AI Website Builders Tested: Lovable, Claude and Codex

A client asked me recently whether they could just use Lovable to build their next website instead of hiring our agency. It's a fair question. The demos are impressive, the pricing is a fraction of what any agency charges, and the marketing makes it sound like the entire web design industry is about to vanish.

I've been building with AI tools for the last two years, both in our agency workflow and as direct experiments. I've used Lovable, Claude Code, Codex, Framer's AI features, Relume, and several others on real projects. My co-founder Emma has been testing the Relume + Webflow workflow specifically with client work in parallel. So I went back and tested them again specifically against the brief a business owner might give them, and pulled in her observations where they're sharper than mine.

The honest answer turned out to be more nuanced than "AI good" or "AI bad". This is what I actually found.

What I Tested and How

I ran the same brief through four categories of AI tool, each representing a different approach to AI-assisted web development.

Lovable is the buzziest of the AI-native website builders. You describe what you want in natural language and it generates a full React application, complete with components, routing, and a working deployment. No code required.

Claude Code and Codex are AI coding assistants. They don't pretend to build a website for you out of thin air. They work alongside a developer, generating code, suggesting architecture, and helping debug. You direct them.

Framer with AI features is a design-first tool with AI generation bolted on, primarily aimed at designers who want to ship quickly without a developer.

Relume is a different beast altogether. It generates AI sitemaps and wireframes that export directly to Webflow, where a designer then refines and ships the actual build. It's not trying to remove the human - it's trying to compress the structure-and-scaffold phase of a designer's workflow.

I tried building a website for a fictional boutique Lake District hotel with each tool. A small design-led property where the website needs to feel distinct, communicate brand experience, and convert browsers into direct bookings rather than feed third-party platforms. Same brief, same content, same business goals. What I learned wasn't about which tool is "best". It was about the gap between what AI tools promise and what they actually require from you.

Lovable Promises Magic, Then Hits a Brick Wall

The first ten minutes of Lovable are genuinely impressive. I described the hotel, and within a few prompts I had a working multi-page site on screen. Hero section, rooms overview, contact form, footer. It looked like a website.

Where Lovable genuinely works

For a basic brochure site (one page, a contact form, maybe an "about us" section), this is genuinely useful. If you're a solo founder testing an idea or you just need a placeholder so people can verify your business exists online, Lovable will get you there in an hour. No complaints.

Where it breaks at scale

The brick wall came as soon as I tried to do anything beyond the demo template. Embedding a booking availability widget broke the layout. Adding a properly designed rooms gallery with bespoke imagery for each room required edits that the AI couldn't quite reason about. When I asked it to optimise the loading speed of the page, it made changes that broke unrelated components. The more I asked it to do, the more it started fighting itself.

The deeper problem was architecture. Lovable's output is fine for what it is, but it's not built around any meaningful project structure. There's no clean separation between presentation and logic, no consideration for how the code might be extended six months later, no proper component reuse strategy. For a one-page brochure this doesn't matter. For anything that needs to scale or be maintained, it becomes a liability fast.

This isn't just my experience. Google Chrome team's Addy Osmani coined what's become known as the "70% problem": AI tools rapidly produce 70% of an app, then the remaining 30% (edge cases, security, production integration) is as hard as ever. Lovable's failures sit squarely in that last 30%.

The security reality

Security is a real concern too. In early 2026, a Lovable-built app vulnerability (CVE-2025-48757) exposed over 18,000 users across 170+ apps due to missing Row Level Security on the underlying Supabase database. A Veracode study published the same year found that 45% of AI-generated code introduces security vulnerabilities, with some languages hitting 70%+ failure rates. These aren't theoretical concerns. They're production incidents involving real businesses.

The verdict on Lovable: genuinely useful for tiny, simple, short-lifespan sites and as an early-stage prototyping accelerator. Real production apps do exist on Lovable (some with meaningful revenue), but every credible success story I've found involves either a technical founder iterating heavily, early export to GitHub with proper engineering taking over for security and scale, or both. The unmanaged "use Lovable to launch your business without code" path is where things go wrong. If you're treating it as a long-term business asset without engineering support, you'll regret it.

Claude Code and Codex Are Powerful AND Demanding

Claude Code and Codex are a completely different proposition. These are AI tools designed to assist developers, not replace them. You're not describing a website and watching it appear. You're directing an AI through specific implementation steps, reviewing its output, refining its architecture, and catching its mistakes.

