CASE STUDY
CarGurus: Assisted Search
Senior Product Designer
Prototype
Responsive Web
Year
2025–2026
Industry
AI Product Design
Space of work
Prototype
Timeline
1 quarter
Consumer Pain Points
Buyers described what they needed in full sentences but the search bar only accepted keywords, forcing them to manually translate their intent into filters they did not always know how to use.
Shoppers had no way to understand why a listing was relevant to them. Results appeared ranked by recency or price, with no explanation of fit, deal quality, or how a vehicle matched their specific situation.
First-time and undecided buyers landed on the Discover tab with no guided starting point. The blank search bar asked for specificity that most buyers in research mode did not yet have.
On the Vehicle Detail Page, buyers had access to specs and pricing but no plain-language interpretation of whether the car was actually a good deal for them personally, not just in the abstract.
Dealer Pain Points
Dealers could see inventory data but had no signal telling them which vehicles needed attention first. Aging stock, high-performing listings, and stale leads all looked the same in the dashboard.
Lead lists showed contact history but not buyer intent. A lead who had viewed a listing eight times and used the comparison tool looked identical to one who had clicked once and left.
Pricing decisions required dealers to manually cross-reference market data. There was no system connecting live market pricing to their active inventory with a specific, actionable recommendation.
Monthly performance data existed in tables and charts but required interpretation before it could inform a decision. There was no mechanism to synthesize the numbers into a prioritized action for the coming period.
Key Insight
Auditing both journeys revealed they shared the same gap: the platform collected behavior on both sides but never interpreted it. Every data point a buyer generated was a signal a dealer needed. The opportunity was to connect them through a single inference layer rather than build two separate AI features.
Consumer Key Features
AI Top Picks ranks every listing by match percentage based on the buyer's search behavior and preferences, surfacing the most relevant vehicles before a single filter is touched.
The AI Advisor flyout lets buyers ask questions in natural language from any page, comparing saved vehicles, questioning a deal rating, or exploring alternatives at a different price point.
The Discover tab replaces the blank search bar with a short preference-capture conversation. Four questions produce a ranked shortlist with a plain-language explanation for each result.
The Vehicle Detail Page AI Summary gives buyers a buyer score, a one-sentence verdict, and a pros and cons breakdown grounded in the vehicle's specs, pricing position, and local market data.
Dealer Key Features
AI Dealer Insights surfaces three plain-language alerts ordered by urgency: aging inventory with a price-drop recommendation, demand signals from buyer search behavior, and lead flags for buyers who have crossed an intent threshold.
AI Predictive Analytics generates a six-month sales forecast and ranks the lead list by close probability, factoring in browsing behavior, comparison usage, and save activity rather than just time since last contact.
Pricing Intelligence plots each active vehicle against the local market average and assigns a competitive status. Clicking AI Advice opens a modal with a plain-language rationale specific to that vehicle's mileage, model year, and local demand.
The AI Executive Summary generates a paragraph from the month's performance data identifying what performed, what underperformed, and what to prioritize next. It is regenerable after adjustments.
Key Insight
The most valuable thing AI can do for a dealer is not surface more data. It is tell them what to do next and explain why, in plain language, before they have to ask.
Both sides read the same signal
Consumer and dealer features were drawing from identical behavioral data. A buyer saving a listing three times is the same signal a dealer needs. Designing both in parallel made that connection visible.
Real AI changes how you design
Base44 ran real inference, not scripted outputs. The Advisor returned answers I did not anticipate, which forced me to design recovery paths I would have otherwise skipped.
Output format is a design decision
The Vehicle Summary went through four rounds of refinement before it stopped reading like a spec sheet. Treating the AI output format with the same care as the UI around it is a content design problem, not a prompt engineering one.
Dealers act on urgency, not categories
The first version of AI Dealer Insights organized alerts by type. Every dealer asked the same question: which one do I do first? Reordering by urgency instead of category made the feature feel like it understood the job.
