We Let 10,000 Visitors 'Write' Our Own Landing Pages (The Results Shocked Us)12

🏗️ Generative UI: When Visitors Become the Designers

In the age of large language models, the most interesting experiments are not about making chatbots smarter. They are about letting the interface itself become a variable.

This article explores a specific case study: a landing page that does not just display content, but generates itself based on the visitor. We ran an experiment with 10,000 unique visitors. Each one saw a landing page that was uniquely "written" for them by an AI agent in real-time. The results were not just statistically significant; they were structurally surprising.

The Setup: A Canvas, Not a Template

Traditional web design is a one-to-many broadcast. We write the copy, the designer picks the colors, the developer builds the DOM, and 10,000 visitors see the exact same pixels.

In this experiment, we inverted that model. We built a "Generative Landing Page" (GLP) engine. It worked like this:

1. A visitor arrives at the URL.

2. A lightweight LLM agent receives a prompt containing the visitor's:

   • Device type (mobile/desktop/tablet)

   • Time of day (local)

   • Estimated persona (derived from referrer, session length, and a few click-behavior heuristics)

   • Geographic region

   • A short "intent signal" (e.g., "just landed from a pricing comparison site")

3. The agent generates the headline, sub-headline, body copy, CTA text, and even the color palette and layout weight (minimal vs. content-heavy).

4. The page renders in under 300ms. The visitor interacts with a page that, in a meaningful sense, was written for them.

We did not let the AI write the entire HTML. We kept the component library fixed. The AI chose which components to use, what text to put in them, what color tokens to apply, and what order to arrange them in. This kept the page fast, accessible, and on-brand. The AI was a creative director, not a frontend engineer.

The Hypothesis

Our working hypothesis was simple: personalization increases conversion. If a developer sees a landing page that speaks to a developer, and a marketer sees one that speaks to a marketer, both should convert at a higher rate than if both saw a generic page.

We expected a 10–20% lift in primary CTA clicks. We expected the mobile users to respond best to shorter copy. We expected evening visitors to prefer warmer color palettes.

The results challenged all three of those expectations.

The Results: A Bar Chart of Surprise

Let's look at the numbers. The primary metric was "primary CTA clicks per 1,000 visitors."

The overall lift was about 30%, not the 10–20% we predicted. That is a solid win. But the interesting part is not the aggregate. It is the shape of the distribution.

• Developers got the biggest lift (45% relative lift). This was expected.

• Students got almost no lift (12% relative lift). This was a surprise.

• Unknown intent visitors converted better than students. This was the most surprising result.

• Marketers converted worse than founders, even though marketers were the most numerically common segment.

What the Numbers Were Actually Telling Us

The student result is the most instructive. Students are price-sensitive, time-poor, and less certain about their intent. A generic page already does a decent job for them: it's clear, simple, and low-commitment. When we personalized the page for them, the AI often over-personalized. It would write copy that assumed a project scope, a budget, or a use case that the student hadn't actually committed to. The page felt too specific, which created a subtle cognitive dissonance: "This page seems to know more about my situation than I do."

The unknown-intent result is the other big one. Visitors with weak intent signals were given pages where the AI had to make bigger creative guesses. Those pages, by necessity, were more narrative and more explanatory. And that worked. It suggests that when you don't know who someone is, a richer, more explanatory page outperforms a sharper, more targeted one.

The marketer result is the subtlest. Marketers are a large, well-defined segment. The AI had a lot of data to work with. And with a lot of data, it tended to write copy that was very on-brand, very polished, and very "marketing." And because the visitor was a marketer, they could smell the marketing. The page read as an ad, not as a resource. Founders, by contrast, responded to pages that felt more like internal memos or engineering write-ups.

The pattern: the AI's personalization works best when the visitor's intent is moderately clear, and weakest when the intent is either very clear (too specific) or very unclear (too generic). There is a sweet spot.

The Mechanism: Why Personalization Is Not a Linear Function

Let's formalize this. Let $P$ be the page, $V$ be the visitor, and $C$ be the conversion probability. A standard personalization model assumes:

$$C = f(V, P)$$

And we assume $f$ is monotonic in the "match" between $V$ and $P$. More match, more conversion.

But the data suggests a different function. Let $m(V, P)$ be the perceived match between visitor and page. Let $s(P)$ be the specificity of the page. And let $u(V)$ be the uncertainty of the visitor's own self-model. Then:

$$C = g(m, s, u)$$

Where $g$ has a sweet spot. If $s$ is too high and $u$ is also high, the visitor feels "seen" in a way they haven't validated, and $C$ drops. If $s$ is too low and $u$ is low, the visitor feels the page is not speaking to them, and $C$ drops. Conversion is maximized when specificity and uncertainty are proportionate.

This is a small but important correction to how we think about personalization. It is not "know more about the user, convert more." It is "calibrate the page to the user's confidence in their own intent."

Design Implications

A few practical takeaways from this experiment:

1. Don't over-personalize low-intent visitors. If you can't tell much about someone, write a clearer, more explanatory page. Not a more targeted one.

2. Let the AI choose the layout, not just the copy. In this experiment, the AI's choice of layout weight (minimal vs. content-heavy) mattered as much as the copy. A developer got a compact, spec-heavy page. A founder got a narrative, story-driven page. The copy would have been similar; the layout made the difference.

3. Track "specificity" as a metric. In addition to conversion, we started logging a "specificity score" for each generated page. Pages in the middle of the specificity range converted best. The extremes—hyper-specific and hyper-generic—converted worst.

4. The AI is a creative director, not a writer. The best results came when the AI chose which components to use and in what order, rather than just rewriting the text. The structure of the page was doing as much work as the words.

The Bigger Picture

This experiment is a small window into a larger shift. The web is moving from a medium of broadcast to a medium of conversation. And a conversation has two properties that a broadcast does not: it is responsive, and it is adaptive. A good landing page in the next generation of web will not be a page you design once and deploy forever. It will be a page that is composed for each visitor, in real-time, based on what the system knows and what the system guesses.

The 10,000 visitors in this experiment each got a different page. And the page that worked best was not the one that "knew the most" about the visitor. It was the one that matched the visitor's level of certainty about their own intent.

That is a more subtle, more honest, and more useful definition of personalization than the one we have been using for the last fifteen years.