Reads the whole site. Not just one page.
Multi-page Extraction captures more of a site's real colors, type, and components — for richer, more complete, more accurate results. And because generation now runs in the background, you never wait on the page for it.
How much of a site's real design ends up in your DESIGN.md.
Colors and type that are well-formed and ready to drop into code.
How often multi-page wins a blind quality comparison (50% = tie).
How often multi-page wins on completeness (50% = tie).
Screenshot-grounded rubric, scored out of 10.
Single-page vs. multi-page
The same sites, generated both ways.
- Reads only the landing page
- Misses tokens used deeper in the site
- 91.5% of design tokens captured
- 90% of tokens well-formed
- Reads across the site's key pages, not just one
- Captures tokens from real product & content pages
- 96.8% of design tokens captured
- 95% of tokens well-formed; preferred 64% of the time
Results by site type
Quality score out of 10, and how often multi-page wins head-to-head.
Multi-page helps most where design lives on pages beyond the homepage (vercel, react). Simpler or already-thorough single pages stay flat or can dip (allbirds, github) — it's a strong default, not a universal win.
Quality holds steady across dimensions
Rubric scores (out of 10) — roughly flat to slightly up, with the biggest gains in visual fidelity and completeness.
Reading more pages, without the wait
Multi-page does a little more work up front. That used to mean waiting — not anymore. Generation now runs securely in the background, so you can close the tab and come back to a finished result. We'll even email you when it's ready.
Figures are the aggregate results of our internal output-quality benchmark: single-page vs. multi-page across five representative sites (SaaS marketing, SPA/marketing, e-commerce, product/docs, docs), up to 4 pages crawled per site, scored by a screenshot-grounded rubric ensemble (Claude Haiku). Gains are largest on content-rich sites and mixed on others; multi-page does a little more work up front, and round-trip brand reconstruction was flat (5.4 → 5.4 / 10). Small sample — directional, not definitive. Your results vary by site.