Search suggestions as killer feature
summary
Auto retailer reports 1.9 million visits/day and up to 10 orders/sec and finds poor autocomplete can hurt revenue; A/B tests confirm negative impact and later uplift. They build an autocomplete using product data as a dictionary, clickstream ingestion via Kafka and Spark, store and query suggestions in Elasticsearch, apply Elasticsearch analyzers/stemming to fingerprint and dedupe variants, filter zero‑result entries, then score suggestions by weighted CTR, search frequency and order value — iterative A/B testing yields measurable revenue gains.
AI-written summary · about