We Don't Want a Beige Internet: The Homogeneity Problem with AI-Built Sites

There are two kinds of AI-generated landing pages. 

The first kind gets built with no design brief at all. You can spot it immediately by the purple-to-indigo gradient in the hero, the headline in an Inter font, the content cards with softly rounded corners, and the CTA button that is some variation of bg-indigo-500, which is the default color Tailwind CSS shipped with, now appearing on what feels like half the internet.

The second kind gets built with a brand color document dropped into the prompt. It's subtler but just as recognizable. It has the right hex values applied to the wrong things, distributed across the page wherever the AI tool decided emphasis was needed. It has no underlying logic about hierarchy or weight or what each color is actually supposed to communicate. 

Both pages look like they were built by a tool rather than a person. The difference is just how obvious it is.

This is the beigeification of the web. And it's accelerating.

What "Vibe Coding" Actually Produces

The term "vibe coding" was coined by AI researcher Andrej Karpathy in early 2025. The idea is simple: You describe what you want in natural language, AI generates the code, and you accept it without reviewing every line. By the end of 2025, Collins Dictionary named it Word of the Year. By early 2026, the numbers had caught up with the hype. In fact, 92% of US developers use AI coding tools daily, and GitHub reports that 46% of all new code is now AI-generated.

The tools doing most of this work—Lovable, v0 by Vercel, Replit, Claude Artifacts, ChatGPT Canvas—have become remarkably capable at quickly shipping something functional, and if you're a founder who needs a landing page by Thursday or a marketer who needs a quick prototype to show a client what you mean, that speed is a real benefit.

The problem is what the AI tools default to when you don't tell them otherwise.

Why They All Look the Same (And Yes, Tailwind Is Part of It)

If you've ever wondered why AI-generated sites share such a specific visual fingerprint, the short answer is training data bias with Tailwind CSS at the center of it.

Tailwind is a utility-first CSS framework that has become dominant in modern web development. Instead of writing custom CSS, developers apply pre-defined class names directly in their HTML—things like p-4, text-xl, or bg-indigo-500. It's fast, readable, and predictable, which is exactly why AI models love it. Tailwind and the shadcn/ui component library are overrepresented in the code across millions of public GitHub repositories, so AI models learned to treat that aesthetic as the statistically most likely correct answer to any design prompt.

The indigo problem specifically traces back to a decision made years ago by Adam Wathan, Tailwind's creator. When building Tailwind UI, the official component library, he used bg-indigo-500 as the default button color because it looked professional and not too plain. In August 2025, he posted an apology on X that got over a million views: "I'd like to formally apologize for making every button in Tailwind UI bg-indigo-500 five years ago, leading to every AI generated UI on earth also being indigo." Abandoned projects, half-finished repositories, tutorials—all of them using bg-indigo-500 as their placeholder—became training data, and now the AI generates it constantly across millions of sites. 

As one analysis put it, the AI isn't designing; it's averaging. The output is the median of every Tailwind-based landing page that existed in its training data, and because AI-generated sites are now being published at scale and folded back into the next round of training data, the distribution keeps shifting toward indigo.

Beyond color, the homogeneity runs deeper. The default structure of an AI-generated page is almost always the same: a centered hero with a headline and subheadline, three to four feature cards below it, a testimonial section, and a footer. It's not a bad structure and it converts reasonably well. But it's also the layout for what feels like every SaaS landing page, every fintech app, and every consulting firm that launched in the last eighteen months.

Research Backs This Up

A March 2026 paper from researchers at the University of Washington, "Interrogating Design Homogenization in Web Vibe Coding", examined this problem directly. The researchers looked at six major vibe coding platforms and found that because these tools are primarily trained on English-centric datasets aligned with Western aesthetic conventions, they consistently reproduce those dominant style patterns. The paper identifies what they call a "good enough" trap: When the AI generates something that looks polished and functional, creators without coding backgrounds accept it. There's no friction in the process that would prompt them to ask whether the design actually reflects their brand or their audience.

The researchers argue this is particularly risky for lay creators—people who don't know enough CSS or design to recognize when the output is generic, and who may not have a designer to flag it. They describe a feedback loop of homogenization: Generic outputs become training data, which produces more generic outputs, which narrow the visual language of the web over time.

Edelman's 2025 Trust Barometer found that trust is now as much of a purchase consideration as quality or price, and that inaction invites commoditization. A site that looks like every other site signals (however unintentionally) that there's nothing distinctive behind it.

What It Feels Like to Be on the Receiving End

If you're a consumer scrolling through SaaS options, the beigeification problem is a friction you may feel without being able to name it. Every tool looks credible. Every landing page uses the same hierarchy of trust signals, and every hero section makes a claim about saving time or reducing complexity.

Design has always done brand work. When a site's typography, color palette, and layout feel specific, when they feel like they came from a decision someone made about who the company is, it registers (even subconsciously). When every site uses the same Inter/rounded-card/indigo-button template, that signal disappears, and consumers are already developing pattern recognition for "this was made by AI," with associations that aren't flattering (think: interchangeable, low-effort, not worth a second look).

For B2B companies, this is especially important. If you're selling enterprise software, or managing a platform whose buyers run compliance checklists before they sign anything, your site is often the first thing a prospect sees after a cold email or a LinkedIn click. A generic layout doesn't communicate that you understand a specific, complex problem. It communicates that you built something fast and hoped no one noticed.

Three Ways to Use AI and Still Have a Site That Looks Like Yours

None of this means you should stop using AI to build things. It means you should stop expecting AI to make design decisions for you. 

