Case study

Lina

A concept-validation site for an AI-concierge idea. Before writing a line of the product, I built the apparatus to find out whether anyone wants it: two full landing-page directions, an A/B split, a real signup pipeline and event tracking — on Cloudflare’s free tier, no framework, no build step.

Cloudflare WorkersA/B landingNo frameworkBilingual RU / ENCanvas animationAnalytics-ready
The problem

Building a product before anyone has asked for it is the most expensive mistake there is.

But a validation page only produces a real signal if it reads like a real product — and if you can tell which pitch, not just which palette, people respond to. Lina tests both.

What’s on the page
Two complete directions

/daylight and /midnight are full landing pages, not a theme toggle: different palette, different type pairing, different headline and positioning line. The root path assigns a visitor to one and remembers it.

Bilingual, no reload

RU / EN swap client-side from an in-page dictionary — auto-detected from the browser, remembered per visitor, applied by data-attribute with no second URL and no round-trip.

A signup pipeline that behaves

Hero and footer forms, a bot honeypot, client-side validation, a real POST to a Worker endpoint, dedup handling and a graceful offline message. On success both forms collapse to a persisted “you’re in” state.

Instrumented for a decision

Variant, UTM and referrer ride every event — page view, signup, language switch, errors — into GTM’s dataLayer and Plausible. The analytics tag is a one-line paste; nothing tracks until it’s added.

A living background (midnight)

The dark hero runs a neural-network canvas: node count scales with viewport width, links form by proximity, it’s DPR-aware, pauses off-screen and switches off entirely under reduced-motion.

Real-product polish

Open Graph and Twitter cards with a distinct share image per variant, favicons at every size, a theme-color per direction, focus-visible rings and an aria-live status region.

Worth a closer look
Details
The honeypot is invisible and kind

A hidden field that bots fill in and humans never see. Trip it as a human and you still get the success screen — never an accusation.

Two type systems, on purpose

Daylight pairs Bricolage Grotesque with a Source Serif body — editorial, warm. Midnight runs a Newsreader mark over Hanken Grotesk — colder, more technical. The fonts carry the pitch.

Degrades on every axis

Reduced motion stops the canvas. Offline returns a retryable message, not a dead end. The page is one static document, so it renders before any script runs.

Honest about itself

The footer says so plainly: “This idea is being validated. Your interest helps us decide whether to build it.” No fake countdown, no invented user count.

Architecture
One Worker, no framework, no build

One Cloudflare Worker on the free tier. Each variant is a single self-contained HTML document — inline CSS, inline vanilla JS, no framework, no bundler, no npm.

  • The root path assigns a variant and redirects; the choice sticks per visitor.
  • /api/interest is a POST-only Worker route that records the email with its variant and UTM context.
  • i18n is a dictionary inside the page — no locale files, no server round-trip.
  • Google Fonts is the only external dependency; everything else is inlined.
  • Fits entirely in the free tier and needs no custom domain.
What I took from it

The point isn’t the page — it’s the decision. Building the validation apparatus properly — two real directions, a working pipeline, honest instrumentation — costs a day and can save a month spent building the wrong thing. And a page that degrades gracefully and respects the visitor is itself a small demo of how the real product would be built.

Validating an idea, or building the real thing?