2026

Checkout

A checkout built for the Amazon region

The Bemol checkout was still a webview while the backend had already outgrown it. This project puts on screen what the system was already doing, in a way the customer can follow.

The prototype tested in store · Phase 2

The file that ran on the phone during the test, in its day-two version, in Portuguese as tested. Besides the purchase, you can switch to pickup (Retirar), use the coupon LOJA or turn on the bonus. The data is fake and payment goes nowhere.

In 30 seconds

The problem
The app's checkout was still a single webview scroll with one generic delivery date, while SAP was already splitting orders across several origins. Pickup and delivery couldn't share a cart, and boat delivery was treated as an exception.
What I did
I designed the flow's architecture and split the rollout into two phases: first four steps without touching the visuals, then the full design system. Packages, in-store pickup and boat delivery now show up inside the flow. When Phase 2 became a native cart, I tested the prototype in store and adjusted it between one day and the next.
Where it landed
Phase 1 is live: the four steps cost no conversion, and once things settled, late deliveries dropped by 20%. In the Phase 2 test, six out of six people finished the purchase, with an ease score of 6.7 out of 7.
Company
Bemol Digital
Role
Senior Product Designer
Scope
Product Design
Channels
App & Web
Status
Phase 1 shipped · Phase 2 native, tested in store

01

Before: the old webview checkout

The old checkout was a single long scroll: a webview handling address, delivery options, payment and order summary with no step structure at all.

The starting point

  • Old checkout as a single long webview scroll

It showed one delivery deadline, calculated from a generic darkstore account rather than the real origin of each item.

The backend already knew how to split the order. The screen didn't.

02

Three problems the interface was hiding

  1. Problem 1

    One deadline for orders that were already split

    SAP was already breaking orders into sub-orders shipping from different origins, from distribution centre to pharmacy, store and marketplace, and about 1 in 9 orders ships in more than one package. The interface showed a single generic deadline, and the customer learned about the split only when the first package arrived alone.

  2. Problem 2

    Pickup and delivery couldn't coexist in one cart

    A customer picking up one item in store and having another delivered had no way to express that in a single flow. The selection hierarchy had to make both explicit without doubling the number of steps.

  3. Problem 3

    Delivery by boat treated as an edge case

    For remote customers, boat delivery is the only option there is. Treated as an edge case, it would have worked in the capital and failed exactly the people with no alternative.

03

The size of the problem

What the flow had to resolve, before any performance number. The postcodes served by boat are in Amazonas, Pará and Acre.

3
Methods in one cart
39
Postcodes served by boat
1 in 9
Orders split

04

The strategic decision: a two-phase rollout

I split the rollout in two: first the structure, then the visuals. Phase 1 broke the single scroll into steps without touching the design system. Phase 2 applies the full system, with what Phase 1 taught us in production. If both had shipped together, there would be no way to tell which change moved the numbers.

Phase 1: structural MVF

  • Checkout broken into four sequential steps

The step logic, still wearing the old visuals.

05

Phase 1 in production

Breaking the scroll into steps cost no conversion: cart to purchase on the site and abandonment on site and app stayed the same. Once things settled, late deliveries fell between 15% and 25% in two windows in a row, but operations also improved over the period, so that gain isn't the checkout's alone. The comparisons use 14-day windows against July.

−20%
Late deliveries, once settled
Stable
Checkout conversion

06

Phase 2: the design

With Phase 1 live, Phase 2 designed what it had only structured: packages, in-store pickup and boat delivery, already inside the design system.

The four steps, in the Phase 2 design

  • The four steps of the redesigned checkout: cart, delivery method, payment method and Pix payment

A mockup built from the design screens: cart, delivery, payment and Pix.

07

Embarcação is no edge case

E-Commerce Caboclo serves riverside communities across the Amazon interior, reachable only by boat. Deadlines there depend on river schedules, not road logistics.

Where the deadline comes from

  • The Port of Manaus, where boats depart to river communities

    The Port of Manaus, where boats depart daily to communities across the interior.

Delivery by boat, in the flow

  • Boat delivery option shown in the checkout

The interface had to say so plainly instead of hiding it behind a generic estimate. A customer who knows the boat leaves on Thursday can plan around it, and one who sees a vague range of days can't.

08

Phase 2 became the native cart

Midway through Phase 2 the scope grew: cart and delivery left the webview and became native app screens. The design changed with them. "Package" became "Delivery 1" and "Delivery 2", and the order summary became a fixed bar at the bottom. The end of payment still runs on VTEX, in a webview, which is why bonus and store credit still sit on the payment screen.

The August benchmark set the bar. Bemol's checkout took 6 required taps, an access code and the CVV again; Amazon's took 1 tap. In the native flow, the fingerprint comes in only at payment, Pix is already selected, the warranty shows up when the person decides to move on, and a fixed line at the bottom shows how much they're saving. Counted in the prototype with the same protocol, the path from cart to payment takes a similar number of taps, but required decisions drop from 4 to 1, the warranty, and typed fields from 3 to 0. That's the flow in the prototype at the top of this page.

09

Tested in store, fixed overnight

I tested the prototype with Linda Souza, from the Design team, at Bemol's Ponta Negra store in Manaus, over two days. There were six people who work at the store, aged 23 to 40, none of them from product or tech. Each one made a full purchase, picked a warranty and went through ten smaller tasks, like applying a coupon, changing the address or picking up in store.

