How the Corpus Was Built

How we got all this information

The list of a million domains is Tranco (tranco-list.eu), which averages several popularity rankings. From there:

Stage Domains
Tranco top 1M 1,000,000
After the domain filter (CDNs, ad networks, a handful of adult domains by name) 995,467
Alive on a fetch 710,284
English 323,568
Content-qualified (has CSS, not thin, not a login wall, not parked) 125,721
Rendered successfully in headless Chrome, including a stealth re-render pass that recovered bot-blocked domains 226,643
Excluded afterwards: one operator’s 35,611-site link farm, 9,263 gambling, 8,138 adult, 9,102 placeholder pages, 3,955 bot walls, 1,085 parked, 923 error pages, 279 redirect shells 68,356
Clean corpus 158,294

Each site was rendered once, homepage only, in headless Chrome, in 17 batch runs between 29 April and 27 May 2026, plus 77 re-rendered in October. The viewport was 1280 by 800 on about two thirds of renders and 1366 by 768 or 1920 by 1080 on the rest, because the recovery pass rotated viewports with its user agents. From each render I kept the HTML, every stylesheet (external and inline, deduplicated by content hash: 1,734,591 unique sheets, 56.7 GB), the browser’s computed style for one representative element of each of eleven roles (body, a, p, h1, h2, h3, button, input, header, nav, footer), and every color used, by role, with how many elements used it. The stylesheets were parsed separately for media queries, units, clamp, container queries and so on.

Sectors come from an LLM pass (DeepSeek via Ollama, 84 labels fixed by hand) that read each page and assigned one of 28 sectors and one of 12 site types. The earlier keyword classifier agreed with it on only 34% of sites, which is why I don’t use the keyword labels.

Caveats, because there are real ones:

  • Homepages only. An article page or a product page has different typography. Everything here is the front door.
  • English only. 323,568 of 710,284 live sites. The CJK and RTL web have different norms and aren’t in this.
  • One moment. April to May 2026. No trend over time; this is a snapshot.
  • Desktop viewport. The computed styles are the desktop styles. Mobile type sizes and spacing are in the CSS but not in the computed-style tables.
  • One element per role. The “h1” is the first H1, and in the Stage 6 tables the “button” is the first button-like element in the DOM, which is often a menu toggle. That’s why one H1 in seven is under 24px; it’s real, but it’s the first element, not the hero. The Buttons section no longer uses that button: it was re-measured with a designed-button definition (border or fill, plus padding, outside header, nav, footer and cookie bars), and the first-pass figures stay only as the comparison column.
  • Sheet counts include libraries. When I say a site has 17 breakpoints or 22 stylesheets, that includes every plugin and library it loads. That’s a true statement about what the browser receives and a misleading one about what the designer decided, and I’ve tried to say which is which.
  • Detection is pattern matching. Framework, CMS, theme and library detection work from file names, class prefixes and known colors. They miss renamed or bundled builds and they can double-count a site that ships a component library.
  • The link farm was found by hand. The 35,611-site template was only noticed because one number looked too neat. There may be smaller ones still in the clean set; none big enough to move a top-ten number, as far as I can tell.

Everything is in three SQLite databases on my desk and queryable. If a number here looks wrong, the query that produced it is listed below and can be re-run.

Sources

All numbers are from the clean corpus of 158,294 sites (design.db domain_tags minus excluded_domains) unless a table says otherwise. Scripts live in pipeline/reportv2/.