When they accelerate experienced developers

In the hands of someone who already understands web development, both tools are genuinely useful. For specific cases (generating boilerplate, debugging known patterns, refactoring familiar code) they accelerate me considerably. A METR randomised controlled trial of experienced developers found participants were 19% slower with AI tools while believing they were 20% faster. That study used early-2025 AI tools, and METR's February 2026 methodology update acknowledged significant selection effects in their data — newer developers using later AI tools showed results closer to neutral, and they note that developers who now refuse to work without AI are dropping out of the study, making estimates hard to interpret. Our experience is that the tools have improved meaningfully. The honest answer is that AI coding tools are situationally powerful, not universally faster.

The productivity that does materialise is entirely contingent on you knowing what you're doing.

What happens when non-experts use them

If you ask Claude Code or Codex to "build me a business website" without specifying the architecture, you'll get something. It might even look reasonable in the browser. But underneath, you'll likely have inconsistent component patterns, mixed approaches to state management, security gaps in form handling, accessibility issues that fail compliance audits, and dependencies that conflict in ways you won't notice until something breaks in production.

I tested this directly by asking Codex to build the hotel site cold, without proper specification. The result compiled and ran. It was also unmaintainable and had three separate ways of handling form submission across different pages. Real-world consequence: any future change would require rewriting large chunks of the codebase rather than editing one component.

When I gave the same tool a properly specified brief with clear architectural rules, defined component patterns, and a real understanding of dependencies, the output was excellent. Production-ready. Genuinely fast to build.

The skill that matters is not prompting. It's architecture, requirements analysis, dependency understanding, and the judgment to know what good code looks like when the AI hands it to you.

Relume and Webflow: A Different Beast Entirely

Lovable and Claude Code/Codex sit at opposite ends of the "AI does everything" spectrum. Relume occupies a third position that deserves its own treatment, because the trade-offs are completely different.

Relume is an AI sitemap and wireframe generator that exports directly to Webflow. It's not pretending to write your app for you. It's giving a designer a fast scaffold and then handing them back the design and development work in Webflow's native editor.

Emma has been using this workflow on real client projects for months, and her honest assessment is consistent with what most working designers report.

What Relume genuinely does well. The sitemap and wireframe generation is impressive. You describe a site, and Relume produces a clear page-by-page structure with sectioned wireframes you can edit as needed and present to a client in real time. That alone compresses the discovery and proposal phase of a project significantly. Designer Anmol S, after 275 days of using it, calls it "a powerful, professional ally for any serious Webflow designer who values speed, structure, and scalability."

Where the time you saved comes back. The Webflow export is the start of the work, not the end. From Emma's experience and consistent with wider reports, you still need to:

• Refine the visual design properly (colour systems, typography, brand alignment) - you can work on design directly in Relume, but getting to a result you're genuinely happy with takes real time spent tweaking fonts and colours. Relume's components are intentionally neutral, which gives you a starting point, but significant customisation is needed to make the site look like your client's brand rather than a Relume site
• Build any non-default animations manually in Webflow - Relume's components are static, and Webflow's native interactions are limited enough that serious agencies often reach for tools like GSAP
• Configure Webflow correctly for SEO - no native schema markup support, weak pagination URLs, no server-side access for things like redirects or custom headers, all of which need workarounds
• Add anything beyond a marketing site via third-party tools - Webflow's native ecommerce doesn't support subscriptions or recurring payments, so booking systems and memberships get bolted on through Memberstack, Outseta, FlowBookings or Stripe Elements

Emma's summary: "It saves you days on planning and structure. You give them back in refinement and customisation, especially if you want the site to look distinctly yours rather than what other designers call 'Relume-ish'."

That word - "Relume-ish" - is the most common criticism in the wider designer community. A widely-cited Product Hunt review notes: "You can tell if a Webflow dev uses the Relume library and has barely customized it." Even Relume's own advocates concede the point - the same Medium piece from Anmol S admits "the real skill lies in taking these well-built components and customizing them to create a unique, brand-aligned website that doesn't look like a Relume site."

Where Relume + Webflow genuinely fits. Designers who already own the Webflow workflow and want to compress the structural phase of every project. Clients who need a polished, content-led site without bespoke functionality.

Where it doesn't. Anything that needs custom interactions, complex integrations (real ecommerce with subscriptions, booking systems, multi-language with proper routing), or visual differentiation strong enough that "looks like a Relume site" would be a problem for the brand.

The Skill Gap Nobody's Honest About

The AI website tools marketing pitch is that you no longer need a developer. The reality is that the skills required have shifted, not disappeared.