Two sessions in the store

  • Christofer running a test session at a desk in the store, with the note sheet and the phone running the prototype
  • A participant, seen from behind, using the prototype on the phone while Christofer follows along on the note sheet

Two of the sessions at Bemol's Ponta Negra store: in each, the participant uses the prototype on the team's phone while I take notes. Faces are blurred.

The prototype went back and forth between AI and Figma. Claude Code generated the first version of the screens straight into Figma, but the AI output alone didn't get where we wanted, so we refined the design by hand. With the Figma file adjusted, Claude Code turned the screens into a single HTML file that runs offline. A hidden menu opened each task at its starting point, so every task started clean, and the end screen logged the task, the time and the number of taps. What the first day showed was fixed straight in the code that same night, and the new version was tested the next morning.

The analysis went through AI too. Each session was recorded with taps visible on screen. Gemini summarised every video and listed every interaction: the screen, the element tapped and whether the person hesitated. A script plotted those taps on the real position of each button in the prototype, which is where the heat maps further down come from. Where the summary and the note sheet filled in on the spot disagreed, the sheet won.

10

What the test showed

After each task, the person answered a single question: from 1 to 7, how easy was it? One is very hard and seven is very easy. It's the SEQ scale, common in usability testing. There were four people on day one and two on day two, with the prototype adjusted in between; with a sample that size, the difference between days works as a signal.

6/6
Completed the full purchase without help, from cart to payment
6.7
Average ease score for the full purchase, on the 1 to 7 scale
4.8 → 7.0
Score for switching delivery to in-store pickup: tabs on day 1, buttons on day 2 (Finding 1)

11

Three findings that changed the screen

  1. Finding 1

    Pickup was hidden in a tab

    Deliver and Pick up were tabs stuck to the header, and three of the four people on day one struggled to find pickup. One rated it 1 and went looking for it on the payment screen. They became segmented buttons, and on day two both people found them right away.

  2. Finding 2

    Once a store was chosen, the deadline vanished

    After choosing pickup, the screen showed only the store, and one person couldn't say when the product would be ready. "Within 4h" now stays on the delivery screen and in the store list. The current VTEX flow has the same gap.

  3. Finding 3

    Nobody opened the summary on their own

    That held on both days. When the moderator pointed it out, people used it to check the totals. The animated arrow on day two helped them recognise the bar, but nobody opened it unprompted. The recommendation was to keep the component and measure discovery in a future test.

Pickup, day 1 and day 2

  • Day-one delivery screen with Pick up as a tab and the chosen store without a deadline
  • Day-two delivery screen with Pick up as a segmented button and pickup within 4h on each delivery

On day one, pickup was a tab and each delivery showed only the store. On day two, it became a button and started saying when the order would be ready.

Tap map on the delivery screen

  • Heat map of taps on the day-one delivery screen, with Deliver and Pick up as tabs
  • Heat map of taps on the day-two delivery screen, with Deliver and Pick up as segmented buttons

Every tap from the recordings, plotted on the screen. With tabs, each person hesitated about three times here on average. With buttons, nobody hesitated.

12

What already worked

All six people found the savings on the green line in the bar, without opening the summary. The warranty, offered when moving on and priced by instalment, bothered nobody. Separate deliveries per package were the difference people mentioned most against the current app.

You sell it, I buy it if I want to.
Test participant, on the warranty (translated)

13

What didn't ship

  • Changing the pickup store after confirmation

    Blocked by the order service: once confirmed, the store is fixed. We set the expectation in the interface instead of promising something the backend couldn't honour.

  • Per-package delivery tracking in Phase 1

    The tracking service returned one status per order, not per sub-order. Showing separate packages without separate tracking would have created a question we couldn't answer.

  • Fully dynamic deadlines at cart level

    VTEX calculates by geographic polygon, and resolving that before an address exists wasn't possible. The deadline appears once the address does.

  • Store credit and bonus in the cart, next to the coupon

    The idea came up in the team, but while the end of checkout is a webview both stay on the payment screen. What the test proposed instead is having them on by default, with an option to use the whole balance.

  • Leaving only payment on the last screen

    We considered dropping delivery, details and summary from it. In the test, people scrolled to the bottom to double-check before paying, and fixed their delivery through the link down there. The review stayed.

14

Who did what

Christofer

  • Flow architecture and the decision to split the rollout in two phases
  • Step structure of the checkout, the piece that carried Phase 1 on its own
  • Interaction design for package splits, pickup and boat delivery
  • KPI definition together with data, before launch
  • Prototype, script and analysis of the native cart's in-store test

With the team

  • Design system components, built with the DS team
  • VTEX and SAP integration, with engineering
  • In-store test sessions, run and observed together with Linda Souza

15

Reflections

Three things that stayed with me from this one.

The picture I keep is a customer in a riverside community you can only reach by boat, finishing a purchase and getting it days later by boat. The checkout was built with that person in mind.

The change that mattered most barely shows in the visuals: breaking one scroll into four steps. It shipped before a single new component existed.

Phase 2 left me with a way of testing. Because the prototype was a single file, I could fix overnight what the first day showed, and the second day was already testing the fix.