Here’s what we’ve learned:

1. Structure is where AI earns its keep.

Vibe coding tools are excellent at scaffolding. They can spin up a functional component library, a working page template, or a data-driven layout in a fraction of the time it would take to do it from scratch. 

Use the AI output as a working prototype. Build something to react to, move things around in, and annotate. But not as the finished product. If you're using Claude Artifacts or v0 to build a landing page, generate the layout and then export it or screenshot it. Use that as the brief for a designer, and not as the deliverable.

2. You have to figure out what your brand looks like before you open a prompt.

One of the most effective things you can do before vibe coding anything is spend 30 minutes in a design tool first—Figma for the full brand mapping exercise, Whimsical if you're moving fast and just need to sketch structure, Google Stitch if you want AI-assisted prototyping that at least starts from your visual reference points. Map out what your brand actually looks like. Your hex values. Your typefaces. The visual weight of your logo. Screenshots of sites or apps that feel right to you, and (just as importantly) ones that don't.

When you bring that reference material into a design tool first, you're giving yourself a filter. You can look at the AI's output against your actual brand language and see immediately where it diverges. You're also creating something that a designer or developer can take and use to override the AI's defaults with specificity.

3. There’s an important handoff most people skip.

What usually happens: someone generates a page with AI, decides it looks good enough, and publishes it. They skip the entire layer of human judgment that usually sits between "the AI built something" and "this communicates who we are."

That layer isn't one person or one pass. Usually, several important people are involved. A content strategist strips out the feature inventory and rewrites it with the buyer experience in mind. A UX strategist reworks the structure so the page makes an argument instead of just listing things. And thanks to the previous steps, a designer has content worth making visually distinctive, because the thinking has already gone into it and they're not being asked to compensate for the absence of strategy.

The AI output is raw material that feeds into that process, not a shortcut around it. A page that skips straight from prompt to publish hasn't been positioned. It's been generated. And before it goes anywhere near a real URL, someone should run a Lighthouse audit, check every link, and test it at something other than 100% zoom on a MacBook. Because the bar for "looked fine when we approved it" is lower than it sounds.

What You Shouldn't Use AI For (At Least Not Without Review)

The decisions that carry the most brand weight, such as your color palette, your typography, your hero section, your navigation, your product page are exactly the ones AI is worst positioned to make, because any competitor using the same tool with a similar prompt could generate the same output. That's the visible problem. 

The less visible problem is that even when the page looks fine, the code underneath it may not be. A 2025 review of Lovable-generated applications found that more than 10% shipped with security vulnerabilities that could expose user data. AI will produce something that looks functional in isolation but indistinguishable on the outside and broken on the inside is a bad combination for any company whose buyers are paying attention to both.

When the Stakes Are Higher Than a Landing Page

The perspective we’ve shared so far applies to every company using AI to build public-facing work, but there's a specific category of company for whom shipping AI slop carries real credibility and risk consequences: established companies selling into industries where buyers are careful, skeptical, and doing real due diligence before they sign anything. 

We worked with a software company that had published an external microsite built in Claude Artifacts before we came on. Their buyers are sophisticated and evaluation driven. The microsite was the first thing a prospect landed on after a cold outreach or a conference conversation.

Here's what was on it: 

1. A stat block in the hero with four numbers, four labels that used a random stat like proof of something but wasn't anchored to any problem their buyers actually experienced

2. A three-column card layout (that is, without exaggeration, the default output of Claude Artifacts when you ask it to explain three options) 

3. Brand colors applied to headline text, button labels, and section markers with no apparent hierarchy logic (no H1, H2, H3, etc.) 

4. Copy that listed features in sequence rather than leading with what a buyer's experience looks like before and after using the product

5. Multiple broken links and no meaningful SEO structure

The design wasn't offensive. That's part of what makes it dangerous. It looked functional enough that someone approved it. But anyone in that buyer's seat who had evaluated more than a handful of vendors in the last two years certainly developed an eye for this—the three cards, the stat block, the color application that doesn't quite cohere. And what they saw when they landed on that page  was that senior-level oversight had not been in the room when it shipped (or worse, it was and didn't care).

For a company whose product touches sensitive data or operates in a context where buyers scrutinize vendors carefully, that signal carries weight beyond aesthetics. Broken links on an external microsite tell a careful buyer that QA doesn't happen here. Feature-heavy copy that reads like a prompt response suggests the company doesn't actually understand the buyer's problem well enough to describe solving it. And WCAG accessibility compliance isn't optional for companies selling into enterprise.

The Internet Doesn't Have to Be Beige

The tools exist, the capability is real, and the speed is genuinely useful. But speed without intention produces a website that looks like it was generated by averaging everything that came before it (which, technically, it was).

The designers who are pushing back on this—and there's a growing movement of them, reaching toward brutalism, pixel art, custom type, tactile textures, anything that resists the default—are doing so because they know something that vibe coding culture tends to skip over: Visual distinctiveness is doing communication work before anyone reads a word. It tells someone what kind of company you are before they've had a chance to decide whether to keep reading.

At Wheels Up Collective, we work with companies who want to use AI without disappearing into it. We use the tools (Claude, Codex, v0, all of it) but we bring the strategy, the brand thinking, and the design sensibility that turns a generated prototype into something that actually looks like you.

AI can sketch. We turn it into art.

We like rainbows, not beige. My two-year-old daughter does too, and I'd really prefer she grow up able to find them beyond just the physical world.