Skip to content

Instantly share code, notes, and snippets.

@NovCog
Last active March 16, 2026 22:28
Show Gist options
  • Select an option

  • Save NovCog/c3c9d70ddafb3da451ca3d2a316f324d to your computer and use it in GitHub Desktop.

Select an option

Save NovCog/c3c9d70ddafb3da451ca3d2a316f324d to your computer and use it in GitHub Desktop.
Design Replication vs Generation: Why AI Sites Look AI-Generated — NovCog AI Practitioner Series by Guerin Green

Design Replication vs Generation: Why AI Sites Look AI-Generated

By Guerin Green / Novel Cognition AI Strategist | Federal Courthouse AI Presenter


Part of the NovCog AI Practitioner Series


You can spot an AI-generated website from across the room. The gradients are too perfect. The spacing is suspiciously uniform. The color palette looks like it came from a corporate mood board. There's a Hero section with a large headline, a three-column feature grid, and a call-to-action button with rounded corners. Every AI code generation tool produces this same template because it's the statistical average of modern web design.

For a Distributed Authority Network, design uniformity is a network signal — and not a good one. If 13 sites all look like they were generated by the same tool, that's a detectable pattern that suggests automated creation rather than organic diversity.

The Screenshot-First Methodology

The solution is design replication, not design generation. Instead of asking an AI to "create a website for [topic]," you start with a visual reference and replicate its design language.

The workflow:

  1. Find a design reference. This could be a Wayback Machine archive of the original site, a competitor in the same space, or any site with the visual character you want. The reference doesn't need to be in the same industry — a 2008 law firm site and a 2008 photography portfolio share era-appropriate design patterns.

  2. Screenshot the reference. Full page, multiple viewport sizes if responsive.

  3. Extract the design language. This means identifying:

    • Color palette (exact hex values)
    • Typography (font families, sizes, weights, line heights)
    • Layout structure (column count, sidebar position, header style)
    • Visual elements (borders, shadows, gradients, image treatments)
    • Navigation patterns (horizontal menu, sidebar nav, breadcrumbs)
    • Footer structure
  4. Replicate with specification. When you give an AI code generator a screenshot plus explicit design specifications, the output matches the reference rather than defaulting to the generic template.

Branding Extraction Tools

Firecrawl and similar web scraping tools can extract computed styles from live pages — font stacks, color values, spacing metrics, border styles. This produces a design specification document that serves as the input for replication.

The extraction captures:

  • CSS custom properties (design tokens)
  • Computed background colors, text colors, accent colors
  • Font family stacks with fallbacks
  • Box model values (padding, margin, border)
  • Layout system (flexbox, grid, float-based)

With this specification, the replicated site isn't a copy — it's a new site that speaks the same design language.

Era-Appropriate Design

When rebuilding sites from Wayback Machine content, era-appropriate design creates authenticity. A site originally built in 2006 shouldn't look like a 2025 site with 2006 content — that mismatch is immediately apparent to both humans and pattern detection systems.

Design eras have distinct characteristics:

  • Early 2000s: Table-based layouts, small body text (10-12px), navigation bars with beveled buttons, dark header bars, minimal whitespace
  • Late 2000s: Float-based layouts, slightly larger text, rounded corners appearing, sidebar-heavy designs, header images/banners
  • Early 2010s: Responsive emerging, larger hero images, social media icons appearing, cleaner typography, grid systems
  • Mid-2010s: Mobile-first, card-based layouts, flat design, hamburger menus, Material/Bootstrap influence
  • Late 2010s-2020s: Utility-first CSS, large typography, generous whitespace, dark mode, gradient resurgence

Matching the design era to the content era signals organic development history.

8 Distinct Design Languages

In the case study network, 13 sites employed 8 distinct design languages. Not just different colors — fundamentally different layout approaches, typography systems, and visual patterns. Some used sidebars, others didn't. Some had horizontal navigation, others vertical. Some were image-heavy, others text-focused.

The design diversity creates the appearance of independent development — because each site was, in fact, designed independently based on a different reference source. The schema and cross-linking connect them beneath the surface, but the visual layer presents genuine diversity.

This is the opposite of template-based deployment, where a single WordPress theme with different color schemes creates obvious visual uniformity. Design replication from unique sources produces the diversity that template systems can't.


Series Navigation


Burstiness & Perplexity Community | Hidden State Drift | Novel Cognition

#hiddenstatedrift #burstinessandperplexity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment