RESEARCH · AI DESIGN TELLS
How many products look like they were generated, measured
A published study by Adrian Krebs scored 1,590 Show HN submissions against detectable AI design patterns and found 22% carried four or more, and 32% carried two or three. Only 46% were clean. The patterns are measurable in CSS and DOM rather than judged by eye, which makes the sameness of generated design an observable fact rather than an opinion.
WHAT WAS FOUND
Just over half of new product launches carry at least two detectable marks of generated design.
In a study published by Adrian Krebs, 1,590 Show HN submissions had their landing pages scored automatically against a list of AI design patterns. The checker loads each page in a headless browser and asks deterministic questions of the DOM and computed CSS — no image analysis, no judgement.
The corpus matters when reading those numbers. Show HN is where indie makers and solo builders launch, which is the population most likely to have used a generator. The equivalent number for funded startups has not been published, and it is the version we care about — that is the corpus we intend to run.
WHY THIS IS NOT AN OPINION
Design criticism is usually taste, and taste is arguable. This is not. Every pattern below is a CSS or DOM check with a yes or no answer, the code is open source under an MIT licence, and anyone can run it against any site including ours. That is the whole reason the finding is worth citing.
THE FOURTEEN TELLS
None of these is a mistake on its own. Each is a reasonable decision that a competent designer might make deliberately. The signal is not any single one — it is how many appear together, on sites that have never met each other.
OUR OWN SCORE
One of fourteen. And we are not going to pretend it is zero.
We scored this site by hand against the same criteria before publishing any of this, because a studio selling differentiation should be willing to be measured by its own rubric.
The one we carry is the FAQ accordion, on the service pages and the FAQ page. We kept it deliberately: it is the format assistants extract from most reliably, and being findable is worth more to us than a clean score. That is the honest version of the trade — and it is the point of the whole exercise. The question is never whether you used a common pattern. It is whether you chose it.
The rest we do not carry: no Inter-centered hero, no purple, no gradients, no glow, no glass, no dark-mode default, no pill badge, no stat banner, no emoji.
WHAT WE ARE RUNNING NEXT
The same method, on the population that actually buys design.
Show HN answers a question about indie launches. It does not answer the one founders ask us, which is whether funded companies in their category look alike. So the run we are preparing is a defined corpus of recently funded AI startups, scored against the same open-source checks, published in full with the method and the raw list.
No numbers from that run appear on this page until it has been run. A benchmark with estimated figures is not a benchmark, and the entire value of this page is that its numbers are checkable.
WHAT TO DO ABOUT IT
Not: avoid all fourteen. That is just a different template.
A site that carefully avoids every pattern on the list is still designed by a checklist. The useful reading is narrower — pick the two or three places where your product is genuinely different from every alternative, and make those the places that look different. Everywhere else, a familiar pattern is doing its job.
The other half is upstream of visual design entirely. Most sites that read as generated also read as generic in their words: a headline that names a category rather than a problem, three feature cards, a stat row. Fixing the sentence usually does more than fixing the gradient.
Source: Scoring Show HN submissions for AI design patterns by Adrian Krebs, and the ai-design-checker tool, MIT licensed.
QUESTIONS
Why do AI-built websites all look the same?
Because generative tools optimise for a plausible result, and plausible is an average. Every product built with the same tools converges on the same average, which is now specific enough to be detected in CSS: one sans-serif, a centered hero, a pill badge above the headline, a purple gradient, permanent dark mode, icon-topped cards.
How can design sameness be measured objectively?
By checking the DOM and computed CSS rather than judging by eye. A headless browser loads the page and a script asks concrete questions: is the accent colour in the indigo-violet range, is there a pill badge directly above the H1, is the body text grey on a dark background. Each check is deterministic and repeatable.
Is using AI to build a site a bad idea?
No. It is the correct choice for shipping fast, testing an idea, or building something internal. The cost is differentiation, and that cost only matters once you are being compared to someone selling the same thing. Ship with it, then decide what should be different once you know what you sell.
Does looking generic actually lose deals?
It removes a reason to choose you rather than creating a reason to reject you, which is harder to see in your own analytics. The observable version is buyers confusing you with a competitor, or investors reading you as earlier-stage than you are.
Can you check our site against these patterns?
Yes. It is part of what a UX audit covers, and the criteria are public and open source, so you can also run them yourself. We would rather you knew the answer than paid us to tell you.