Scoring storefront fidelity, section by section
"Pixel-faithful" is easy to write on a landing page. It only means something when there's a number behind it, and a gate that stops low-fidelity work from shipping. Here's how the judge works.
Any tool can claim it clones a store "faithfully." The hard part isn't the claim. It's proving it, every build, without a human eyeballing each result. If you can't score fidelity, you can't gate on it, and "brand-true" is just marketing. So the last stage of a GemBoss build is a judge: it scores what was produced against the reference, and decides what's allowed to ship.
Score sections, not pages
The most important choice is the unit. Scoring a whole page gives you an average that hides the problem. A store can be "88% right" overall and still have one broken hero, and the hero is the section that decides whether a shopper stays. Averages launder the weak link.
So GemBoss scores section by section. Each captured-and-rebuilt section gets its own grade against its counterpart in the reference. A weak header can't hide behind a strong footer; it surfaces as a low score on that specific section, which is exactly the thing you'd want to fix first.
An average launders the weak link. Section-level scoring surfaces the one block that's actually costing you.
Two lenses: structure and vision
Fidelity has two failure modes, so the judge looks through two lenses.
Structure. Did the rebuild preserve what the section is? The right number of columns, the right information clusters (eyebrow, heading, body, CTA), the right container rhythm and role. A section can be visually close and structurally wrong. A three-up grid rebuilt as a two-up will "look fine" in a thumbnail and convert differently in real life.
Vision. Does it actually look like the reference? This is a perceptual comparison. A visual-similarity measure over the rendered section, plus a perception pass that reads the section the way a person would and flags the ways it drifts. Type weight that's off, spacing that's collapsed, an image treatment that lost its crop.
A section has to satisfy both to earn a high score. Structure without visual match is a skeleton; visual match without structure is a screenshot.
The gate
Scores only matter if something acts on them. Each section's grade is checked against a bar. Sections at or above it pass. Sections below it are flagged. Surfaced for a rebuild or a human look, and are not quietly presented as finished. That's the difference between a score and a gate: the gate can say no.
This is why a GemBoss store can carry a 90+ brand-match number and have it mean something. The number isn't a vibe printed on a card; it's the floor the gate enforces before a build is called done.
Grading the grader
A judge that silently gets worse is more dangerous than no judge at all. It keeps stamping "90+" on work that's slipping. So the scoring system is itself held to a regression net: a fixed set of reference sections with known-good outcomes that the perception-and-scoring stack is run against, so a drop in the grader's own accuracy shows up as a failed check rather than as quietly worse stores.
It's the same discipline all the way down. Capture is gated by fidelity; the two renderers are gated by parity; and the judge that runs those gates is gated by its own eval set. Each claim the product makes has a check standing behind it that can fail loudly.
Why this is the whole game
Generative store builders are easy to demo and hard to trust, because the interesting failures don't show up in a cherry-picked screenshot. They show up on the fifth store, in the one section nobody looked at. Section-level scoring, two lenses, and a gate that can say no are how "faithful" survives contact with real, messy reference sites. It's less glamorous than the generation, and it's the reason the output is worth shipping.
Watch the judge score a real build
Bring a store you admire. We'll capture it, score every section, and show you what passes the gate.
Request a demo →