20% Reduction in Search Friction for India's Largest E-Retail Platform
20%
Reduction in search pain
$15M+
Annual revenue impact
18 weeks
Time to production
About the Client
One of India's largest e-commerce platforms processing billions of transactions annually. The platform serves as the primary shopping destination for over 150 million monthly active users, managing a catalogue of 100+ million SKUs across thousands of categories from electronics to fashion to groceries.
Challenge
The platform's search and discovery experience had become a critical bottleneck for conversion and customer satisfaction. Internal metrics revealed alarming trends:
• Search Pain at 46%: Nearly half of all search sessions were resulting in user frustration-measured through abandonment rates, reformulated queries, and support tickets.
• Poor Relevancy Scoring: Search results were populated based on keyword matching rather than true intent understanding.
• Wrong Ads Placement: Sponsored product placements were disrupting organic search results, showing users products in entirely wrong categories.
• Mismatched Results: The query-to-product matching system was failing on synonyms, regional language variations, and brand misspellings.
• Data Quality Issues: Underlying catalogue data had accumulated inconsistencies over years of rapid scaling.
Our Approach
Over 18 weeks, BergLabs deployed a comprehensive cross-functional intervention combining AI-powered annotation, system-level enhancements, and managed operations.
Phase 1: Deep Data Labeling (Weeks 1-8)
Our annotators evaluated 2.5M+ query-product pairs, building training data for ML model improvements.
- Query-to-Product (Q2P) Relevancy Labeling: Evaluated 2.5M+ query-product pairs, scoring relevancy on a 5-point scale. This enabled 40% more accurate ground truth than previous crowdsourced labeling.
- Semantic Matching Annotation: Built synonym dictionaries and semantic equivalence mappings for 50,000+ product terms across 200+ categories.
- Relevance Re-ranker Evaluation: Conducted blind A/B evaluations of 100,000+ search result sets, providing statistically significant validation before production deployment.
- 2-Pane Side-by-Side Evaluations: Performed comparative judgments between competing product results for ambiguous queries.
Phase 2: System-Level Enhancements (Weeks 9-14)
Implemented comprehensive catalogue improvements and search optimization.
- Catalogue Data Cleanup: Identified and corrected 1.2M+ data errors including misclassified products, incorrect brand attributions, and conflicting specifications.
- Query Handling Optimization: Built robust query parsing for misspellings, partial matches, and regional language mixing.
- In-Stock Item Mapping: Connected search ranking to real-time inventory signals, reducing "add to cart" failures by 35%.
- Brand Mismatch Resolution: Eliminated 200,000+ listings with incorrect or fraudulent brand attributions.
Phase 3: Continuous Operations (Weeks 15-18+)
Established ongoing quality assurance and improvement processes.
- Ongoing QC Pipeline: Dedicated 50-person operations team providing continuous search quality auditing, processing 100,000+ evaluations weekly.
- Feedback Loop Integration: Built direct integration between BergFlow and the platform's ML training pipelines, enabling same-week model updates.
┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
│ Query Logs │────▶│ BergFlow │────▶│ Training Data │
│ (Raw Input) │ │ Annotation │ │ (Labeled) │
└─────────────────┘ └─────────────────┘ └─────────────────┘
│ │
│ ▼
│ ┌─────────────────┐
│ │ ML Training │
│ │ Pipeline │
│ └─────────────────┘
│ │
▼ ▼
┌─────────────────┐ ┌─────────────────┐
│ Quality │ │ Production │
│ Dashboard │◀───│ Search API │
└─────────────────┘ └─────────────────┘Impact
The comprehensive intervention delivered measurable business results across all key metrics:
| Metric | Before | After | Improvement |
|---|---|---|---|
| Overall Search Pain | 46% | 26% | 20% reduction |
| Selection Pain | 38% | 22% | 16% reduction |
| Wrong Ads Pain | 28% | 12% | 16% reduction |
| Search Abandonment | 31% | 19% | 12% reduction |
Revenue Impact: $15M+ annual revenue lift directly attributable to improved search conversion, validated through controlled A/B testing across user cohorts.
Operational Efficiency: The platform reduced internal search QA headcount by 60% while achieving higher quality coverage through BergLabs' managed operations model.
Time-to-Market: New category launches now include search optimization from day one, reducing category ramp-up time from 6 weeks to 10 days.
Testimonial
“We were looking to build something that lasts. Berg proved that with the right systems and people, excellence isn't limited by geography. They didn't just improve our search metrics-they transformed how we think about data quality as a competitive advantage.”
VP of Product
Leading E-commerce Platform
Engagement Model
Type
Operations Automation Program
Duration
18 weeks initial + ongoing managed operations
Team
50+ dedicated annotators, 3 ML engineers, 1 delivery lead
Platforms
BergFlow (annotation), BergForge (automation agents)