We Let AI Design Our Landing Page — Conversion Rate Jumped 340% in a Week

By Dr. Elias Thornwood

The Experiment

Last month, our product team decided to stop guessing. For years, our landing page design process followed a familiar rhythm: a designer drafts a mockup, stakeholders review, feedback loops iterate, and the page ships. Then we convert. Or don't. The cycle repeats.

We wanted to know something most teams assume but rarely test: What if the AI does the designing, and we just curate?

The result surprised even us. Our landing page conversion rate went from 2.1% to 9.3% within seven days — a 340% relative jump. Not a small A/B test nudge. A genuine, sustained, repeatable improvement.

This article walks through what we did, what the AI actually produced, where it stumbled, and the principles that made the difference.

Setting Up the Baseline

Before touching any prompts, we needed a clean baseline. Over three weeks, our existing landing page collected:

  12,400  sessions
    262    signups
     2.1%  conversion

Traffic sources were stable (68% organic, 22% paid, 10% referral), so we could attribute the lift to the page itself. We instrumented with standard funnel events: page load, hero view, form start, form submit, success.

The Prompt Architecture

We didn't ask the AI to "make a nice landing page." That's a designer's prompt, not a researcher's. We structured the task in four layers:

Layer 1 — Audience model. We fed the model a compressed ICP (ideal customer profile): roles, pain points, current workarounds, and the specific job-to-be-done.

Layer 2 — Conversion constraints. We told it the single primary action (start free trial), the secondary action (book demo), and a hard rule: no more than two CTAs above the fold.

Layer 3 — Voice and tone. Plain, concrete, first-person-plural. No "seamless" or "cutting-edge." If a sentence could appear in a SaaS brochure, rewrite it.

Layer 4 — Output format. Standard markdown with section headings, copy blocks, and a separate layout spec. No HTML. No commentary. Just the page.

Each layer was a distinct prompt. We iterated the audience model twice before the output quality stabilized.

What the AI Got Right

Three things stood out in the first usable draft.

1. Inverted the headline logic. Our old headline led with the product name. The AI's version led with the outcome:

"Ship your next release 4x faster — without hiring a platform team."

That's a claim about the user's life, not our feature list. Small change, but it reframes the entire page.

2. Killed the feature grid. The old page had a 12-tile feature grid below the fold. The AI proposed three problem-solution pairs, each anchored to a specific workflow. Users scan vertically; they don't read grids.

3. Made the proof concrete. Instead of "Trusted by 200+ companies," the AI wrote:

"Median customer sees first deploy in 11 minutes. P95 onboarding time: 3 days."

Numbers with units beat logos. Logos say we're popular; numbers say we're fast.

Where It Stumbled

Not everything was gold. Two failures were instructive.

Emotional flatness. The AI's copy was precise but sterile. We spent about four hours of editing to add rhythm — short sentences, a couple of rhetorical questions, one line of light humor. The AI writes like a good consultant; we had to teach it to write like a good writer.

Understated risk-reversal. The old page buried the free-trial terms at the bottom. The AI moved them up but kept them generic. We rewrote the trial line to be specific: "14 days, full access, no credit card, cancel in one click." Specificity is trust.

The Iteration Loop

The workflow that worked looked like this:

  1.  Prompt  →  AI draft  →  Team review (30 min)
  2.  Edit copy in markdown
  3.  Re-prompt with edits as constraints
  4.  Repeat until the copy reads like us
  5.  Hand to designer for layout (AI gave the spec,
      designer handled spacing, type, motion)

Total time from prompt to shipped page: four business days. Our typical cycle was three to five weeks.

The Numbers

Seven days after launch, the new page outperformed the old on every metric:

The conversion jump wasn't uniform. It was driven mostly by the hero section and the first problem-solution pair. The lower sections performed comparably. In other words, the top of the page did most of the work, which is exactly what a good landing page should do.

What Generalizes

A few principles that we'd apply to any team considering this:

• Treat the AI as a strong junior copywriter, not a designer. You still curate, edit, and own the voice. The model gives you 80% of the structure in minutes; you supply the 20% that makes it feel human.

• Constrain the output. Open-ended prompts produce open-ended pages. Specific CTAs, specific audience, specific tone — constraints are what make the draft usable.

• Keep the baseline honest. If you don't know your current conversion rate, you can't measure the lift. Instrument before you experiment.

• Let humans handle the visual layer. The AI gave us a layout spec — hierarchy, section order, copy placement. A designer translated that into pixels. The division of labor matters.

• Expect to edit for rhythm. LLMs optimize for clarity. Writers optimize for clarity and cadence. That second axis is where brands live.

A Note on Causation

A 340% jump in a week is impressive, but a single landing page is not a controlled experiment. Seasonality, traffic mix, and the novelty effect all contribute. We're running the old page again in a separate traffic slice to confirm. The direction of the lift is robust; the exact magnitude is still being pinned down.

What we can say: the AI didn't design the page. It proposed the page. The team selected, edited, and shipped. That's the right mental model. The model is a very fast, very consistent brainstorming partner that never gets tired, never defends a bad choice, and never says "let's just keep the grid."

The Bigger Picture

Landing pages are a small artifact, but the pattern is general. Anywhere we have a structured, language-heavy, goal-oriented task — onboarding copy, support macros, release notes, email sequences — the same loop applies. Prompt, constrain, draft, curate, ship.

The question for most teams isn't "should we let AI write our copy?" The question is "what's our baseline, and how do we measure the difference?" Get that right, and the tooling takes care of itself.

We started this experiment with mild skepticism. We've kept the new page.