Case study
Harbey
Find it. Know it’s safe to buy.
An app and responsive website that helps grocery shoppers find products in the store, and understand whether they’re safe for their household.
- UX Research
- UX Design


About this project
Key results
- Usability score, first to second round of testing5 people in each round
- +13%
- Tasks completed in the second roundup from 33 of 35 in the first
- 30/30
- Testers with low vision found products on their ownin the second round
- 2/2
Overview
- The problem
- Shoppers lose time hunting for products in stores they don’t know well. People shopping around a food allergy then face a second job: reading small print in the aisle to check it’s safe.
- The goal
- Help shoppers answer two questions in the store: where is it, and is it safe for my household? No account needed, and usable with low vision.
- What I did
- Interviewed 5 shoppers, audited 3 competitors, defined two personas, went from paper to hi-fi, and ran two rounds of testing with 5 people each, redesigning in between.
- The result
- In the second round, every tester completed every task and the usability score rose 13%. Both testers with low vision found products on their own.
The problem
Finding a product is only half the problem.
- First: finding it
A product you can’t find
In a store you don’t know well, or after sections have moved, one product can take several laps. Asking rarely helps: it feels like interrupting, and staff may not know either.
- Second: checking it
A label you can’t read at a glance
For someone shopping around a food allergy, finding the product is only half the job. Then they read small print in a dim aisle and decide whether it’s safe.
A wrong answer in the second job costs far more than time. That’s why Harbey answers both questions in one place.
The goal
Brief
Help people find it, and know it’s safe.
The task I set first
Help shoppers find products in the store with less walking.
The task after research
Help shoppers find products in the store, and understand whether each one is safe for the restriction they’ve saved.
Research doubled the task. The process below shows why.
What success had to look like
- Find it
A product found on the first try
Without extra laps around the store, even in a store the shopper doesn’t know well.
- Check it
A safe answer people can trust
Clear enough to act on, with the reason behind it, before the product goes on the list.
How I measured it: task completion, observed errors and usability scores (SUS) across two rounds of usability testing.
Out of scope, on purpose: cart, checkout and accounts. Low vision was a requirement from the first sketch, not a later pass.
The process
How each step led to the next.
01Research
Finding the product was only half the job.
- The question
- Why do shoppers struggle to find products, and what do they do about it?
- What I did
- Interviewed 5 shoppers of different ages, habits and familiarity with their store.
- Decision
- The project grew from one job to two: Find it and Check it.
I interviewed five shoppers of different ages, habits and familiarity with their store. My questions were about finding products, but three of the five independently described a second problem: even after finding a product, they still had to work out whether they could safely buy it. That changed the direction of the project.
- 01
Store layout
People lose time in stores they don’t know well, especially when sections have moved.
- 02
Product labels
Checking whether something is safe means reading small print: slow, and hard in a dim aisle.
- 03
Scattered products
Special-diet products appear in several places, so it’s hard to know you’re holding the right version.
- 04
Asking for help
Most wouldn’t ask staff. It feels like interrupting, and staff may not know the answer either.
“I needed tahini last week and walked past it three times.”
02Competitors
Nobody connected both problems.
- The question
- Is someone already solving both problems, and where do they fall short?
- What I did
- Ran Walmart, Kroger and Fig through the same six tasks.
- Decision
- Harbey connects both jobs and works without an account.
I ran Walmart, Kroger and Fig through the same six tasks. Each one solved finding or dietary checking, but none of them connected the two.
| Product | Find product | Dietary safety | Connected |
|---|---|---|---|
| Kroger | Yes | Yes | No |
| Walmart | Yes | No | No |
| Fig | No | Yes | No |
| Harbey | Yes | Yes | Yes |
Harbey works without an account. All three competitors asked for an account before giving any useful information, and Fig lost my saved restriction after a reload. In Harbey, the restriction lives on the device.
03Personas
Two shoppers needed almost opposite things: certainty and speed.
- The question
- Five interviews, five very different shoppers. Which needs repeat, and who am I designing for?
- What I did
- Made an empathy map for each interview, grouped what repeated, turned the two strongest patterns into personas, and mapped a journey for each.
- Decision
- Two personas, and one rule for every screen: a shortcut can never cost confidence.
I made an empathy map for each of the five interviews. Two patterns held across several shoppers, so each became a persona. Neither is a single real participant.
Ananya
Age 42Has low vision and shops for a household member with a peanut allergy. For her, being wrong is not an option.
Built fromA parent managing a peanut allergy, a retiree who reads labels with a magnifying glass, and a shopper whose gluten-free items are split across the store.
Success meansConfidence and accuracy, over speed
Amari
Age 23Eight months at the same store and still no clear mental map. A knee injury makes every extra lap costly.
Built fromA shopper with a knee injury who passed the tahini three times, and a student who gives up on an item after about two minutes.
Success meansFewer steps, first-attempt finds, shorter trips
A journey map for each showed where the trip breaks: for Ananya, a moved section and labels in a dim aisle; for Amari, hunting one item and backtracking. Those moments became the features: an exact location, a safety answer with its reason, and a list sorted into a route.
On every screen I asked two questions: would Ananya trust this, and would Amari walk less?
04Structure
One vocabulary for one physical place.
- The question
- How should location and safety be described, so people never have to relearn them?
- What I did
- Mapped the information architecture around the places a shopper actually walks.
- Decision
- Every location reads Aisle → Side → Shelf height. Safety has three answers, always worded the same.
Inside a store, a category and an aisle both point to the same thing: a physical location. So every location in Harbey is written the same way, everywhere.
- Aisle
- Side
- Shelf height
The same rule applies to safety: three possible answers, always in the same language and structure.
- Safe
- Not confirmed
- Not safe
Seven branches from Home in the app, including Scan, with one vocabulary throughout.
05Sketches
I needed to validate my ideas before polishing anything.
- The question
- Which layout gets someone from the list to the right shelf fastest?
- What I did
- Sketched the flows on paper, moved the strongest into lo-fi wireframes, and prototyped the whole journey.
- Decision
- Search and scan share the header, and the shopping list stays on Home.
Paper let me explore the flows quickly. The strongest ideas moved into digital wireframes, and the first prototype covered the whole journey.






