All Case Studies
E-CommerceMarch 20268 min read

AI-Powered Enterprise Search for MENA's Largest Online Retailer

A top-3 online retailer in the Middle East with 20M+ monthly active users and a catalog of 12M+ SKUs across 40+ categories.

Search Relevance

91%

+168%

Conversion Rate

2.4x

+140%

Query Latency (P99)

120ms

-73%

Annual Revenue Impact

$18M

+$18M

The Problem: Search Was a Dead Feature

When we first audited the platform, the data told a damning story. Only 18% of users were using search — the rest had learned to browse categories manually. For an e-commerce platform with 12 million SKUs, that's catastrophic. Users searching for "كريم واقي شمس" (sunscreen cream) were getting results for kitchen cream dispensers. Arabic morphology, regional dialect variations, and the prevalence of Arabizi (Latin-script Arabic) meant their keyword-based Elasticsearch setup was functionally broken for the majority of their user base.

Our Approach: Hybrid Search with Cross-Lingual Understanding

We took inspiration from the approach described in Google's 2023 paper on multilingual retrieval and adapted it for Arabic e-commerce. The architecture has three layers: Meilisearch handles the first-pass retrieval with typo tolerance, faceted filtering, and sub-50ms response times. A custom bi-encoder model — fine-tuned on 2M+ Arabic product queries — generates cross-lingual embeddings that map Arabic, English, and Francoarabic queries into a shared semantic space. A final re-ranking transformer scores the top 100 candidates for purchase intent relevance. We used Meilisearch as the primary engine specifically because its typo tolerance and speed characteristics are perfectly suited to Arabic transliteration patterns, where users routinely misspell or transliterate product names.

Search Relevance Score (%) Over Time

34
Week 1
41
Week 2
52
Week 3
64
Week 4
72
Week 5
79
Week 6
86
Week 7
91
Week 8

Query Language Distribution

Arabic 58%
English 27%
Francoarabic 10%
Mixed / Other 5%

The n8n Backbone: Automated Reindexing & Monitoring

One of the most impactful parts of the project was the n8n-powered automation layer. We built workflows that automatically detect catalog changes, trigger re-indexing pipelines, monitor search quality metrics in real-time, and alert the team when relevance scores drop below thresholds. This meant the system stayed healthy without manual intervention — critical for a catalog that changes 50K+ SKUs per day. n8n also orchestrates the A/B testing pipeline, routing 10% of traffic to experimental ranking models and automatically promoting winners.

Daily Search Query Volume (K)

320K345K410K485K540K610K680K740KWeek 1Week 2Week 3Week 4Week 5Week 6Week 7Week 8

Measuring Impact: From Vanity Metrics to Revenue

We instrumented everything. Search relevance is measured using NDCG@10 against human-labeled query-product pairs. Conversion attribution tracks search → product view → add-to-cart → purchase within a session. The 2.4x conversion rate improvement translates to measurable revenue because we can prove the causal chain. The $18M annual revenue impact figure was validated by the client's data science team using a 6-week holdback experiment.

Key Results

Search relevance jumped from 34% to 91% within 8 weeks. Search-driven conversion rate increased by 2.4x, and the client estimated $18M in additional annual revenue attributable to improved search quality. The system now handles Arabic, English, and Francoarabic queries natively.

Technology Stack

MeilisearchPythonTransformersAWS BedrockKubernetesElasticsearch

Want similar results for your business?

Book a free 30-minute consultation — no pitch deck, just a conversation.

Get in Touch →