Customer-facing genAI search POC
I led a proof of concept exploring how generative AI could help customers ask natural-language questions, compare product information, and move from vague intent to confident results.
Before
running shoes
Trail runner, size filters hidden
Best match unclear across variants
Customer scans, refines, repeats
After
I need shoes for city walks in the rain
Recommended direction
Prioritize waterproof uppers, visible grip, and all-day cushioning. Show trade-offs before the product grid.
waterproof
commute
comfort
01
Best for wet city walking
02
Lighter option with less traction
Project details
Role
Product design lead
Team
Product, engineering, search relevance, data science, and legal
Responsibilities
Discovery, conversational UX, prototype flows, evaluation criteria, stakeholder readouts
Key contributions
Search intent framework
Mapped vague customer language into actionable intent categories, so the AI response could clarify needs without overwhelming the shopper.
Answer pattern system
Created reusable response structures for summaries, trade-offs, follow-up prompts, and product ranking explanations.
Risk and trust alignment
Partnered with legal and product stakeholders to define guardrails for confidence, attribution, and failure states before scaling the concept.
Opportunity • Customers often describe needs before they know product terminology. • Search results surfaced products, but not the reasoning behind the match. • The team needed a low-risk POC to learn where AI could improve discovery without replacing familiar browsing patterns.
Goals • Prototype a conversational layer that clarifies intent, not just keywords. • Define answer patterns that balance helpfulness, trust, and product merchandising. • Establish evaluation criteria so the concept could move from demo to roadmap decision.
My approach
1
Map natural language intent
I reviewed real search logs, support questions, and on-site filters to identify how customers described needs before choosing product attributes.
2
Create an answer hierarchy
Before exploring interface polish, I defined what a useful answer needed to contain: interpretation, confidence, trade-offs, and a next action.
Intent
Reasoning
Results
3
Prototype the search moments
I built flows for query entry, AI summary, follow-up refinement, product ranking, and the handoff back to normal browsing.
Query → Interpret → Explain → Rank → Refine
Each moment had a fallback path so the POC could fail gracefully when confidence was low.
4
Evaluate trust and feasibility
The final concept review focused on where AI created clarity, where it introduced risk, and what data the team needed before a pilot.
Comprehension
Trust and attribution
Merchandising fit
Content strategy decisions
Prompt framing
We tested customer-language prompts, not system prompts, to keep the concept focused on what shoppers would actually ask.
Explainability
Every recommendation needed a short reason so customers could trust the ranking and correct the AI when needed.
Governance
The POC documented what the AI should never infer, when to hand off to classic search, and how low-confidence answers should appear.
Outcomes
3
testable search moments defined
12
intent scenarios reviewed with stakeholders
1
roadmap recommendation for pilot planning