Search and scan share the header
They’re the two main ways into a product, so both stay reachable from the same place.
The shopping list stays on Home
The trip is visible the moment the app opens, instead of hiding behind another tab.
- Set a restriction
- Search
- Check the answer
- Scan
- Find a safe alternative
06Testing · Round 1
- Tasks completed
- 33/35
- SUS score
- 66.5
- Observed errors
- 20
The completion rate looked encouraging. The behavior didn’t.
- The question
- Can people get through the journey, and do they read a safety answer correctly?
- What I did
- Unmoderated test of the lo-fi prototype: 5 participants, 7 tasks.
- Decision
- I stopped trusting completion alone and fixed four behaviors.
Low fidelity · 5 participants · 7 tasks · Unmoderated
- 5/5acted on a safety answer without checking the evidence→ Added the information behind each answer
- 4/5couldn’t figure out how to start a shopping trip→ Added a clear starting action
- 3/5thought a route stop meant one item, not one aisle→ Restructured the route and added item counts
- 4/5struggled to choose between safe alternatives→ Gave each one an aisle, side, shelf height and distance
“Safe according to who, and when?”
07Iteration
I made the answer explain itself.
- The question
- How can a safety answer earn trust instead of asking for it?
- What I did
- Rebuilt the product screen around the evidence, and gave every alternative a location.
- Decision
- Every answer shows what was checked, which restriction, who it’s for and when.
Instead of just “Safe,” every answer now shows:
- What was checked
- Which restriction was used
- Who it applies to
- How recently it was checked
Alternatives became useful in the store too. Not “here is another safe product,” but “here is another safe product, and here’s where to find it.”


08Testing · Round 2
- Tasks completed
- 30/30
- SUS average · 79.2 checking, 68.8 route
- 75.0
- Observed errors, 4 of them on one control
- 9
The flow I had rebuilt around safety scored comfortably higher. The route, which I had already fixed once, was still close to the benchmark, and testing showed why: both route testers pressed Start shopping trip, landed on a list instead of the route, and stopped.
“It gave me the same screen twice, so I assumed I had done it wrong.”
Button
Was: Start shopping tripNow: Plan Route
The wording wasn’t the point. The button also had to actually open the route.
Trip card
Was: 5 items · 4 aislesNow: 5 items · 3 stops
The card now describes the trip, not the shopping list.


My first fix had addressed the finding, not the problem. A design fix is still a hypothesis until I test it.
- The question
- Did the fixes work, and what still breaks?
- What I did
- Tested the hi-fi prototype with 5 participants: 3 checking products, 2 walking the route.
- Decision
- Start shopping trip became Plan Route, and it now opens the route.
High fidelity · 5 participants · 3 checking products · 2 walking the route
09Refining
Every answer, one structure.
- The question
- Does Harbey check products differently depending on the result? One participant couldn’t tell.
- What I did
- Gave all three answers the same order: the allergen and who it’s for, the sources, who checked it.
- Decision
- Every answer names its source plainly: “Checked by Harbey”.
One participant couldn’t tell whether Harbey checked products differently depending on the result. So all three states now follow the same order, and name the source plainly.
- Allergen + who it’s for
- Sources
- Who checked it
Was: “Label information updated”Now: “Checked by Harbey”



A small moment
The screen that changed a decision.
One tester was moving fast and went straight for Add it anyway on a product that wasn’t safe. At the confirmation step they stopped, read it, and backed out.
What changed their mind was that the message named the person the restriction applied to, not just the peanut. It was the clearest sign that the confirmation step was doing real work, not decoration.
10Accessibility
Accessibility changed what the product says, not just how it looks.
- The question
- Can someone with low vision shop with Harbey without relying on the map or on color?
- What I did
- Wrote every location in words, gave each safety state text, an icon and a color, and annotated focus order and labels for the build.
- Decision
- When a warning opens, the safe option gets focus first.
The most important accessibility decision wasn’t a font size or a color. It was writing every location as Aisle → Side → Shelf height, and giving every safety state text, an icon and a color, so meaning never depends on color alone.

