data, and Eppo for experimentation, which replaced dashboards she once built by hand.
Growth at Gojek began as a crew filling gaps the core team left. The core team was busy building features like phone number masking. Growth took onboarding gaps, such as choosing which SMS provider sends the OTP for each telco.
Drivers with thousands of rides assumed every customer knew how Gojek worked, so the fix was telling the driver a customer was new. Growth PMs were later embedded in teams and became owners of product areas.
She says a separate growth team works only with insane product-market fit. When demand outruns the core team, a side team can act as the cleanup crew and plug in what was forgotten. Her team was heavy on statistics graduates who cared about numbers, odds and probabilities.
At that scale almost anything does something, so the question becomes what has the most impact and unlocks the future.
She hires the first growth person for statistics, not flash. The hire must be intuitively good at sampling, selection bias and not misreading ratios from small bases. She would hire anyone willing to take an intro statistics course, young or experienced.
She gives them a known gap to own rather than asking them to model all of growth. "the worst possible thing is to have a growth person who thinks they are doing the right thing and is measuring things wrong."
She screens growth candidates with a take-home experiment design. She presents something observed, asks how the candidate knows it is true, then has them design an experiment. Candidates get five days for about four hours of work, because live doesn't work.
She looks for random sampling and a measured approach, not build and hope. She likes hearing that candidates had to google and read white papers.
She favors quick, hacky experiments over heavy tool rollouts. Gojek's first growth experiment was a Python script on the Twilio API that texted drivers from a CSV, telling them to accept more rides. It raised acceptance by 2%, and more for brand new drivers, which led to work with driver onboarding.
She warns that onboarding many tools that don't integrate can eat six months with nothing shipped. "growth is an opportunity cost of time trade off with what you could have been doing to the product in that time."