AI-generated websites look the same because a model built them: every model draws from the same training data and reaches for the same component libraries. The look got its own parody this week. Slopstrap, a satirical CSS framework that makes any site look hastily vibecoded, is on Hacker News with 528 points and 95 comments as of November 21 (discussion).
Mike Lacher built it as a joke (site) and describes Slopstrap as a “load-bearing accelerant to make anything look like it was made by an LLM.” It ships fake status dots, gradient headlines, and testimonials from a Dr. Sarah Chen who does not exist. Every component is a default that agent-built pages reach for constantly, which is why the joke landed.
TL;DR: The parody packages every tell of AI-generated websites: glowing orbs, gradient headlines, stat counters reading “14.8M+ Users, probably,” fake testimonials. The tells converged because the models and component libraries converged. Stripping them is cosmetic. What restores trust is verifiable specificity: dated numbers, named sources, claims a reader could check.

What are the tells of an AI-generated website?
Slopstrap’s component list reads as a field guide:
- Blinking status dots (“AI Engine Active”, “Thinking Really Hard”)
- Glowing purple orbs over a dark grid background
- Gradient headlines in Inter, body copy in low-contrast gray
- One italic serif word in a headline (“Build the future of work”)
- Fat colored left borders on cards nested inside cards
- Bento grids fracturing paragraphs into glowing tiles
- Scroll-animated stat counters (“+420% YoY”)
- Logo marquees and testimonials from companies you never worked with
The site’s own alert copy lands the punchline: “Your landing page now looks identical to 12,400 other YC batches.”
Why does every agent produce the same page?
Three layers stack. Models share training data, so they converge on the same taste. Component libraries share defaults, so a card-layout request returns the same rounded, glowing, left-bordered card everywhere. And the loop is tightening: agents read other agents’ output as examples, which pulls the average toward whatever shipped last month. Slopstrap leans into that loop: it ships an /agents page so an AI agent can apply the whole look on purpose.
This is the build side of a problem we’ve covered from the distribution angle: vibe coding made shipping fast, and when everyone ships the same default, looking generated stops being neutral and becomes a signal.
What is the joke actually mocking?
The orbs and gradients are the recognizable part, but the sharpest jokes target the fake data: stat counters with “probably” baked into the label, and testimonials from people who were never customers. That is where the trust problem actually lives.
The site’s demo makes it in one toggle. The before is a plumber’s page: serving Springfield since 1998, licensed technicians, a phone number, weekend hours. Every line is checkable. Flip Slopstrap on and the same facts drown in orbs and stat counters. The information didn’t change. The verifiability did.
So removing gradients fixes nothing: you can de-slop the cosmetics and keep publishing unverifiable claims, and readers will still smell it, because the smell tracks the claims. The style is just where they look first.
What actually fixes it?
Verifiable specificity, and it applies to copy as much as design. Put a date next to every number, a name next to every source, and ask whether a skeptical reader could check each claim in under a minute. One real, small number beats a big invented one.
That is the discipline this blog runs on. A written rubric requires every post to carry real numbers with context and named-source links; two drafts sit unpublished right now because they didn’t clear that bar. We publish the small numbers too: this blog runs at single-digit daily pageviews, and our full SEO timeline is public. The demand side pays it back. This week our two highest-traffic organic landing pages are narrow, query-exact posts with hard numbers in them.
The coding agent that built your site is optimized to ship, and shipping was never the trust problem. Making the output checkable is a separate pass, and no model runs it by default.
If your agent-built site has started to smell like everyone else’s, that’s fixable in one pass. CrossMind audits your pages against a verifiable-claims checklist in one afternoon. The output is a page-by-page list of claims that need a real number, a date, or a source. https://agent.crossmind.io