The JB Hi-Fi PDP is the highest-traffic, highest-stakes page in the purchase funnel. Small improvements in conversion here have outsized commercial impact. Working with product and analytics, I identified hypotheses, designed the variants, and interpreted results to guide shipping decisions.

JB Hi-Fi is exploring how to integrate AI-powered search into its retail experience. Before designing anything, we needed to understand the mental models and trust drivers that would determine whether customers would engage at all.

Before committing to any direction, working with the Product Owner, I led a structured exercise to map every potential AI use case across JB Hi-Fi — evaluating each against business effort, customer impact, risk, and competitive landscape. The goal was to separate genuine opportunities from noise, and ensure the team invested in AI initiatives with real commercial and customer value.

The matrix below plots use cases from low to high on both business impact and customer value axes. Highlighted in green are the initiatives already in motion — including AI search, which sits in the high-value quadrant alongside personalised search and shopping assistant. This exercise directly shaped the decision to prioritise AI search as the first live experiment.

AI use case scoping — business effort vs customer value Click to zoom

To understand how customers perceive, trust, and adopt AI in a retail context, I ran two studies in parallel.

Across both studies, three clear themes emerged that shaped every design decision that followed.

A key output of the research was understanding where AI fits in the shopping journey. I mapped the actions customers expected AI to complete across discovery, research, purchase, and post-purchase — ranked by importance. Discovery and research tasks consistently ranked highest; purchase and post-purchase actions came with significantly more anxiety around trust and accuracy.

Actions customers expect AI to complete — ranked by importance across journey stages Click to zoom

This finding directly shaped the decision to focus the first experiment on search and discovery — the highest-value, lowest-risk entry point — rather than attempting to automate purchase or post-purchase actions before trust had been established.

The four takeaways became four design constraints. These weren’t preferences. Each one was a direct response to a specific finding, and any concept that violated them was ruled out before we explored it further.

This project is ongoing. In parallel with the live experiment, I am designing across the full end-to-end AI experience — mapping where AI assistance can be introduced at every stage of the customer journey, from home page through to post-purchase. The wireframes below represent current ideation across the home page, PDP, and search results, with the full journey still being defined.

AI entry point concepts — home page, PDP & search pages Click to zoom

These concepts are early-stage ideation. The live experiment on the search page entry point provided the data foundation to evaluate which direction to progress.

The research principles were validated through a live opt-in experiment — an AI-powered conversational search experience, explicitly labelled as beta, framed as "a new way to search." Every design decision I made reflected the research: opt-in entry, transparent framing, visible control, conversational language.

Control — Standard search flow

AI flow — Opt-in conversational search

The AI flow introduced an opt-in entry card within the standard search overlay — "Explore the new way to search · Beta" — inviting customers to try a conversational search experience without replacing the existing flow. Customers who engaged entered a natural language search interface, framed as a JB Hi-Fi team member, returning curated results with similarity and comparison actions.

The experiment ran for 11 days at a total AI message cost of $134. Incremental revenue attributable to the AI flow was $334K over the period — a $262 return for every dollar spent. This established the commercial baseline for the next phase of investment.

The experiment validated the core thesis — opt-in, transparently framed AI search drives higher-intent engagement and meaningful commercial return. The next phase focuses on scaling the experience: improving result relevance, reducing the opt-in friction, and testing embedded entry points earlier in the search journey. The research foundation I established in phase one continues to guide every design decision going forward.

Accessories, warranties, and complementary products are some of JB Hi-Fi's highest-margin commercial levers — and the attachment recommendation system is what drives them. The issue had existed for a while.

Before proposing any solution, I needed to answer a specific question: what is FBT actually supposed to do — and what should be doing the jobs it wasn't? I mapped every major recommendation model across key competitors, documenting the job-to-be-done, placement logic, and use cases for each. The goal wasn't a catalogue of what exists. It was to build an evidence-based case for which models JB Hi-Fi needed, which were being misapplied, and why the logic needed to change — not just the design.

Competitive mapping — Miro board Research phase

Click to zoom

Mapping Similar Items, Buy Again, FBT, On Sale, Recently Viewed — each with JTBD, market examples, placement logic, and use cases documented.

The mapping confirmed what the data suggested — FBT and Similar Items are fundamentally different tools serving different jobs, and conflating them was causing both to fail. Two decisions followed directly:

• Keep FBT — but redesign it. FBT was the right model for JB Hi-Fi's context, but the visual design made tiles indistinguishable from other product tiles across the site. A distinct visual treatment was needed so customers could immediately understand they were in a complementary purchase moment, not browsing.
• Introduced Similar Items and View Alternatives. These models were absent from the experience entirely, despite being standard practice. Adding them gave customers a structured way to explore alternatives without leaving the PDP.

Redesigning the UI without fixing the underlying logic would have been cosmetic. The root cause was algorithmic. FBT was trained to surface same-category products, so a TV PDP showed more TVs. Working with engineering, we changed the logic to surface complementary items: a TV PDP now shows a soundbar, a wall bracket, an HDMI cable. Items that complete the purchase, not compete with it. This was as much a product decision as a design one, and getting it across the line required the competitive mapping to make the case.

Three changes came directly from the diagnostic: fix the FBT logic, visually differentiate FBT from generic product tiles, and introduce the missing models. Each change had a specific job.

Before FBT showing same-category products. No Similar Items. No View Alternatives.

FBT surfaces more laptops on a laptop PDP — same category, not complementary. No Similar Items. No in-tile discovery.

After FBT showing complementary categories. View Alternatives on each tile.

TV PDP: FBT shows a wall mount ($69) and soundbar ($399). "View alternatives" on each tile for in-page discovery without leaving the PDP.

Fixing the FBT logic and introducing new models solved the product problem. The design challenge was making the system feel coherent across the full journey. Not a collection of modules — a single considered experience. Four principles shaped every decision:

• Sequence matters more than placement. Attachments appear only after the customer has resolved their primary product decision — not during it.
• Group, don't scatter. Consolidate complementary options into a single decision moment rather than distributing across multiple surfaces.
• Make relevance visible. Language, hierarchy, and visual design should signal why a recommendation is being made.
• Complement, don't compete. Every recommendation should make the primary purchase feel more complete — not introduce doubt.

Four variations were tested against the control — combining the new UX with varying levels of relevancy reranking. Variation 1 (new UX, no reranking) and Variation 4 (new UX, 80% relevancy reranking) both reached 93% statistical significance. The result validated the core thesis: fixing what the system recommended mattered more than how much it recommended.

Product Recommendation Redesign — Var 1 · 93% stat. sig.