Is stripe.com vibe coded?

Reading this score

stripe.com landed on 15/100, which puts it in the hand-written band. That number is not a percentage and not a probability — it is strength of evidence: 38 points of tells, damped by craft, less 38 points of deliberate decisions, mapped through one logistic curve.

What pushed it up

• Large language model house style +16 — 2 stock LLM marketing phrases.
• Statistics nobody measured +13 — Round, unsourced, and identical across thousands of sites. Nobody measured any of it.
• Dependency sprawl +8 — 3 third-party asset hosts and 79 script files. Each sub-problem got whatever library was most popular in the training data, and nothing was ever consolidated.
• Fixed pixel widths that will overflow a phone +8 — Hard pixel widths wider than a phone. Layouts break at sizes nobody prompted for, because desktop was the only viewport that got looked at.
• The rule-of-three cadence +6 — 14 triplet lists and 0 “not just X — it's Y” constructions. Rhythm borrowed from marketing prose, with none of the content.
• The database is addressed straight from the browser +6 — Supabase is queried directly from the page, so row-level security or rules are the only thing standing between a visitor and the table. This is not proof they are off — it is the reason to go and check, because "allow all" is what makes the demo work.
• A bento grid because bento grids exist +5 — Asymmetric tiles used as a style rather than because some tiles matter more. If shuffling them changes nothing, the grid is decoration.

What pulled it down

• Deliberate social preview −9 — Open Graph title and image were both authored.
• Hand-written ARIA −7 — 143 explicit accessibility attributes.
• Structured data −6 — JSON-LD is present — somebody thought about search.
• Semantic landmarks −6 — 5 landmark elements instead of a pile of divs.
• CSS custom property system −6 — 206 CSS custom properties — a real token layer.
• Canonical URL −4 — Canonical link is set.

Craft we looked for and did not find

These are the actionable half of the report. None of them are hard, and every one of them would have moved the number.

• Skip-to-content link
• prefers-reduced-motion honoured
• Self-hosted typography
• Hand-authored stylesheet
• No utility-class soup

What this scan cannot see

One request, no JavaScript executed. So everything below is invisible to us, and a page can be excellent or awful at all of it without moving the score:

• Domain registration age — needs WHOIS.
• Whether the nav's /about and /pricing actually resolve — needs a crawl, not one fetch.
• package.json dependency count, orphaned code, clashing abstractions, test coverage — needs the repo.
• Backend security shortcuts: string-built SQL, verify=False, missing auth on endpoints — needs the source.
• Whether row-level security or Firebase rules are actually switched on — needs the project.
• Console errors, hydration warnings and runtime 404s — needs a browser executing the page.
• Real Lighthouse numbers and horizontal scroll at 375px — needs a rendered viewport.
• Whether a JS-handled form posts anywhere — needs a submitted request.