Customer segmentation is one of the highest-ROI applications of data science in business. It replaces the "treat all users the same" approach with targeted strategies for distinct groups. This article uses Grab — Southeast Asia's leading super app — as a case study, walking through the strategic framework from business objectives to go-to-market execution.
All data in this article is hypothetical, used solely to illustrate the methodology. Grab's actual business data is not publicly available.
What Is Customer Segmentation?
Customer segmentation divides users into meaningful groups based on behavioral, demographic, or value-based similarities. Users within each group should be as similar as possible; groups should be as different as possible.
All Users (100%)
│
┌────────────────┼────────────────┐
▼ ▼ ▼
┌─────────┐ ┌─────────┐ ┌─────────┐
│ High │ │ Growth │ │ Dormant │
│ Value │ │ Users │ │ Users │
│ 15% │ │ 45% │ │ 40% │
└─────────┘ └─────────┘ └─────────┘
│ │ │
VIP program Cross-sell Win-back
Dedicated care Upgrade path Low-cost reachThe core principle is simple: not every user deserves equal investment. Concentrate resources on the most valuable and highest-potential groups, and overall ROI improves naturally.
Segmentation vs. Personalization
| Dimension | Segmentation | Personalization |
|---|---|---|
| Granularity | Group-level (thousands to millions) | Individual |
| Complexity | Medium | High (requires recommendation systems) |
| Use cases | Marketing strategy, product planning, pricing | Recommendations, dynamic pricing, push notifications |
| Data requirements | Moderate | Large-scale real-time behavioral data |
| Time to value | Fast (1-2 weeks) | Slow (full ML pipeline needed) |
Practical advice: Do segmentation first, personalization second. Segmentation is the 80/20 rule in action — 20% of the complexity captures 80% of the business value.
Business Value
- Marketing efficiency: Invest heavily in high-value users, use low-cost channels for low-activity users
- Product decisions: Different segments have different needs, changing feature priorities
- Retention strategy: Identify at-risk groups before they churn
- Pricing strategy: Different segments have different price sensitivities
Why Grab?
Grab is Southeast Asia's largest super app, spanning ride-hailing (GrabCar), food delivery (GrabFood), digital payments (GrabPay), grocery delivery (GrabMart), and financial services (GrabFinance).
┌─────────────────────────────────────────────────────┐
│ Grab Super App │
│ │
│ 🚗 GrabCar 🍔 GrabFood 💳 GrabPay │
│ Ride-hailing Food delivery Digital payments │
│ │
│ 🛒 GrabMart 🏦 GrabFinance │
│ Groceries Lending/Insurance │
│ │
│ ───────────────────────────────────────────── │
│ A single user may use 3-5 business lines │
│ → Rich behavioral combinations → Higher value │
└─────────────────────────────────────────────────────┘A single user might commute with GrabCar on Monday, order GrabFood for lunch, use GrabMart for groceries on the weekend, and pay for everything with GrabPay. This cross-business behavioral richness makes segmentation more challenging — and far more valuable.
The Complete Segmentation Process
┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐
│ Business │ → │ Data │ → │ Method │ → │ Interpret│ → │ Execute │
│ Objective│ │ Prep │ │ Selection│ │ Personas │ │ Strategy │
└──────────┘ └──────────┘ └──────────┘ └──────────┘ └──────────┘
▲ │
└──────────────── Iterate (monthly/quarterly) ──────────────┘A common mistake is starting directly with method selection. But segmentation starts with the business objective. Are you optimizing marketing budget allocation? Improving retention? Finding cross-sell opportunities? Different goals lead to different feature choices and segmentation strategies.
Step 1: Define the Business Objective
For Grab, common segmentation goals include:
| Goal | Key Question | Segmentation Focus |
|---|---|---|
| Marketing budget optimization | Which users deserve more investment? | Value-centric (spend, frequency) |
| Retention improvement | Which users are about to churn? | Activity change-centric |
| Cross-selling | Can ride-only users be converted to food delivery? | Business line combination-centric |
| New user nurturing | What's the path from trial to loyalty? | Lifecycle-centric |
This article's goal: Identify cross-business behavioral combinations across Grab's lines to design cross-sell strategies.
