Don't Ask AI to Design Your Landing Page. Ask It to Tear the One You Have Apart.

There are two ways to point AI at a landing page, and the internet overwhelmingly pushes the worse one.

The popular way: prompt a tool to "design a high-converting landing page" and watch it produce something generic — hero, three feature cards, testimonial strip, CTA. It's fast, it's plausible, and it's indistinguishable from the ten thousand other AI-generated pages out there. This blog has covered how to make that approach actually work, including which landing page tools survive being driven by Claude instead of clicked by hand.

The unpopular way, which is what this post is about: paste in the landing page you're already running and ask Claude to take it apart. Not generate a new one. Not "give me five variations." Rip into the thing you have — specifically the three places where money leaks out of a paid campaign: the gap between what your ad promised and what the page says, the above-fold clarity, and the friction in your CTA.

The generate-new approach works best when you have nothing. The tear-apart approach works best when you have traffic and a page, which is the situation most people paying for ads are actually in.

Why critique beats generation

Generation is an exercise in averages. An LLM trained on millions of landing pages produces a landing-page-shaped object — competent, forgettable, wrong in exactly the ways your specific audience doesn't want.

Critique is an exercise in specifics. You're handing the model your actual headline, your actual ad copy, your actual button text, and asking it to act as the skeptical visitor you paid $4 per click to acquire. The failure modes are concrete: it can tell you your headline makes a promise your hero image contradicts, or that your form asks for a phone number at the moment of maximum skepticism.

A critique also has a feedback loop that generation doesn't. When Claude generates a page and conversion rate stays flat, you learn nothing. When Claude critiques your page, then you ship its fixes, then conversion moves — you've established a signal. If you've read how to build a weekly AI iteration loop from campaign data back to Claude, you'll recognize this as the first half of that loop, with a critical detail: the loop only works if the first audit is brutal enough to be worth acting on.

The three-audit prompt

Here's the structure I use. Paste in your page copy, your headline structure, and (this matters most) the exact ad copy that sends traffic to the page. Then run this prompt:

You are a conversion consultant with no loyalty to the person who wrote
this page. I'm paying for traffic to this page. Assume the traffic is
expensive and the visitor is one back-click away from leaving.

Audit three things:

1. AD-PROMISE MATCH. Here is my ad copy: [paste]. Quote every element
   of the ad promise and trace where on the landing page that promise
   is fulfilled or contradicted. Flag anything the ad implies that the
   page delays, buries, or ignores.

2. ABOVE-FOLD CLARITY. A visitor sees only the hero before deciding
   to stay. In one sentence, what do they think this page is offering?
   Write that sentence. Then tell me if it's wrong, vague, or if it
   buries the specific thing my ad promised.

3. CTA FRICTION. List every demand the page makes of the visitor
   before conversion: form fields, account creation, scroll distance,
   claims they must believe without proof. Rank them by cost to the
   visitor. Name the single most expensive one.

Then: give me the three highest-leverage changes, ranked, each with
the specific copy to replace and what to replace it with.

That last line — replace and replace-with — is what turns an audit from commentary into work product. A consultant's opinion costs $150/hr; a rewrite of your hero headline is something you can ship in ten minutes.

Audit 1: The ad-promise gap

This is the highest-value check and the one people skip, because it requires giving the model two artifacts instead of one. Most landing page failures on paid traffic aren't page failures — they're continuity failures. The ad says "free audit, no credit card required" and the page's form asks for a work email plus phone number. The ad shows a specific use case; the hero speaks in category-level abstractions.

Every back-click you paid for is usually a discontinuity, not a design flaw. Claude is good at this because pattern-matching "what does sentence A promise and does sentence B deliver it" is exactly what LLMs do. Paste both, ask it to quote and trace, and you'll typically get back the specific line where trust broke.

A practical rule the audit should enforce: the first headline the visitor sees should be recognizably a continuation of the ad's language, not a restatement of your homepage positioning. If your ad says "cut invoice processing to under a day" and your hero says "the future of accounts payable automation," Claude will catch it — but only if you told it what the ad said.

Audit 2: What the fold actually says

The above-fold test in the prompt is deliberately blunt: write the sentence a visitor would use to describe this page after seeing only the hero. Then compare that sentence to what you think you're selling.

This works because it forces the model to simulate comprehension rather than evaluate craft. A hero can be beautifully written and still produce the wrong sentence. "The intelligent platform for modern revenue teams" is fine prose and a blank sentence — the visitor's description of it would be "some B2B software thing," which is not why they clicked.

Ask for the sentence in the visitor's plain words. If it doesn't contain the specific noun or outcome from your ad, you've found your problem. (If your page was built by an AI tool in the first place, this audit pairs well with the loop that keeps rewriting it weekly — the fold test is the fastest signal check each week.)

Audit 3: The friction ledger

The CTA audit reframes form fields and requirements as costs the visitor pays. That framing matters because teams add friction one reasonable step at a time — a phone number field here, a company-size dropdown there — and never re-add up the total.

When you ask Claude to rank every demand by cost, the usual suspects surface immediately: the field nobody knows how to answer (job title, "company revenue"), the proof gap (a bold claim with no testimonial or number near the button), the button label that describes the software ("Get Started with Flowstate") instead of the payoff ("Show Me My Sample Audit").

You don't need to fix everything. Fix the single most expensive demand. Re-run the page against the same prompt. That's the cycle.

One thing Claude can't see

An honest limitation: text-in, text-out audits miss what the visitor actually sees. Layout, load time, mobile spacing, whether the CTA is below the fold on a phone — none of that lives in your copy. The audit fixes the words; it can't tell you the button renders half off-screen on mobile or that the hero image loads in four seconds.

So the workflow is: run the copy audit with Claude, then look at your own page on a slow phone connection with fresh eyes, and separately check where your leads actually go once the form works. The audit is cheap enough to run every time you ship a change, which is the real point — it's not a one-time consultant report, it's a standing critic you can summon in ninety seconds before every dollar of spend.