Home › Research › Design Diversity
Manva Research · methodology

Why AI websites all look the same — and how we engineer against it

A transparent account of our design-quality method: the 50-sites-per-industry bar, the weakest-5 loop, and the rule that only visual review counts.

Last updated 2026-08-12 · by the Manva team

The problem with AI websites in 2026

Everyone can generate a website now. Look at fifty of them side by side and you see the actual problem: they're the same website. Same hero layout, same three-card section, same safe palette — a template factory with an AI accent. For a business, sameness is fatal: your website's first job is to not look like your competitor's.

This page documents, transparently, how Manva's generation pipeline is built to fight sameness — and the review method we use to keep it honest. It's an engineering story, not a benchmark; where we publish numbers in future they'll come with methodology attached.

The acceptance bar we set (verbatim, from our internal rules)

Our design pipeline is developed against a written standard: generate 50 websites per industry — jewellery, restaurants, salons, doctors, fashion, interiors, manufacturers — put them side by side, and require that they look like they were designed by different professional agencies. None may feel like a palette swap of another. Every site must fit its industry and brand positioning. A merchant should be proud to publish it immediately.

Two internal rules worth quoting because they shape everything:

  • "Do not stop when the selector is diverse; stop when the OUTPUT is diverse." — engineering metrics (how many palettes, fonts, layout combinations exist) are diagnostics, never completion evidence. Only visual review counts.
  • "The weakest generated website defines the quality of the product, not the average." — after every improvement we re-generate, review visually, find the weakest five, and fix the pipeline until those disappear. We call it the weakest-5 loop.

How the pipeline creates difference (not variation)

  • A Creative Director layer, not a template picker. Positioning first: a heritage Banarasi house and a Gen-Z streetwear brand should not share a design language even within "fashion". Layout, typography hierarchy, color harmony, image treatment, navigation style, card style, CTA placement and motion are composed per business.
  • Design systems over fixed templates. There are no "template #1–40" files to rotate through — composition happens from design-system parts, which is why two jewellery sites can differ structurally, not just chromatically.
  • Preview equals live. What the merchant reviews is exactly what publishes — an internal invariant we treat as launch-critical, because confidence to publish IS the product.
  • SEO ships with every site. Each published site carries its own sitemap.xml, robots.txt and per-page metadata — discoverability isn't an upgrade tier.

What we'll publish next (and how to hold us to it)

As the review corpus grows we intend to publish the side-by-side galleries themselves — per-industry grids of generated sites, with the weakest-5 iterations visible — plus aggregate, anonymized findings about what Indian SMB websites need (section patterns by industry, language mix, mobile behaviour). Methodology will be published with every number. If you're a journalist or researcher who wants early access to the galleries, contact us.

Judge the output yourself

Describe your business and look at what the AI designs — the review loop above is the reason it won't look like anyone else's.

Start free — build your site in minutes

Keep reading

Frequently asked questions

Is this a study with published data?

Not yet — it's a transparent account of our design-quality method (the 50-sites-per-industry bar and the weakest-5 review loop). Published galleries and aggregate numbers come next, with methodology attached; we deliberately don't publish figures before then.

Why do AI-generated websites all look the same?

Because most generators select from templates and vary colours/fonts — variation without difference. Structural difference requires composing from design systems with a positioning step first, then rejecting look-alike outputs in review.

How is design quality measured at Manva?

Visually, side by side, against the bar that 50 sites per industry must look like the work of different agencies — with the explicit rule that engineering diversity metrics are diagnostics, never completion evidence.