Agencies are building websites in a matter of hours. Should you?

My boss has lowkey gone AI crazy. But in the most efficient and genuinely optimal way possible.

He recently rebuilt our entire website using AI vibecoding.

I’m talking: full rebrand, new site, done in a fraction of the time traditional development would take. And, according to Adweek, many major agencies are doing the same thing - building bespoke GEO tools, client platforms, even entire products using AI coding assistants.

‍

Big agencies… like Havas, Broadhead, and Supergood are vibecoding their own tools on top of large language models, often in a matter of hours.

One VP built their agency's first GEO monitoring platform in a single evening. An upgrade that layered in audience personas? Two hours.

This is absolutely wild.

Efficient? Yes. But my first question was: is it good? So, I wanted to break down what actually works about this and where the risks live. Because I’m sure there’s a few.

‍

First off, what actually is vibecoding?

Vibecoding means using AI-powered coding assistants to build software without being a developer yourself. You describe what you need in natural language, the AI generates the code, and you refine through prompts until you get what you want.

Tools like Claude Code, Cursor, Lovable, and Replit are making this possible. You can literally build anything: landing pages, tools, entire platforms by essentially vibing with an AI about what you want until it materialises. No traditional coding required.

‍

The case for: Why it works

Speed is the obvious advantage.

What used to take weeks or months now happens in hours or days. Broadhead's VP built a competitive intelligence feature in two hours that would have required weeks of traditional development and multiple team members.

Control matters too.

Agencies building in-house rather than using off-the-shelf tools can tailor features for specific client needs. So, you're building exactly what you need for how your team works, instead of adapting to someone else’s platform.

Cost efficiency is real.

Havas has avoided signing enterprise agreements that can run into multiple millions annually by building their own tools. Smaller teams can compete with bigger budgets because AI handles the heavy lifting.

Iteration becomes natural.

When changes take two hours instead of two weeks, you can experiment way more. Test features, kill your darlings, evolve based on actual usage rather than being locked into whatever you built six months ago.

Democratisation of development means marketing teams aren't bottlenecked waiting for dev resources. If you can articulate what you need, you can potentially build it yourself.

‍

The case against: Where it breaks down

Well, firstly, quality control becomes a real concern when non-developers are shipping code.

AI-generated code works, but does it work well? Is it secure? Is it maintainable? Will it scale? These questions obviously require developer expertise to answer.

Also, technical debt accumulates fast when you're moving at vibecoding speed.

Quick iterations mean you might be building on shaky foundations. Something that works fine for 100 users might collapse at 10,000. And security vulnerabilities you didn't know to look for might be sitting in your code.

The limitation of AI capabilities matters.

Vibecoding works great for standard use cases and common patterns. When you need something truly custom or complex, AI assistance hits walls. You end up needing actual developers anyway.

Brand consistency and design quality can also suffer when speed is prioritised over craft.

This is not the case for our rebrand. But I can’t help but wonder for others rebuilding company websites: does it feel as considered and cohesive as something a designer and developer collaborated on over weeks? Sometimes yes, sometimes... questionable.

‍

AI vibecoding genuinely works for certain applications.

Building internal tools, prototyping features, creating simple platforms - these are all viable use cases where speed and iteration matter more than perfection.

Agencies are proving this at scale. Havas rolled out their Brand Insights AI globally across 100 countries and 60+ languages. Supergood is using AI to deliver more software than documents. This isn't theoretical - it's happening and working for specific applications.

‍

But it works best with guardrails.

Developer oversight for security and quality. Proper testing before shipping to clients. Understanding that vibecoding is a tool for execution, but strategy and creative direction still require human expertise.

The future Supergood's founder predicts is probably right: everyone's making software now. In two years, agencies will deliver more software than documents. And Vibecoding is enabling that shift.

‍

The question you’ve all been waiting for: should you vibecode your rebrand?

Maybe.

I know that’s not super helpful.

If you need something built quickly, have clear requirements, and can live with iterating as you go, vibecoding can work. If you're building internal tools or prototypes where speed matters more than polish, abso-freaking-lutely.

But if you're creating something customer-facing that represents your brand, represents months of strategy work, needs to scale, or requires deep technical complexity, I would proceed with caution. At minimum, get developer review of what AI generates. The agencies succeeding with this aren't just vibing blindly.

Yeah, it’s powerful. But like any tool, it works best when you know both its capabilities and its limits. Best you brush up x

‍