Step 2: Data Preparation and Feature Design
Available Data Sources
┌──────────────────────────────────────────────────────────┐
│ Grab Data Landscape │
├──────────────┬──────────────┬──────────────┬────────────┤
│ Ride Data │ Food Data │ Payment Data│ User Attr │
├──────────────┼──────────────┼──────────────┼────────────┤
│ Order count │ Order count │ Txn count │ Tenure │
│ Total spend │ Total spend │ Txn volume │ City │
│ Last order │ Last order │ Online/offline│ Device │
│ Common routes│ Cuisine pref │ Top-up freq │ Referral │
│ Avg distance │ Avg AOV │ Card linked │ Tier │
└──────────────┴──────────────┴──────────────┴────────────┘Feature Design Principles
Raw data can't be used directly. Good feature engineering determines segmentation quality.
Principle 1: Business-meaningful features
Avoid features like "standard deviation of GrabFood orders over 6 months." Good features should be immediately understandable:
- ✅ Total orders (engagement)
- ✅ Total spend (value)
- ✅ Days since last activity (churn risk)
- ✅ Ride share vs. food share (business line preference)
- ✅ Average order value (spending level)
- ✅ GrabPay usage rate (ecosystem stickiness)
Principle 2: Consistent time windows
All features should be based on the same time window (e.g., past 6 months).
Principle 3: Handle outliers
Extreme users (corporate accounts, test accounts, API users) can severely skew clustering results. Winsorize or remove before clustering.
Step 3: Choose the Segmentation Method
Methodology Comparison
| Method | Principle | Pros | Cons | Best For |
|---|---|---|---|---|
| RFM | Score on Recency, Frequency, Monetary | Simple, no modeling needed | Transaction-only, fixed dimensions | Quick wins, e-commerce |
| K-Means | Minimize within-cluster distance | Fast, scalable | Requires predefined K, outlier-sensitive | General purpose, large datasets |
| DBSCAN | Density-based clustering | Auto-determines K, finds outliers | Parameter-sensitive | Anomaly detection |
| Hierarchical | Agglomerative/divisive tree | Dendrogram visualization | Computationally expensive | Small datasets, exploration |
| LTV-based | Segment by customer lifetime value | Directly maps to business value | Requires LTV model | Subscription businesses |
Recommended Strategy: RFM First, Then K-Means
┌─────────────────────────────────────────────────────────────┐
│ Segmentation Strategy Path │
│ │
│ Week 1: RFM Quick Segmentation │
│ ├── 3 dimensions: Recency × Frequency × Monetary │
│ ├── Each dimension split into 5 quantiles │
│ ├── Output: Champions / Loyal / At Risk / Hibernating │
│ └── Purpose: Build intuition, give teams something │
│ │
│ Week 2: K-Means Multi-dimensional Clustering │
│ ├── 8+ dimensions: add business line prefs, payments │
│ ├── Elbow Method + Silhouette Score to choose K │
│ ├── Output: 5-7 refined clusters │
│ └── Purpose: Discover hidden patterns RFM can't see │
│ │
│ Week 3: Translate to Personas + Design Strategies │
│ ├── Name each cluster, build user portraits │
│ ├── Assign action plans per persona │
│ └── Purpose: Enable business teams to execute │
└─────────────────────────────────────────────────────────────┘Step 3a: RFM Quick Segmentation
RFM is the classic framework with three dimensions:
- R (Recency): How recently did the customer transact? More recent is better.
- F (Frequency): How often? Higher is better.
- M (Monetary): How much? Higher is better.