Verdict
- Label
- “Not safe. Checked against peanuts for Sam.”
- Role
- Status message, announced on load
Read as text, never conveyed by color or icon alone.
Safe Alternatives
- Label
- “Safe Alternatives”
- Role
- Button
First action after the evidence, so the safe path comes before the override.
Add Anyway
- Label
- “Add Anyway. Opens a confirmation.”
- Role
- Button
Focus moves into the dialog; Escape closes it and focus returns here.
Allergen information
- Label
- “Allergen information, collapsed”
- Role
- Disclosure button
Announces expanded or collapsed, on focus and after activation.

When a warning opens, the safe option gets focus first.
Dialog
- Label
- “This product may contain peanuts”
- Role
- Modal dialog
Focus is trapped inside. Escape closes and returns focus to Add Anyway. The heading is the dialog’s name.
Safe Alternatives
- Label
- “Safe Alternatives”
- Role
- Button, first in the dialog
Receives focus when the dialog opens, so the safe path is the default.
Add Anyway
- Label
- “Add Anyway. Adds a product that is not safe for Sam.”
- Role
- Button
The label states the consequence, since this is the only way past the warning.
I started out treating accessibility as something to apply to the interface. I finished understanding that it can change the content and the information architecture of a product.
11Website
Same experience, different context.
- The question
- What does the same service look like at home, before the trip?
- What I did
- Kept the structure and the words, and reworked the layout for planning instead of walking.
- Decision
- The website plans the trip; the route goes to your phone, or the list downloads by aisle.
I kept the structure and vocabulary so people could move between app and website without relearning anything, but I didn’t just scale the mobile UI up. The website is less about walking the store and more about planning the trip before you arrive.

No scanner
Scanning makes sense with a phone in an aisle, not at a laptop.
No walking route
Send the route to your phone, or download the list grouped by aisle.
More space
Dietary restrictions get room to be shown in full.
The outcome
What changed, what didn’t, and what I’d do differently.
Did it work?
Check it worked. Find it improved, but isn’t finished.
| Part of the task | What happened | How I know |
|---|---|---|
| Check it | The rebuilt safety flow scored 79.2, and a tester stopped before adding an unsafe product. | SUS and observation, round 2 |
| Find it | All route tasks were completed, but the route scored 68.8, close to the benchmark. | SUS and observation, round 2 |
| Low vision | Both low-vision testers found products on their own. Neither used the map. | What they said and did in testing |
| Overall | Usability score rose 13% (SUS 66.5 → 75.0), and observed errors dropped from 20 to 9. | Round 1 vs round 2, 5 participants each |
Outcome
The words did more work than I expected.
Both low-vision testers said Harbey gave them enough information to find products on their own. Neither of them used the map.
“Nobody has ever told me the shelf height before.”
What I got wrong
The mistakes taught me more than the finished screens.
I trusted the wrong number
Round 1 hit 94% task completion and I almost moved on. Behavior, errors and ease told the real story.
I over-corrected
To stop the route losing track of items I added counts, then repeated them three times on one screen. One clear count would have done it.
I gave the map the best spot
It sat on top of every route stop. Neither low-vision tester touched it, and one scrolled past it every time. Moving it is first on my next iteration list.
“The picture I cannot use is sitting on top of the words I can.”
What I learned
- 01Define success before designingSpeed isn’t always the metric. Ananya and Amari needed different things.
- 02Don’t let one metric tell the story94% completion looked like success. The behavior behind it said otherwise.
- 03Retest your own solutionsA fix isn’t done because it looks right in Figma. The route entry proved it.
- 04Accessibility starts with informationAisle, side and shelf height made Harbey more accessible, and more useful for everyone.
Next steps
Screen reader testing
Both low-vision testers used magnification. The annotations still need validating in a real build.
Understanding “Not confirmed”
Neither study produced this state. How do people read uncertainty, and what do they expect to do?
Completing the journey
Today it’s Find → Check. Next: Find → Check → Collect → Buy.
I thought I was designing a better way to find groceries. I finished it thinking about confidence.
Good UX isn’t getting the interface right the first time. It’s making something, finding where it breaks, understanding why, and being willing to change it.
— Iakov Grechukhin, UX researcher & designer

