When Claude Builds a "Marketing Website" Instead of a Landing Page: 5 Things That Give It Away

People search "marketing websites" roughly constantly — probably something near two thousand times a month — and almost none of them are asking what a marketing website is. They're asking why the one they just generated looks wrong. You asked Claude for a marketing website, it gave you something that technically qualifies — header, hero, three feature cards, footer — and yet the moment you imagine a stranger landing on it, something feels off. You can't name it yet.

I can. I've spent the last few weeks auditing real sites that agencies shipped straight out of LLM chats, holding each one against the rubric a marketing site actually gets judged on: message match, conversion paths, CMS-ability, and an SEO baseline. (This is deliberately not another build tutorial. If you want the deployment side, that's covered elsewhere — start with Claude Built It, ChatGPT Coded It: The Three Paths From AI-Generated Code to a Live Domain.)

Here are the five things that give an LLM-built "marketing website" away, every time.

1. The navigation was copied from every demo Claude has ever seen

Ask Claude for a "marketing website" and you will get the same skeleton back, near verbatim: Home, Services, About, Blog, Contact. Maybe a "Pricing." That skeleton isn't marketing architecture — it's the average of ten thousand tutorials, docs pages, and boilerplate templates in the training data.

Real marketing-site navigation is shaped by the buying decision. A local services business might need "Get a Quote" before anything else. A SaaS product might need "Product / Solutions / Pricing" as top-level groups with proof pages nested underneath. The nav is an argument about how visitors decide — and Claude didn't make that argument, because you never told it who was visiting or what they needed to believe.

The fix is unglamorous: before generating anything, write down the two or three things a visitor needs to accomplish, and the one action that counts as success. Then say: "Navigation should exist only to serve those paths. Propose a nav structure and justify each item." Claude is good at this when asked directly. Left alone, it defaults to the demo skeleton.

2. Message match is broken at the page level

On a single-purpose landing page, this is the classic problem — your ad says one thing, the page says another, and Quality Score quietly bleeds. We covered that failure mode in depth in The Ad Is the Brief: How to Feed Your Ad Copy Into Claude So the Landing Page Mirrors It Word for Word.

But on a multi-page marketing website, the problem mutates. Each page gets generated in a fresh chat session, or at least a fresh prompt, and each one optimizes itself in isolation. The result: your homepage says "We help brands grow through data-driven storytelling," your Services page says "Full-stack digital solutions for modern businesses," and your About page says something even vaguer because nothing pinned it down. Four pages, four value propositions, zero coherence.

A marketing website is supposed to sound like one organization. The audit question is blunt: read the H1 of every page in sequence — could a stranger tell these belong to the same company? If not, the fix is a positioning paragraph written once, pasted into every generation prompt, and treated as non-negotiable. Claude will happily abandon your positioning for a smoother-sounding sentence unless you forbid it.

3. There is no conversion path, only CTAs

Here's the subtle one. LLM-generated marketing sites always have buttons — confident, rounded, gradient-filled buttons. What they don't have is a path: a sequence of pages where each one moves a specific kind of visitor one step closer to a specific action.

Run this test on any generated site. Pick your primary conversion — a booked call, a filled quote form. Now walk the site as a skeptical visitor. Where does the homepage send you? Does that page continue the story or restart it? Is there a single page where the visitor actually converts, and do all the other pages route toward it? In most of the sites I audited, every page had a CTA, but the CTAs pointed at different destinations, several of which were "Contact" pages with a form that asked eleven questions. Nothing was wrong with any individual page. The site as a system went nowhere.

The fix lives in the prompt, not the design: tell Claude the name of the money page, and require that every internal CTA either lands there or on a page that explicitly hands off to it. One destination. Everything feeds it.

4. Content that can't be edited — the CMS illusion

This is the giveaway that bites hardest a month later. The generated site looks like a marketing website — and marketing websites, everyone knows, have content that changes: new case studies, fresh testimonials, seasonal offers, a blog. So the client asks their agency rep: "Where do I add a testimonial?" And the answer is: open the code Claude wrote and edit a JSX array by hand.

In my audits, this showed up as testimonials hardcoded into a component file, service descriptions duplicated across pages with tiny differences, and blog sections linking to nothing because Claude generated the card layout without any actual posts behind it. The site presents as CMS-driven and operates as a static artifact.

Two honest options here. Either accept the static reality and price updates accordingly — which is fine for a small brochure site with a once-a-quarter change cadence — or build on a stack where the content layer is real. The first post to read is You Built It With ChatGPT. Here's How to Point Your Domain at It., because the deployment choice (a plain static host versus something with a content layer) often determines whether editing is possible at all. Just know which situation you're in before the client asks, not after.

5. Zero SEO baseline — the site is invisible by construction

LLM-generated sites frequently ship with missing or duplicated title tags, meta descriptions that are either absent or identical across every page, no meaningful heading hierarchy, no schema markup, and image alt text that is literally the filename. None of this affects how the page looks, which is why nobody catches it in review — but it means the "marketing website" cannot do the single most important thing a marketing website does: get found.

Worse, some of these failures compound. Duplicate titles mean search engines can't distinguish your Service A page from your Service B page, so neither ranks. Missing alt text hurts both accessibility and image search. And because generated sites are often client-side-rendered SPAs by default, some crawlers see an empty shell where your content should be.

Run the baseline yourself before launch: every page has a unique title and description, one H1 per page, headings in logical order, alt text on every image, and a sitemap. If that sounds like a pile of boring checks, it is — and Before You Deploy: The 12-Point Checklist for Code ChatGPT or Claude Handed You exists precisely because boring checks are the ones AI code fails. For the deeper crawl-visible issues, Your AI Landing Page Is Raising Your CPC: How Generated Code Quietly Wrecks Quality Score covers the paid-search cousin of the same problem.

The pattern underneath all five

Every one of these failures comes from the same root: Claude built the appearance of a marketing website because that's what marketing websites look like, not because it understood what one does. Appearance is easy to generate. Function — message match, conversion paths, editable content, findability — has to be specified, checked, and enforced by you.

So stop asking for "a marketing website." Hand over the positioning, name the conversion, define the content-update rhythm, and demand the SEO baseline in the same breath as the design. The difference between a demo and a marketing site isn't what Claude generates. It's whether anyone audited it against the job it was supposed to do.