AI Website Builder Hands-On Test: One Prompt Generates a Landing Page, But Three Pitfalls Lurk

We run MODEL, a small weekly magazine dedicated to reviewing AI tools. Our principle is simple: anything we haven't personally tested gets labeled "untested." What follows is a real test log—the good and the bad.

Test Setup: One-Sentence Prompt, Zero Cost

We tested Bolt.new's free plan at zero cost throughout. The prompt we entered, word for word, was:

a simple landing page for a weekly AI tools newsletter with an email signup box

The goal was clear—a "simple" landing page with an email signup box. This is also a real need for many indie developers and small teams: no complex features, just something quick to ship that collects emails.

What Actually Happened: Fast, Functional, But More Than "Simple"

From hitting enter to the first version being generated took about 2.5 to 3 minutes. We entered a test email in the signup box, and the page returned "Subscribed! You're on the list!"—the feature genuinely worked.

But what it generated went well beyond "simple": a hero section, a stats bar, a features grid, a sample newsletter, an FAQ, a second call-to-action button, and a footer. On top of that, it automatically created a subscriber database table (Bolt Database, powered by Supabase) and pulled in a stock image. The whole process ran without errors.

From an engineering standpoint, this was a complete "prompt to working product" demo. The problems were at the content layer.

Bolt.new is an AI full-stack app generator from StackBlitz, built on WebContainers technology that runs a complete Node.js environment in the browser—no local installation required to generate, preview, and deploy apps in real time. After users enter a natural language description, it automatically selects a tech stack (usually React + Vite + Tailwind CSS), generates code, installs dependencies, and starts a dev server. Its free plan is billed by tokens, but the user-facing interface typically only shows a remaining-quota percentage or rough status rather than precise per-call token breakdowns, making it hard for developers to accurately assess the cost of a single generation. Bolt also integrates one-click database and third-party service connections, further lowering the barrier from prototype to working product—but also making the volume of generated content and resource consumption harder to estimate.

Three Pitfalls to Watch For

Pitfall 1: Fabricated Social Proof

The page boldly claims "12,000+ readers." That number has no source. If you publish it as-is, you're lying to your visitors. AI website builders tend to fill empty space with "confidently fake numbers"—this was the most alarming finding of this test.

Pitfall 2: Promise of Unbuilt Features

The copy promises a welcome email will be sent, but there's no email-sending logic in the backend at all. After subscribing, users only see a success message and receive no follow-up email. This "copy-functionality disconnect" is quite insidious in AI-generated products because the page looks complete and credible.

Pitfall 3: "Simple" Isn't Simple, and Consumption Is Invisible

The extra sections and database table all consume tokens on the free plan, which doesn't display exact token usage. You asked for "simple," but the AI gave you a "rich" page—not a bad thing, but the cost isn't transparent.

Supabase is an open-source Firebase alternative offering hosted PostgreSQL databases, authentication, file storage, and edge functions, with auto-generated REST and realtime subscription APIs. In AI website builders, Supabase is often the default "backend-as-a-service" option: the tool simply calls its management API to create tables, configure row-level security policies, and generate frontend data-access code—no manual server setup needed. This integration makes "one-sentence generation of a page with a database" possible, but it also means the generation process implicitly creates cloud resources. For free plan users, the creation of these resources and subsequent reads/writes count toward platform quota or generate separate usage on the Supabase side, and the two platforms' billing metrics aren't aligned—further aggravating the cost transparency problem.

Conclusions and Practical Advice

Bolt.new is great for getting a working, good-looking signup page in minutes. But before publishing, you must review every line of copy, because AI website builders fill empty space with confidently fake data.

Practical tip: add constraints directly in your prompt, e.g., "don't invent stats, testimonials or features that aren't built." This one sentence saves a great deal of cleanup work later.

In AI-assisted development, prompt constraints (negative constraints) refer to explicitly stating in the instructions "what not to do," e.g., prohibiting fabricated data, banning descriptions of unimplemented features, limiting the number of sections. Compared with positive descriptions, negative constraints more effectively compress the model's room to improvise, because large language models under open instructions tend to "complete" a page that looks whole—including filling in statistics, customer testimonials, and feature promises. Writing these prohibitions into the prompt is like pre-installing a content review checkpoint for the generated result, significantly reducing pre-publish cleanup work. This technique also applies to copywriting, code comments, and API documentation—it's one of the lowest-cost quality control measures available when using generative tools today.

What We Couldn't Test

We also tried Synthesia's free plan, but registration got stuck at the email verification step, so no video was produced. We labeled it "untested" rather than guessing at a conclusion. Every week we publish similar hands-on logs, with our testing methodology and labeling rules publicly documented at readmodelmag.com/how-we-test.

This test also reveals a more general problem: the real risk of AI website builders isn't code quality—it's content truthfulness. Working code and good-looking pages are the lower bound of what these tools can do; whether the copy is credible and the features are delivered is the upper bound users need to police. Treat AI as an assistant that quickly builds the skeleton, not as a content author that fills in the flesh—that's probably the most pragmatic way to use it at this stage.