RFM Applied to Grab
High Frequency (F)
▲
│
At Risk │ Champions
Inactive but │ Recent, frequent, high-spending
formerly │ → VIP program
frequent │
─────────────┼──────────────→ Recent Activity (R)
│
Hibernating │ New Customers
Dormant │ Recent but low frequency
users │ → Onboarding nurture
│Score users 1-5 on each dimension, then assign labels based on the combination:
| Segment | R Score | F Score | M Score | Action |
|---|---|---|---|---|
| Champions | 4-5 | 4-5 | 4-5 | Reward loyalty, VIP treatment |
| Loyal Customers | 3-5 | 3-5 | 3-5 | Upgrade experience, cross-sell |
| New Customers | 4-5 | 1-2 | — | Drive second purchase |
| At Risk | 1-2 | 3-5 | 3-5 | Win-back campaigns, churn surveys |
| Hibernating | 1-2 | 1-2 | 1-2 | Low-cost reach or sunset |
RFM Limitations
RFM is effective but limited:
- Transaction-only: Ignores business line preferences, payment behavior
- Arbitrary boundaries: Threshold choices directly affect results
- Cannot discover hidden patterns: Confined to the R/F/M space
This is why we need K-Means.
Step 3b: K-Means Multi-dimensional Clustering
K-Means finds natural cluster boundaries across arbitrary dimensions.
Core Concept
Step 1 Step 2 Step 3
Random K centroids Assign each user Recompute centroids
to nearest centroid
· · ★ ····· ★ ····· ★
· · ···· ····
· · ····· ·····
· ★ ····· ★ ····· ★
· · ···· ····
· · ····· ·····
Repeat Steps 2-3 until convergenceThe algorithm automatically groups similar users together without manual threshold setting.
Key Decision: Choosing K
K too small (e.g., 2-3): Groups too coarse, can't distinguish meaningful differences K too large (e.g., 15+): Groups too fragmented, business teams can't design strategies for each
The practical sweet spot is usually K = 4-7.
Two diagnostic tools:
| Method | Principle | How to Read |
|---|---|---|
| Elbow Method | Plot inertia for different K values | Find the "elbow" bend point |
| Silhouette Score | Measures within-cluster tightness vs. between-cluster separation | Higher is better (0-1) |
K-Means Applied to Grab
Input 8 feature dimensions:
┌─────────────────────────────────────────────┐
│ K-Means Input Features │
├─────────────────────┬───────────────────────┤
│ Engagement Features │ Preference Features │
│ • Total orders │ • Ride share │
│ • Total spend │ • Food share │
│ • Days since active │ • GrabPay usage rate │
│ • Avg order value │ • Account age │
└─────────────────────┴───────────────────────┘
│
┌────┴────┐
│ K-Means │
│ K = 5 │
└────┬────┘
│
┌─────┬───────┼───────┬─────┐
▼ ▼ ▼ ▼ ▼
Cluster Cluster Cluster Cluster Cluster
1 2 3 4 5Important: K-Means is scale-sensitive. Spend might range 0-5000 while ride share ranges 0-1. Always standardize features first so each dimension contributes equally to the distance calculation.
Step 4: Cross-Business Segmentation — Grab's Unique Advantage
This is the unique value of a super app. Single-business companies can only see one behavioral dimension, but Grab can analyze ride, food, and payment behaviors simultaneously.
Business Line Activity Matrix
Food Activity
Low Medium High
┌─────────┬─────────┬─────────┐
High │ Commuter│ Balanced│ Super │
Ride │ │ User │ User │
Activity Med │ Light │ Moderate│ Food │
│ Rider │ User │ leaning │
Low │ Light │ Food │ Food │
│ User │ curious │ Lover │
└─────────┴─────────┴─────────┘Layer GrabPay usage on top for even more precise user identification.
Step 5: Persona Development and Go-to-Market
Translating numbers into business language is the make-or-break step for adoption.
