Three more bookings. Every hundred visits.
LalaFood wanted to move conversion from 25% to 28% by the end of 2024. That is a 12% relative increase, starting from its October baseline.
The question: what would make more people finish an order? Across 329,590 sessions, the strongest clues were who was browsing, what they saw, and how easy the choice felt.
A better discovery experience is the starting point. A second order is the next challenge.
Follow the volume, not just the highest rate.
Segment B brings 43% of sessions and converts at 30.42%. Segment D brings another quarter, but converts at only 18.59%. Both deserve attention: one already performs well; the other has room to improve.
Cuisine tells a similar story. Thai food converts at 46.67%, but from just 90 sessions. Rice dishes account for 28.34% of traffic and convert at 24.40%.
Small improvements in a large audience can matter more than impressive rates in a tiny one.
Nearby. Well reviewed. Priced to fit.
Restaurants within 2 km convert at 27.98%. Beyond 10 km, that falls to about 10%. More than half of sessions are already in the nearest band.
Trust matters too: restaurants with 4-5 stars and over 100 reviews convert at 32.62%. Strong price bands include IDR 40-50k and 110-120k, while most browsing happens at lower prices.
Put nearby, trusted restaurants first, then match the price range to the customer.
An offer needs to be seen.
Sessions with a promotion shown convert at 26.92%, compared with 1.75% without one. RPL & MFP promotions lead the offer types at 31.03%.
GoJek Plus members also convert more often: 27.13% versus 24.91% for non-members. Yet members represent only 18.06% of sessions.
Make relevant offers visible and test membership incentives. These are associations, not proof of causal lift.
Dinner is the moment to win.
Weekday dinner converts at 27.01% and accounts for 23.06% of sessions. It combines strong intent with substantial traffic.
Weekend lunch has a similar conversion rate, 26.80%. The opportunity is to pair these high-intent moments with nearby restaurants and time-limited offers.
Start with a Weekday Dinner Deals test, reaching customers before the 6-9 PM window.
Find more likely bookings. Know the trade-off.
With roughly three non-bookings for every booking, accuracy alone is not enough. Logistic regression finds just 22.52% of actual bookings; tuned XGBoost finds 75.62%.
The tuned model reaches 82.30% accuracy and 61.98% precision. It captures more bookings, but produces more false positives than logistic regression.
Tuned XGBoost is the stronger candidate when identifying more bookings is the priority.
Co-branding and promotions stand out.
The model's strongest signal is co-branding activity over the past six months. Promotion availability and visibility follow, with distance adding another important signal.
SHAP measures how much each feature influences predictions. The ranking shows average influence size; the feature explanations describe the patterns behind it.
Make the prediction useful: connect the strongest signals to choices the product team can test.
The first order is only the beginning.
Across cohorts, only 11-29% return in Month 1. The steepest loss comes immediately after the first transaction, with further erosion around months 4-5.
New cohort sizes also decline, from 275,386 in June 2024 to 47,587 in March 2025. March is only partially observed, so it is not a full-month comparison.
Make the second order easier, then re-engage customers before the later drop-off.
Make discovery easier. Give people a reason to return.
Test a Flexible Nearby Restaurants experience: prioritize restaurants within 2 km, surface trusted options, and keep relevant promotions visible. Pair it with Weekday Dinner Deals and a GoJek Plus trial.
For retention, reward the second transaction, personalize re-engagement around months 4-5, and offer loyal customers lasting benefits.
Measure conversion and repeat orders in A/B tests before a wider rollout. The goal is 28%, not an assumed result.