Anyone can spin up a website with an AI coding tool in an afternoon now. That is not the problem. The problem is that thousands of people are spinning up the same website, because they are all handing the same decisions to the same handful of models. A recent conversation between Y Combinator general partner Aaron Epstein and Raphael Schaad, the founder of Cron (acquired by Notion), walked through a batch of “vibe coded” sites and found the same seven failure patterns showing up again and again.

None of these AI website mistakes are fatal, and none require design school to fix. They are mostly a matter of knowing what to watch for before you ship.

Letting the model pick your look and feel is why every site looks the same

Ask an AI tool for something “modern” and it will hand you the statistical average of every modern-looking site in its training data. That is fine in isolation and forgettable in aggregate. Epstein and Schaad pointed to a specific pattern that keeps surfacing: a fake software dashboard with red, green, blue and purple callouts, icons rendered in a slightly darker shade of the background, and a color scheme Schaad half-jokingly called “the Fisher-Price primary.” It looks polished. It also looks like every other AI-built dashboard mockup on the internet.

Purple gradients are the other tell. A style choice that once took months to spread from one standout site to its imitators now propagates in a week, because the model gets trained on whatever gets linked to and copied often enough. The feedback loop that used to slow trends down no longer exists, so originality decays fast unless someone actively pushes back on the defaults.

Watch for these specific patterns creeping into your build:

• Purple gradients used as a default rather than a brand choice
• Bento-box grid layouts applied without a reason
• Generic dashboard mockups with rainbow callout colors
• Icon sets and emoji-style elements that could belong to any product

None of these are inherently wrong. They became wrong the moment every AI-assisted site started using them at once, which is exactly the dynamic we cover in why your website needs to work as a source, not a megaphone.

Interaction feedback that surprises the visitor breaks trust instantly

Good interface design tells a visitor, instantly, that a click registered and that an element is clickable. When AI-generated interactions ignore that convention, the result feels like walking into a glass door in a room you thought was open. It is disorienting precisely because it violates an expectation the visitor did not know they had until it was broken.

The discussion flagged a specific set of interaction habits worth auditing on your own site:

• Decorative lines that trail the cursor down the page
• Cursor-following light or glow effects
• Background animation with no functional purpose
• Automatic fade-ins on every section
• Buttons or UI elements that shift position unexpectedly
• Hover effects that move things without communicating anything
• Animation that pulls more attention than the product itself does

Each of these individually is a small annoyance. Stacked together on one page, they train the visitor to distrust the interface.

Scroll hijacking was the single most common complaint in the review

Of every UX problem the pair encountered, hijacked scrolling came up most often — four separate times on four separate sites. This is when native browser scrolling gets intercepted by JavaScript so the page can choreograph a scroll-triggered animation, and it frequently backfires by locking the visitor in place mid-scroll while an animation finishes building itself.

Schaad described the experience bluntly as “going through molasses.” Epstein’s complaint was more specific: on one site, an elaborate line-drawing animation captured his full attention while the actual explanatory copy sat unread in the corner of the screen. The animation was working exactly as coded. It was also actively preventing the page from doing its job, which is communicating what the product does.

Beyond scroll jacking, the broader category of broken UX patterns includes:

• Non-standard or unpredictable navigation
• Menus that jump or behave inconsistently between pages
• Elements that look clickable but do not respond as expected
• Buttons that move or auto-advance without user input
• Functionality hidden entirely behind hover states
• Sticky headers that behave inconsistently across sections

The common thread is that the page stops behaving the way visitors expect a website to behave, and every deviation from that expectation adds friction the visitor did not sign up for.

A beautiful page that doesn’t explain the product has failed at its job

This is where AI-assisted design and AI-assisted copy compound each other. A visually striking landing page paired with vague, jargon-heavy copy leaves the visitor no better informed than before they arrived. The pattern is familiar from years of B2B “human slop” writing, except AI can now generate it faster and in greater volume.

The tell is language that sounds sophisticated but does no actual work: words with multiple meanings used loosely, sentences that describe a category instead of a product, and copy that never quite states who the product is for or why it matters to them. A landing page exists to convert a visitor into a lead or a customer. If someone cannot tell what you sell within a few seconds of arriving, the design succeeded and the page still failed. Every element on a page — text or visual — needs to be doing work, not just occupying space that looks intentional.

Too much competing structure is as bad as too little

AI tools are good at generating structure: headers, subheads, callout boxes, badges, dividers. What they are not good at is judging whether that structure serves the reader or just looks busy. The result is often a page with too many competing text styles, decorative labels that add no real meaning, and a weak visual hierarchy between the logo, the headline and the supporting copy.

The fix is subtractive, not additive. Before adding another visual element, ask whether it clarifies something the reader would otherwise miss. If the answer is no, it is decoration pretending to be information architecture.

A brand needs one visual system, not several stitched together

Vibe-coded sites frequently look like they were assembled from different projects: a hero section in one style, a features grid in another, product visuals that clash with the landing page around them. Individually, the assets can be attractive. Together, they signal that no one made a deliberate call about what the brand should look like — the AI was simply asked for “modern” or “trendy” repeatedly, section by section, with no unifying direction behind it.

This is the same failure mode showing up at the visual layer that shows up at the messaging layer: default choices standing in for brand strategy.

None of this is really about the tools

Every failure mode above traces back to the same root cause: AI removes technical friction, not the need for judgment. The tools make it trivial to produce something that looks finished. They do not make it easy to know whether that output actually serves the visitor or the business.

The habits that produce weak sites are recognizable: accepting every AI suggestion without editing it, letting the model set brand direction by default, starting from generated output instead of a strategy, and treating saved time as a license to add more effects rather than think more clearly about the page. The person driving the tool is still the editor. If nobody plays that role, the slop just arrives faster than it used to. This connects directly to a broader shift worth understanding: as AI agents increasingly browse and evaluate sites on a user’s behalf, the question of whether your website is actually ready for that traffic matters as much as whether a human visitor likes it.

Frequently asked questions

What are the most common AI website mistakes?

Generic design defaults, jarring interaction feedback, scroll jacking, weak product messaging, cluttered information hierarchy, an inconsistent visual system, and over-reliance on AI judgment without human editing. Each is easy to catch once you know to look for it.

Is vibe coding a website a bad idea?

No. It removes a real barrier for people without design or development backgrounds. The risk is treating AI’s first output as final rather than as a draft that still needs judgment, editing and a clear brand direction applied on top of it.

Why do so many AI-built sites look alike?

Models are trained on whatever gets linked to and copied most. A distinctive design choice that used to take months to spread now reaches every similar prompt within days, because the model absorbs it into its defaults rather than one designer copying another by hand.

What is scroll jacking and why is it a problem?

Scroll jacking is when JavaScript intercepts native browser scrolling to control a page-triggered animation. It frequently locks visitors in place mid-animation, which reads as broken rather than intentional, and it was the single most common complaint raised in the review discussed here. See also: Google ads budget misallocation is more common than you.

How do I know if my AI-generated site has a messaging problem?

Ask someone unfamiliar with your product to look at the homepage for ten seconds and explain what it does and who it’s for. If they cannot, the copy is doing decoration instead of communication, regardless of how polished the design looks.

What’s the single biggest fix for these mistakes?

Treat AI output as a first draft, not a finished product. Every section, animation and line of copy should have to justify its presence rather than being kept just because the tool generated it.