We let AI write the words, not build the page

Opening thesis

When you test a value proposition, you want the market reacting to your promise — not to a broken layout. So we do not ask AI to invent the whole page. It writes the words; proven structure holds the shape. Here is why that division of labor is what makes a test mean something.

A generated page can quietly poison the test

It is tempting to ask AI to spin up the entire landing page, ad, or survey from scratch. The output looks plausible. But generated structure breaks in ways you do not notice — hierarchy, spacing, accessibility, the things that read as “this is a real company.” You think you are testing your value proposition; you are actually testing a slightly broken artifact.

Early tests are fragile enough already: small audience, noisy signal, decisions made on thin response. Add an avoidable layout defect and you cannot tell whether the market rejected your idea or just could not take the page seriously. A bad container makes a good offer look bad and a bad offer look like noise.

Assembly, not generation

So we split the job along the grain of what AI is actually good at. It writes the parts that are language — headlines, the way a pain is framed, the call to action, the tone a given channel expects. Tested structure owns the parts that need to be consistent: layout, the platform’s rules, the shape of the thing.

This is not AI-skepticism; it is a better division of labor. Because the container stays stable, the only thing changing between tests is the promise — so when results differ, you can actually read why. Freehand every asset and you have changed the structure and the hypothesis at once, and learned nothing clean.

Honest limits: a fixed set of building blocks means some experiments — the ones that need video, interactive demos, or elaborate comparisons — are not in the kit yet, and some channels still need a human in the loop until the connections are in place. We would rather ship a clean, honest test than a flashy, uncontrolled one.

Protect the test from noise

When you run a test, protect it from avoidable noise. You want the market responding to the problem, the promise, and the ask — not to hidden design defects. Use AI to sharpen the message and generate alternatives; use tested patterns for the shape. Change too many things at once and you lose the ability to learn what mattered.

The payoff is that even a failed test teaches you something. When the container was controlled, a flat result lets you ask sharp questions — wrong audience, weak promise, too big an ask, wrong channel — instead of wondering whether the page just rendered badly.