How a marketing team runs an AI-built website without a developer
Benjamin Libor7 min read
A marketing team runs an AI-built website by splitting the work in two: design and structure change through an AI tool that writes code from a plain-language request, and content changes through a visual CMS, with every change shown in a preview and approved before it goes live. Nobody hand-designs pages in a visual editor and nobody waits on a developer for a new section. What makes it safe is not the AI, it's the rules around it: a design system the AI must use, one named approver, and a full history of every change.
In short
- The old reason to stay on a visual builder, needing a developer for every change on a code site, no longer holds.
- Design and structure change by prompt to an AI tool like Claude Code; content changes by clicks in a visual CMS.
- A design system and a rules file in the repository keep the AI on brand; a preview link and one approver keep mistakes off the live site.
- Everything is versioned in the company's repository, so any change can be undone and nothing is locked into one vendor.
- A developer or studio is still useful for the set-up and for larger features, just not for day-to-day work.
Why did code sites need developers, and what changed?
For years the trade was simple. A visual builder like Framer let a marketer change a page without help, at the cost of control and lock-in. A code site gave full control, at the cost of a developer for every new section, every layout fix and every form.
Two things changed. AI coding tools such as Claude Code and Cursor now write good front-end code from a sentence, working inside an existing project with its components and rules. And visual CMS layers for code sites let non-technical people edit content stored as files in the repository, the way they edited a CMS before. Put together, the marketing team gets what it liked about the visual builder without the platform around it.
If you're weighing that trade in general, Framer vs custom code for AI search compares the two stacks in detail.
Who does what on an AI-built website?
The work splits cleanly by kind of change.
| Change | Who | How | Approval |
|---|---|---|---|
| Fix a typo, update a headline | Anyone with content rights | Visual CMS or edit on the page | Publish |
| New blog post or case study | Content marketer | CMS item, or ask the AI to draft it | Preview, then publish |
| New landing page from existing sections | Marketer | Prompt to the AI tool | Preview link, approver |
| New section or component | Marketer or designer | Prompt to the AI tool, with a sketch or reference | Preview link, approver |
| Redirects, page titles and descriptions | SEO owner | CMS settings | Publish |
| New feature (calculator, gated content) | Marketer with a developer or studio | AI tool, reviewed by someone technical | Review and approver |
| Hosting, domain, repository access | Admin | Account settings | Two admins |
Notice the shape: most changes never need anyone technical, and the ones that do are the larger, rarer ones.
What does a design change look like in practice?
Take a made-up company, Fernwick, a scheduling tool. Its marketing lead wants a new pricing comparison section on the pricing page.
- She opens the AI tool connected to the website's repository and writes: add a comparison table under the plans with these rows, in our usual section style.
- The AI tool builds the section from the existing components and design tokens, and the host creates a private preview link.
- She checks it on her phone and asks for one change: make the first column narrower.
- The preview updates. The approver, the head of marketing, opens the link and approves.
- The change goes live, and the commit records who asked, who approved and what changed.
No ticket, no sprint, no designer laying out boxes by hand. The design work happened once, when the design system was defined; the AI applies it.
What does a content change look like?
Content is where a visual CMS earns its place, because most content changes are small and frequent.
- A marketer opens the CMS, picks the collection (blog, case studies, jobs), edits the item's fields and saves a draft.
- Or, signed in, she opens the live page and edits the text in place.
- Someone with the publish right reviews the draft and clicks Publish, which writes the change to the repository as one commit.
- The host builds and deploys it, and the sitemap and listings update themselves.
The content lives as files in the company's repository, not in the CMS vendor's database. That matters for two reasons: AI tools can read and write it directly (a prompt can draft an article into the right collection, or translate a page), and nothing is lost if you change tools. How to write content that AI answers then quote is in writing content that AI answers quote.
What keeps an AI-built site on brand?
Left alone, an AI tool will invent a new shade of blue for every section. Three things stop that.
A design system the AI must use
Colours, type, spacing and components are defined once, as tokens, in the repository. The AI is told to use only those, and new sections are assembled from existing components. Ask for something new and it extends the system rather than going around it.
A rules file in the repository
AI coding tools read a file of instructions at the root of the project (often called AGENTS.md or CLAUDE.md). It says where copy lives, which components exist, what never to change, how pages must be structured for search, and which checks must pass. It's the brief that every request starts from, so every person on the team gets the same behaviour from the AI.
Automatic checks
Before a preview is offered, the project runs checks: broken links, missing alt text, missing titles, accessibility problems, pages that got slower. A change that fails doesn't go live. This is cheap to set up once and catches the mistakes people miss in review.
How do approvals and rollback work?
Approvals are what make a non-technical team comfortable with a code site.
- One named approver in marketing for design changes. Not a committee; one person who checks the preview and says yes.
- Rights per person for content: who may edit, who may change SEO settings, who may publish. Most people edit; few publish.
- Drafts by default. Nothing reaches the live site until someone publishes it.
- Undo is a new change. Every publish is a commit in the repository with its author. Undoing it creates another commit that puts the old version back, so the history stays complete.
Compared with a visual builder, the history is better, not worse: you can see every change to every page, by whom, and when.
When do you still need a developer?
Less often than before, but not never.
- The set-up. Repository, hosting, the CMS, the design system, the rules file and the checks. This is a one-off piece of work for a developer or a studio.
- Larger features. Anything with logic, data or security: sign-ins, gated content, calculators, integrations with your product. The AI writes most of it, and someone technical should review it.
- When something breaks in a way the AI can't explain. Rare, but worth having someone you can call.
A good arrangement is a studio or freelancer on a small retainer, with the company owning every account. The team runs the site; the expert is there for the big pieces.
Questions people ask
Can non-technical people really edit a code website?
Yes, with a visual CMS on top. Content changes are form fields and in-page editing, much like Framer or WordPress, and layout changes are plain-language requests to an AI tool with a preview link. Nobody on the marketing team edits code by hand.
Is it safe to let AI change the website?
It is with the right guardrails: a design system the AI must use, a rules file, automatic checks, a preview for every change and one approver before anything goes live. Every change is versioned and can be undone.
Which AI tool should we use to build and change the site?
Claude Code and Cursor are the common choices for working inside a website's repository. What matters more than the tool is the rules file and design system in the repository, which make any of them behave consistently.
What happens if we stop paying for the CMS?
If the content lives in your own repository, the site keeps running and the content stays yours. You lose the visual editing screens, not the website.
Written by Benjamin Libor, founder, echo.
Something wrong or out of date? Write to hello@echo-aeo.com and we'll fix it.