Configurator ConversionProof of concept
November 2026

One night, one fictional brand

One watch. Four ways to order it. Which one sells more?

We invented a watch brand, modelled its watch, and built four complete order flows around the same product. Not to show off 3D. To show that the way you let people configure is a choice, and that the choice can be tested.

Tidewright T-39 in steel with a blue sunburst dial
Live 3D. Drag sideways to turn it.
1product, modelled in Blender, live in the browser
4complete order flows on the same options
7choices per watch, one of them required
0personal details asked. Every flow stops before payment

The promise

More orders from your configurator, proven in an A/B test.

Audit the flow. Walk every step the way a buyer does, with evidence for each finding.

Redesign the weakest step. One change at a time, so you know what worked.

Test it against what you have. The test decides, not the designer.

Assumption, stated as one: which flow wins differs per product and per buyer. That is why this page shows four candidates and no predicted uplift.

Try them on your phone

Same watch, same options, four flows

Each one is a bet about how people decide. Open them, build a watch, and notice where you hesitate.

Flow A: one choice per screen Flow A

Step by step

The bet:
one decision at a time keeps people moving.
The risk:
seven screens before you see the total add up.
Open flow A Flow B: everything on one page with a running total Flow B

One page

The bet:
seeing everything at once builds confidence and speed.
The risk:
a long page hides the one choice that is required.
Open flow B Flow C: tap the watch itself to change a part Flow C

Tap the watch

The bet:
the product is the menu. Touch what you want to change.
The risk:
options you cannot see on the watch are easy to miss.
Open flow C Flow D: three questions, then a suggested watch Flow D

Guided

The bet:
most buyers want a good suggestion, not twenty decisions.
The risk:
a wrong suggestion costs more trust than no suggestion.
Open flow D

An agent walked all four

What the audit found

Two scripted buyers, the key screens photographed, every tap counted. These numbers describe the interface, not people. They are the reason for a test, not a replacement for one.

14taps to order in flow A, step by step
8taps in flow B, one page, for the same watch
20taps in flow C, tap the watch
14taps in flow D, guided

Three flows noticed a missing wrist size only at the very end

The repair took 3 to 4 extra taps, against 1 in flow A. Found, partly fixed, re-run.

In flow A the error message was off screen

The flow stopped correctly, but the buyer saw nothing happen. Found on a screenshot, fixed, re-run.

A second agent found what the first one missed

It opened the one panel the audit never opened, and the page froze. Fixed, and now part of the test.

First test we would run: B against A

The classic wizard against the single page, 14 taps against 8. A hypothesis, with no predicted uplift.

Read the full audit, with before and after screenshots

The product

Tidewright T-39

A fictional build-to-order watch from Hong Kong. Own design, own name, modelled from nothing in ten rounds. The stills are rendered in Blender, the 3D above is the same model in your browser.

Macro of the blue sunburst dial with applied indices Steel case, blue dial, steel bracelet Gold case, green dial, stone-set bezel, stitched leather Black case, salmon dial, rubber strap Rose gold caseback with a personal engraving

A second model, stills only

Tidewright T-36

A dress watch from the same build script, changed in 40 places: slimmer case, leaf hands in blued steel, small seconds. It is not in the configurator on purpose. A new option group would change every number in the audit, so it stays a picture until the flows are tested.

Close-up of the opaline dial with blued leaf hands and a small seconds sub-dial Dress watch in rose gold with an opaline dial on stitched leather Dress watch in steel with a black dial

Making of

The night, start to finish

From an empty scene at 22:29 to an audited set of order flows at 23:34. No sound needed.

The process cut for social, 45 seconds

Not the watch, the work: the brief, Blender by script, live 3D, four flows, the audit agent, the Figma library, twenty banners and the log. Vertical, with a 4:5 version for feeds.

The full making-of, 76 seconds

The walkthrough, 2 minutes 19

All four flows played start to finish, then the audit and the first test.

The teaser, 30 seconds

How it got there

Ten rounds, nothing hidden

Build, render, look, criticise, improve. Every round was saved, including the ones that went wrong.

Round 1
1. First build. A studio light cuts the dial in half and the case looks like plastic.
Round 2
2. Domed crystal, finer bracelet, end links. The lugs still look stuck on.
Round 3
3. Light moved. The sunburst dial shows for the first time.
Round 4
4. Lugs now grow out of the case. Metal still too white.
Round 5
5. Lighting rebuilt with thin edge lights. It reads as steel.
Round 6
6. First variant. Leather looks like plastic, stones look like holes.
Round 7
7. Stitched leather and cut stones. The lug tops are still wrong.
Round 8
8. Calmer reflections on crystal and steel.
Round 9
9. Caseback engraving added, bracelet gets its own light.
Round 10
10. The real cause of the white lugs found and fixed: an empty material slot.

From live code to Figma

The design system, ported in one pass

The order flows were built in code first. Then their tokens and core components were ported into a Figma library: 39 variables with their CSS names, 8 text styles, 6 components with variants. Every fill, stroke, radius and gap is bound to a variable.

Figma overview: a phone screen built from library components next to the button, option row, swatch tile and summary row states

The gaps, stated plainly

Not ported yet: bottom sheet, hotspots, question card, engraving field, chips and header. Swatch colour is an override, not a property. One colour mode only. The Figma file is private and shared on request.

Brand assets in bulk

Twenty banners from one template, checked on the pixels

Four builds, five formats, nine brand rules. Contrast is measured on the rendered image, and what fails is corrected and rendered again.

Landscape banner: Built to order. Not to stock. Square banner with the gold and green watch Square banner with the dial close-up

See all twenty, with the rules and what went wrong first

Something to pass around

Drop-off, the game

A small watch dodging labels such as Hidden fee and Pop-up Under a minute

You are an order. Try to reach checkout.

Seven steps, a patience bar, and every bad configurator pattern coming at you. It makes the point faster than a slide does.
Play it

Working title: Configurator Conversion

Three names to react to

Proposals only. None has been checked for trademarks or domains.

Buildrate

Names the number you improve: the share of started builds that become orders. Easy to say in a boardroom, hard to mistake for a 3D studio.

Stepwise

Names the method: every step of the configurator measured, one step improved at a time. Friendly, less ownable.

Order Bench

Names the place: a workbench where order flows are built and tested against each other. Fits made-to-order brands and craft.