The Challenge
Customers struggled to articulate subjective aesthetic preferences using traditional technical text filters, leading to low engagement with an existing ML recommendation engine.
Role - Behavioral loop framework, visual discovery interaction model, usability testing scenario architecture, and cross-team stakeholder alignment.
Team - 1 Senior Designer (Me), 1 Product Manager, 1 Content Designer, 1 UX Researcher, Engineering, Data Science.
Timeline - 3 Months (Concept through Shipped MVP and Feature Scoping).
The Strategy
Build customer confidence in our AI recommendation engine by designing an interface that translated raw swiping gestures into immediate visual feedback, showing users exactly how the algorithm adjusted to their preferences.
The Outcome
$27
10%
+9%
12%
In Revenue
Engagement
Conversion
Add to Carts
The Problem
Shopping for visually driven products, fixtures, décor, and finishes is inherently subjective.
Customers often know what they like when they see it, but struggle to describe that preference using filters, specs, or technical language.
Visual Scout already existed as a machine-learning powered experience that allowed customers to like or dislike products and receive updated recommendations.
However, engagement was lower than expected.
Customers didn’t fully understand how their actions influenced results.
-
Product reordering felt unpredictable.
-
The system felt algorithmic rather than responsive.
-
This wasn’t a product availability problem.
-
It was a translation and trust problem.
The Design Challenge
The challenge wasn’t to redesign screens.
It was to make a machine-learning system feel human, responsive, and intentional.
We needed customers to confidently answer one question:
-
“Is this actually helping me find something I like?”
At the same time, the experience had to:
-
Work across mobile web and app.
-
Integrate with an evolving recommendation engine.
-
Scale across visually driven categories.
-
Avoid overwhelming customers with explanation.
This required balancing clarity, system intelligence, and cognitive load.
Phase 1
Engineering the Behavioral Loop
The Strategy: Replacing traditional search criteria inputs with a continuous interaction cycle based on instinctive visual reactions.
Traditional filtering assumes customers know what they want.
Visual Scout challenged that assumption with a different premise:
-
Customers react more easily to options than articulate criteria.
Through low-fidelity sketches and behavioral flow mapping, I pressure-tested three core questions:
-
How does a customer enter the experience?
-
How does the system respond to signals?
-
How does exploration transition into evaluation?
Rather than starting with polished UI, I defined a behavioral loop:
Signal → System Response → Reinforcement → Refined Signal
This loop became the foundation for the system.
We validated that the core issue wasn’t product availability, it was translation. Customers struggled when asked to explain what they liked, but responded quickly and confidently when reacting visually.
Trust depended on visible feedback.
Phase 2
Prototype
North Star Vision
The long-term vision was to help customers move seamlessly from preference expression to confident comparison without forcing traditional filtering too early.
The North Star experience envisioned customers being able to:
-
Save liked items for later review.
-
Restart or refine preference sets.
-
Quick View products without leaving discovery.
-
Compare up to four items side by side.
This model created a scalable path from exploration to evaluation while allowing delivery to be sequenced intentionally.
What We Validated
-
As we tested early concepts, we validated that the core issue wasn’t product availability, it was translation. Customers struggled when asked to explain what they liked, but responded quickly and confidently when reacting visually.
-
We also learned that trust depended on transparency. When customers could clearly see how their likes and dislikes influenced results, confidence in the experience increased significantly directly shaping how feedback and system response were surfaced in the final design.
Phase 3
MVP Scoping & Shipped Experience
The Strategy: Managing a strategic feature deferral to validate the core swipe-based preference model before investing in complex evaluation tools.
To reduce risk and validate the core behavior, we intentionally launched a focused initial experience.
The first release centered on validating visual preference expression and system feedback the foundational behavior the rest of the experience depended on.
The shipped experience prioritized:
-
Clear swipe right / swipe left interactions.
-
Accessible tap-based alternatives.
-
Immediate, visible system response.
-
Reduced cognitive load during early discovery.
This allowed us to test whether customers would meaningfully engage with a visual-first discovery model before investing in deeper evaluation features.
Tradeoffs & Delivery
The Tradeoffs: Testing confirmed users genuinely wanted features like saved sets and side-by-side product charts. However, we made a deliberate strategic call to defer these secondary evaluation tools to answer a more fundamental question first: would a swipe-based interaction model actually work in a high-consideration home improvement context?
The Delivery Strategy: By prioritizing a focused MVP centered strictly on preference expression, the initial release favored raw exploration over deep evaluation. This limitation gave us clean data, and the resulting 10% engagement spike secured the evidence needed to scale.
What This Taught Me
I came into this project assuming the ML model was the primary problem. What the research revealed was that the algorithm was working, but customers just didn't trust it.
That shifted how I think about intelligent systems: the design layer isn't decoration on top of the technology, it's the trust infrastructure that makes the technology usable. I now push earlier in discovery to understand whether a system has a perception problem versus a performance problem.
More Case Studies