Grab's Seven Personas
┌─────────────────────────────────────────────────────────────┐
│ Persona │ Behavioral Profile │ Est. Share │
├───────────────────┼───────────────────────────┼────────────┤
│ 🌟 Super User │ High rides + food + pay │ ~8% │
│ 🚗 Commuter │ High rides, low food │ ~15% │
│ 🍔 Food Lover │ High food, low rides │ ~20% │
│ 💳 Pay Power User │ High payments, few orders │ ~12% │
│ ⚖️ Balanced User │ Moderate across all lines │ ~18% │
│ 💡 Light User │ Low across all lines │ ~15% │
│ 📊 Moderate User │ Medium, no strong pref │ ~12% │
└───────────────────┴───────────────────────────┴────────────┘Strategy Blueprint Per Persona
Super User (~8% of users, ~30% of revenue)
┌─ Goal: Retain + Increase ARPU ────────────────────────────┐
│ │
│ Strategy: │
│ • VIP program (dedicated support, priority dispatch) │
│ • Cross-business subscription ($29.9/mo rides + food) │
│ • Surprise & delight (birthday perks, milestone rewards) │
│ │
│ Risk: Losing even one is extremely costly │
│ KPIs: Monthly retention > 95%, ARPU quarterly growth > 5%│
└───────────────────────────────────────────────────────────┘Commuter (~15%)
┌─ Goal: Cross-sell food delivery ──────────────────────────┐
│ │
│ Insight: Daily ride commuter, rarely orders food. │
│ May not realize how convenient GrabFood is. │
│ │
│ Strategy: │
│ • "Order dinner on your way home" contextual push │
│ • Lunchtime GrabFood 50% off first order │
│ • Post-ride prompt: "Grab dinner too?" │
│ │
│ KPIs: Food delivery activation rate, 30-day retention │
└───────────────────────────────────────────────────────────┘Food Lover (~20%)
┌─ Goal: Cross-sell rides + boost GrabPay ──────────────────┐
│ │
│ Insight: High-frequency food user, almost never rides. │
│ Likely remote/hybrid worker. │
│ │
│ Strategy: │
│ • Weekend ride discounts ("Get out this weekend? $3 off")│
│ • Extra GrabPay cashback on food orders │
│ • GrabMart cross-recommendation ("Restock while at it?") │
│ │
│ KPIs: Ride activation rate, GrabPay binding rate │
└───────────────────────────────────────────────────────────┘Pay Power User (~12%)
┌─ Goal: Convert to transactional user ─────────────────────┐
│ │
│ Insight: Uses GrabPay at convenience stores and │
│ restaurants, but rarely orders rides or food. │
│ High financial stickiness. │
│ │
│ Strategy: │
│ • GrabPay-exclusive ride/food discounts ("Pay & save 15%")│
│ • Spending milestones unlock ride credits │
│ • GrabFinance product recommendations │
│ │
│ KPIs: Order activation rate, GrabPay volume growth │
└───────────────────────────────────────────────────────────┘Light User (~15%)
┌─ Goal: Activation ────────────────────────────────────────┐
│ │
│ Insight: Occasional use or recent sign-up. │
│ Hasn't formed a usage habit yet. │
│ │
│ Strategy: │
│ • Onboarding quests ("Complete 3 rides, unlock $5 off") │
│ • Time-limited new user offers (3 free deliveries in 7d) │
│ • Referral rewards ($5 each for inviter and invitee) │
│ │
│ Caution: Control subsidy costs — avoid pure deal-seekers │
│ KPIs: 3rd order completion rate within 30 days │
└───────────────────────────────────────────────────────────┘Balanced User (~18%)
┌─ Goal: Stabilize retention + upgrade to Super User ───────┐
│ │
│ Insight: Moderate across all lines. Highest upgrade │
│ potential to Super User. │
│ │
│ Strategy: │
│ • Loyalty points across all business lines │
│ • Multi-line milestone rewards ("Use 3 lines this month")│
│ • Personalized recommendations by usage pattern │
│ │
│ KPIs: Cross-line usage frequency, Super User upgrade rate│
└───────────────────────────────────────────────────────────┘Step 6: Measurement and Iteration
Segment Migration Matrix
Segmentation doesn't end when the model runs. The most critical measurement tool is the segment migration matrix — tracking how users move between segments over time.
