Turning Used‑Car Search into a Smart Recommendation Engine

Search is compared to buying a car at a dealership: when the exact model (e.g., a black BMW X6) is unavailable, the salesperson recommends an alternative (e.g., a snow‑white version) with persuasive reasons, illustrating how recommendation works when the original query returns no results.

The unseen part of search involves backend processes such as query analysis, tokenization, inverted indexing, and ranking, which transform the user's input into information presented on the front end.

Users obtain information through two channels: the search engine itself and the recommendation system, both of which capture search intent and display curated results.

In the used‑car domain, conversion from search is 40% lower than overall business conversion, prompting an investigation into why the search funnel underperforms.

By analyzing 500 sampled search terms, the team classifies user intent into two major types: target‑clear (specific brand, series, price, etc.) and target‑vague (general usage or model categories), revealing regional variations and a lack of precise goals.

Four key problem areas are identified: the search box offers generic prompts, the search results page displays minimal information, suggestion words are scarce and poorly related, and the result page suffers from secondary filters that reduce precision.

The design goal is to educate vague users, gradually guide them toward clearer intent, and provide more accurate recommendations for target‑clear users.

Proposed improvements include: enriching the search box with contextual guidance and a rotating suggestion‑word carousel; redesigning the search page to showcase diverse, personalized recommendations instead of a blank canvas; introducing authoritative ranking lists based on accumulated big‑data; optimizing suggestion‑word logic by expanding types and improving semantic relevance; removing unnecessary secondary filters and moving type selection to the results page; enhancing result precision with multi‑dimensional secondary filters (brand, series, price, mileage) and dynamic inventory displays; and adding informational cards for market and dealer data to support multi‑type searches.

In summary, starting from user search terms, the team reconstructs search scenarios, uncovers issues at each node of the search chain, educates vague users, and incrementally guides them toward precise intent, thereby strengthening both the recommendation and search logic.