The rebuild test
Can an AI assistant rebuild a Kiekkist design from a screenshot, and what goes missing? We tried it on 30 October 2026 with three of our own designs. Because they are our designs, the rebuild is allowed.
Result
The look came across. The quality, the rights and the scope did not.
First screen: original and rebuild
Desktop, 1440 px. We captured the rebuilds the way a visitor sees them, with their Google fonts loaded.
-
Fieldnote
Original Check 99 
Rebuild Check 45 -
Roast Ledger
Original Check 99 
Rebuild Check 81 -
Specimen
Original Check 99 
Rebuild Check 36
Note: the Fieldnote and Roast Ledger rebuilds show photos the agent cut out of our screenshot. The photos belong to our designs; we only show them small here.
Method
Selection
Three gallery designs: Fieldnote, Roast Ledger and Specimen.
Input
One full-page screenshot of the home page each, at 1440 and 390 px.
Assistant
A fresh session of a current AI assistant for each, same environment. Task: “Rebuild this website as faithfully as you can … make it a good, production-ready website.” No access to sources, skills or rules. Home page only.
Measurement
Kiekkist checker (rules 2026-10-30.3, profile BFSG, mobile lab run on a Moto G profile, median of 3 runs) and kk compare 2026-10-29.1. Original and rebuild treated the same, both served from a local HTTP server.
The numbers
| Design | Original | Rebuild | Hard errors in the rebuild, by kind | Similarity (kk compare) |
|---|---|---|---|---|
| Fieldnote | 99/100 (A) 0 hard errors | 45/100 (D) 15 hard errors |
| too similar blink 11/22, same first screen, 54 % same H2 |
| Roast Ledger | 99/100 (A) 0 hard errors | 81/100 (B) 8 hard errors |
| borderline blink 11/13, same H1 |
| Specimen | 99/100 (A) 0 hard errors | 36/100 (E) 11 hard errors |
| too similar blink 16/17, same first screen, 100 % same H2 |
Measured on 30 October 2026. “Hard errors” are findings the checker counts as errors (not warnings). One kind can occur more than once, for example on several viewports.
What it means
The look is copyable.
kk compare rates two of three rebuilds as “too similar” and one as borderline. A convincing home page comes quickly. If a look is all you want, you don’t need us.
The quality doesn’t come along.
All three rebuilds have hard errors and the originals have none: serious axe violations, fonts from Google Fonts (the visitor’s IP address goes to a third party; a Munich court ruled this a breach without consent, LG München I, 3 O 17493/20), text spacing under WCAG 1.4.12 in all three, no reflow at 320 px and sideways overflow in two, a missing reduced-motion variant in one.
The rights don’t come along.
Two rebuilds simply cut the photos out of the screenshot: someone else’s photos without a licence, at screenshot resolution only.
The scope doesn’t come along.
The rebuild is one page. The original is 9 to 10 pages, tokens (DTCG, OKLCH), DESIGN.md, 3 social tiles, 6 slides, exports to Astro, React and WordPress, check reports and a licence.
We draw no conclusion on load time.
The checker blocks third parties in the lab. So the rebuilds never loaded their Google fonts and look faster in the lab (LCP 0.8 s to 1.7 s against 2.2 s to 2.6 s for the originals).
Limits
- One assistant; others may do differently.
- Three designs, one model, one prompt.
- Lab load time is not comparable, because third-party fonts were blocked.
- A user who explicitly asks for accessibility, self-hosted fonts and their own photos gets better rebuilds. That is exactly the point: you have to know what to ask for, and be able to check it.
- Automated checks cover only part of WCAG 2.2.
Run the test yourself
The checker is open source. Run it on your own rebuild, or on any other page.
About the checker Proof: 36 websites measured
What we won’t do
- use a weak prompt so the rebuild looks bad
- show only the runs that suit us
- have other people’s websites rebuilt