Current Quarter
Super Balanced Light At Risk
Prev Super [85%] 10% 2% 3%
Qtr Balanced 8% [72%] 12% 8%
Light 2% 15% [65%] 18%
At Risk 1% 5% 20% [74%]
Diagonal = stayed in same segment
Off-diagonal = migrated to another segmentHow to read this:
- Super → Super 85%: Strong super user retention
- Light → Balanced 15%: Activation campaigns converting 15%
- Balanced → At Risk 8%: 8% of balanced users starting to churn — investigate
- At Risk → Light 20%: 20% of at-risk users deteriorating further
This table directly tells you which strategies work and which segments need more attention.
A/B Testing Framework
Every persona's strategy should be validated through A/B testing:
┌──────────────────────────────────────────────────┐
│ A/B Test Design │
│ │
│ Target: Commuter segment (~15% of users) │
│ Hypothesis: "Evening GrabFood push" increases │
│ food delivery activation │
│ │
│ Control (A): No push notification │
│ Treatment (B): 5-7 PM GrabFood offer push │
│ │
│ Metrics: │
│ • Primary: Food activation rate (7-day) │
│ • Secondary: 30-day food retention │
│ • Guardrail: Ride frequency doesn't drop, │
│ unsubscribe rate doesn't increase │
│ │
│ Sample: 2,000 per group, 2-week runtime │
└──────────────────────────────────────────────────┘Common Pitfalls
1. Feature Selection Bias
Using only spend amount mixes "low-frequency high-AOV" with "high-frequency low-AOV" users.
→ Fix: Include frequency, recency, behavioral type, and other multi-dimensional features.
2. Too Many Clusters
K=15 looks precise, but the marketing team can't design 15 different strategies.
→ Fix: K=4-7 is usually the sweet spot. The goal is actionability, not minimizing mathematical error.
3. Ignoring the Time Dimension
A heavy user from 6 months ago may have already churned. Static segmentation still labels them "high value."
→ Fix: Include Recency as a key feature; re-run segmentation periodically (monthly or quarterly).
4. Non-Actionable Results
"Cluster 3 has an avg_order_value of $23.7" — the marketing team can't do anything with this.
→ Fix: Every segment must have an understandable label and a clear action plan. The data scientist's job doesn't end when the model runs — it ends when results are translated into business language.
5. Running It Once and Forgetting
User behavior changes over time. Last quarter's Super User might be this quarter's Light User.
→ Fix: Build a segmentation pipeline that re-runs periodically, tracking segment migration.
6. Ignoring Cluster Size Balance
If one cluster has 60% of users and another has 2%, the segmentation lacks discriminatory power.
→ Fix: Check cluster sizes. If one is too large, consider further sub-segmentation.
Conclusion: Segmentation Is an Operating System, Not a Report
┌─────────────────────────────────────────────────────────────┐
│ │
│ Week 1 RFM quick segmentation → Build intuition │
│ ↓ │
│ Week 2 K-Means multi-dimensional → Discover patterns │
│ ↓ │
│ Week 3 Translate to Personas → Design strategies │
│ ↓ │
│ Week 4+ A/B test → Measure impact → Iterate │
│ ↓ │
│ Monthly/Quarterly Re-segment → Track migration │
│ │
└─────────────────────────────────────────────────────────────┘The most important takeaway: Segmentation's goal is not minimizing inertia or maximizing Silhouette Score. It's enabling business teams to understand, act, and measure. A K=5 segmentation where every group has a clear strategy will always beat a K=20 segmentation that nobody knows how to use.
Great customer segmentation is not a report — it's a continuously running decision engine that tells your team who to target, what to do, and what results to expect.