Walmart US Search & Discovery

The Challenge

Over 6% of all Walmart.com search sessions were performing poorly due to query ambiguity — the same search term meant different things depending on user context, purchase history, and session intent. A user searching 'apple' could mean fruit or electronics. Keyword-based ranking had no mechanism to resolve this ambiguity at scale, resulting in high session abandonment, low add-to-cart rates, and missed conversion for Walmart's highest-value discovery categories including apparels and home appliances.

My Role

Lead PM for Search & Discovery, Walmart US (3.5 years) — owned query understanding and ML relevance strategy for Walmart.com search. Recognised with the Bravo Award for data-driven analysis and product strategy.

• Led the query-session graph initiative for ML-driven personalisation across 6%+ of search sessions
• Ran A/B experiments to validate relevance improvements, partnering with data science and ML on design and interpretation
• Coordinated across merchandising, engineering, and data science to align the roadmap

The Solution

A query-session graph built from crowd-engagement signals — clicks, dwell time, add-to-cart events — that connects what users search for to what they ultimately buy, rather than relying on query text alone:

• Establishes correlations between queries, categories, and items from session-level behaviour
• Becomes an input feature layer for Walmart's ML ranking models
• Blends a session-graph-informed recall layer into the existing ranking system, enabling contextual ranking that adapts to individual session intent

The Outcome

• +6.5 percentage point lift in session conversion rate
• ~$180M in annualized GMV contribution
• Became the foundational ML feature layer for Walmart.com's search personalisation strategy, underpinning subsequent ranking model iterations

Recognised with the Bravo Award for the data-driven product strategy behind this work.

Key Learnings

• Session-level signals (clicks, dwell time, add-to-cart) are more predictive of intent than query text alone — they reveal what the user meant, not just what they typed.
• A/B testing at Walmart scale requires statistical discipline: even a 0.1% conversion lift represents tens of millions in annualized GMV, so false positives are costly. Run experiments for longer than feels necessary.
• The hardest part of ML-driven ranking is the feature pipeline, not the model — getting clean, low-latency session signals into the ranking system took more engineering investment than the graph itself.

Technologies & Methods

• ML Ranking Models
• Query-Session Graph
• A/B Testing
• Python
• Spark
• Hadoop