To build a real business website with AI in 2026, you still need to know:

• What the architecture should look like before you write the first line. The AI won't decide this for you, and if it does, it'll usually pick something brittle.
• Which dependencies work with which, and how they'll conflict. AI tools confidently install packages that don't play nicely together.
• How to debug when something breaks. AI can help, but you have to read the code well enough to know whether the AI's suggested fix is right or just plausible-looking.
• What good code looks like for your specific tech stack. AI defaults to popular patterns, which may not be appropriate for your case.
• How to think about scalability, security, accessibility, performance, SEO, GEO and AEO at the architectural level. AI won't bring these up unless you ask — unlike a professional agency, the tool won't proactively look out for your visibility in traditional or AI-driven search. You need to direct it, and you need to know enough to ask the right questions.
• What the user actually needs from the site, in business terms. This is the work the AI fundamentally cannot do, because it doesn't understand your business.
• What a well-designed site actually looks like. AI-generated design has improved, but it's still often easy to spot — and it doesn't yet compare with what a professional designer produces. The visual gap is real.

A professional developer or designer brings all of these to a project as a baseline assumption. A business owner using AI for the first time brings none of them. That gap is the whole game.

The tools are powerful, but they amplify what you already are. If you're a careful developer, AI makes you faster. If you're a beginner who treats it as magic, AI builds you something that looks like a website and behaves like a technical debt landmine.

When AI Tools Genuinely Are the Right Choice

Being honest about where AI tools work matters as much as being honest about where they don't.

Placeholder sites and brochures

If you need a quick low-effort site as a placeholder or a brochure (a single page with your services and a contact form, nothing more), Lovable and similar AI builders will do the job. The investment is low, the lifespan is short, and nobody's depending on this site to convert visitors into customers.

Experienced developers and designers

If you're a developer or designer with proper professional grounding, Claude Code and Codex are essential parts of the modern workflow. They make experienced builders faster and more capable. Most of our Digital Otter work in 2026 involves these tools at some stage. They're tools, not replacements.

Idea validation and prototyping

If you're testing an idea before committing to a real build, AI tools let you prototype faster than ever. Build something rough, see if the concept holds, then invest properly once it does.

The common thread is that AI tools work when used by people who already understand the underlying craft, or when the outcome doesn't really need to be polished.

The Trap Most Business Owners Fall Into

The trap is buying the demo and shipping the demo.

A Lovable demo looks brilliant in the marketing video because the marketing video uses the exact use case Lovable handles best. The moment you push the tool toward your real business needs, the gap appears. By that point, you've usually invested enough time that it's hard to walk away. Sunk cost takes over, and the result is a website that almost works but never quite does the job.

The honest cost of "free AI website" tools tends to look something like this. You spend an afternoon getting the basics up. You spend the next two weeks trying to make it look right, integrate with your CRM, fix the mobile experience, and get the contact form actually sending emails. You upgrade the plan to remove branding and add custom domains, which runs you £30-£50 per month. You spend evenings fixing things you can't quite identify.

After two months, you have a site that mostly works, took you 40+ hours of your own time, and still has the conversion design and SEO weaknesses I covered in our piece on why custom web design matters. The "free" route turns out to have cost you a substantial portion of your working time plus monthly platform fees.

A properly built site costs more in cash but less in everything else. The result actually does the job. We see this pattern repeat across websites in Cumbria, Kendal and the wider UK: business owners reach the conclusion that the cheap route cost them more than the considered build ever would have. Research from the Baymard Institute consistently shows that conversion design is one of the highest-leverage factors in website performance, and AI builders consistently get this wrong. Google's own research on Core Web Vitals shows the same pattern for performance.

If you want to understand where AI does add genuine value to web work in 2026, we cover that in our piece on how UK businesses are using AI on their websites in 2026.

The Honest Answer

The client who asked whether they should use Lovable instead of hiring us got a straight answer.

When AI was the wrong call

For their specific case, Lovable was the wrong tool. The architectural limitations meant the site would underperform on conversion and search visibility from day one, and would become harder to fix the longer it ran. The hidden time cost would eat any apparent savings.

When it would have worked

For a hobby project or a placeholder, the answer would have been different. AI tools genuinely do those jobs.

The category error

The category error is treating "AI website builder" as a single thing. Lovable tries to remove the human entirely (works for trivial cases). Claude Code and Codex amplify a skilled developer (work brilliantly in the right hands). Relume compresses a designer's structural workflow into Webflow (works at agency volume, not for solo founders). Framer's AI sits across categories with design-first generation. They're not interchangeable, and they don't solve the same problem for the same buyer.

The skill of building a working business website hasn't gone away. It's just shifted from "writing code" toward "understanding architecture, requirements, and what good output looks like". That shift makes experienced developers more productive. It does not make non-experts into developers.

Anyone telling you AI lets you skip the underlying expertise is selling you something. The expertise hasn't disappeared. It's been hidden inside the tool, and it shows up in the gap between the demo and the result.

And if you've decided you've outgrown the DIY approach and want to talk through what a proper build looks like for your business, get in touch. We build websites in Kendal, Cumbria, and across the UK. We use AI tools every day in our own workflow, and we're happy to tell you when they're the right answer and when they're not.