Ask a model to design a landing page and it will hand you something that looks finished in about nine seconds. That is the problem. The page looks like a page, so the obvious next step is to ship it, and the thing that would have made it convert never gets done.

A generated page is an average of every page in the training data. Your buyer is not an average. They are at a particular stage, carrying a particular frustration, and comparing you against a particular alternative they already know by name. None of that reaches the model through a one-line prompt, so it fills the space with copy that sounds right and commits to nothing.

The expensive kind of plausible

Plausible copy is worse than bad copy. Bad copy gets caught in review. Plausible copy passes, goes live, spends budget, and produces a flat result that nobody can quite explain. The post-mortem usually blames the offer or the audience, because the page looked fine.

What it lacked was a point of view about one specific buyer, and that was never a design problem.

Give the model the research instead

On Copy Chief's Escape Velocity campaign, the design work stayed with us and the research went to AI. The campaign returned 55.03x average ROAS across ten days, with 130.05x on the strongest ad set and a landing page converting at 10.1%.

The sequence that produced it:

1. Define the segment narrowly. Not copywriters. Freelance copywriters stuck under six figures. A stage, not a job title.

2. Research outward with citations. Perplexity's research feature rather than a model's own opinion, because every finding arrives with a source you can open and check.

3. One pain, one headline. Five researched problems became five headlines, running on both the ads and the page so the click and the landing agree.

4. Build lean. Figma to WordPress, Canva for the Meta creative, Keap for follow-up. The build is quick when the thinking is already done.

5. Spend where the relationship exists. Eighty percent of budget to retargeting the warm email list through Meta Conversions API custom audiences, twenty percent cold.

Step five returned most of the money. Step one is why step five had anything worth saying.

What this means if you are hiring for it

The skill that moved the number was not prompt writing and it was not visual craft, though the page still had to look credible to a professional audience. It was deciding what the page had to prove, to whom, before anything was designed.

That is the part to check for when you bring someone in to run paid creative. Ask what they found out about the buyer and where they found it. If the answer is a description of tools, the work will produce pages that look finished.

The full step-by-step version of this workflow, with the reasoning behind each decision, is written up here: How to use AI in marketing campaigns. The design work itself sits on the Copy Chief project page.

The short version

AI did not make the design good. It made sure we were designing the right thing. Generation is the cheap half of what these tools do. Research is the half that changes the result.