  • Fonts. Top families per role, fonts per site, pairings, H1 weight and size bands, uppercase shares: compute_v2.py → results_v2.json (top_fonts, font_count, pairings, weights, h1_size, uppercase, body_mode), from computed_styles. Google Fonts, Typekit and Bunny links, system-stack shares, H1 letter-spacing, H1-to-paragraph ratio, exact size histograms: article_probe.py → article_results.json (google_fonts_link_sites, typekit_sites, bunny_sites, body_font_system, site_font_system, letter_spacing, h1_p_ratio, h1_px, p_px), from stylesheet_links and computed_styles. Font-size units per sheet and site: fontsize_units_probe.py → fontsize_units_results.json, from css_features.db blob_property_units. The pairing template check (Roboto over Verdana is tagDiv Newspaper on 97% of 824 sites; two thirds of the 1,021 clean Newspaper sites keep its default fonts, 674 by href; Linux Libertine over sans-serif is MediaWiki or a Wikipedia mirror on 243 of 246): pairing_template_probe.py → pairing_template_results.json (pairings, cohorts, newspaper_corpus_wide, roboto_verdana_classification, linux_libertine_classification, poppins_over_system_verdict, comparison_with_27oct), from stylesheet_links (hrefs under /wp-content/themes/Newspaper/, td-composer, td-standard-pack, td-cloud-library; load.php and /homepage/media/style.css for the Libertine cohort) and computed_styles for the pairing cohorts, clean corpus. The check was first run as a scratch session on 27 October 2026, which counted 1,045 and 675; the probe’s reproducible href rule gives 1,021 and 674 (1,031 counting child theme directories), and those are the figures the essay uses.
  • Breakpoints. Occurrence counts: results_v2.json (breakpoints) from css_features.db blob_breakpoints. Per-site distinct widths, direction, site shares of each value: article_results.json (bp_*), joining blob_breakpoints to stylesheet_links on css_hash. Site counts for 1440, 1600, 1920, 1536 and 1366px: the breakpoints side session’s cache, built with the same join (hand-over note in the shared notes folder). Bootstrap and Tailwind breakpoint scales: their current documentation. The Tailwind 4 range-syntax gap: pipeline/stage8a_parse_css.py matches only min-width: and max-width:. Stylesheets per site by platform: article_results.json (css_weight_by_cms).
  • White space. Leading ratios, letter-spacing, padding and margin values, 4px/8px grid shares: article_results.json (leading, letter_spacing, spacing), from computed_styles.
  • Buttons. The designed-button pass (a button is a border or a fill plus padding; first visible match outside header, nav, footer and cookie or consent containers; 157,549 clean sites re-rendered offline from the stored markup and stylesheets with JavaScript off, 94,424 buttons, 82,123 with a fill): designed_button_probe.py render-offline-v3 all → designed_button_raw_offline_v3.jsonl and designed_button_results.json (clean_corpus_offline_v3_no_footer_no_cookie for every figure in the section; clean_corpus_offline_v3run_v2_rule for the footer and cookie accounting). The exclusion lists are EXCLUDE_ANCESTORS, FOOTER_ANCESTORS and COOKIE_ANCESTORS in the probe. Run with the Stage 6 python; about two hours at eight pages at a time. The offline render reproduces Stage 6’s stored computed styles on 93 to 95% of sites (render-offline-validate). Hover, focus, active, transition and animation shares come from a second pass, render-offline-fp → designed_button_raw_offline_fp.jsonl, which walks each site’s stylesheets and tests every pseudo-class rule’s selector against the chosen button. The ten most repeated buttons are grouped by exact computed fill, text color, radius, size, weight, padding, border, case, shadow and letter-spacing. The standard page’s button and other first-pass figures are the Stage 6 button role: results_v2.json (btn_radius, btn_shadow_pct, weights.button) and article_results.json (button_contrast, button_bg_top, button_size, button_border).
  • By industry. Sector cards: sector_look_probe.py → sector_look.json (modal font, size, weight, leading, color and button per role, per sector). results_v2.json (sector_table, dark_by_sector, btn_radius_by_sector, fonts_by_sector, fw_by_sector) and article_results.json (sector_typo, sector_body, sector_button, sector_fontcount, sector_gf, sector_bp, clamp_by_sector), sectors from domain_classification.
  • Frameworks, themes, stacks. results_v2.json (css_framework, cms, both_fw, fingerprints, combos_bands, scale); theme slugs: theme_probe.py → theme_results.json from stylesheet_links; CSS weight and single-font share by CMS and framework: article_results.json (css_weight_by_cms, css_weight_by_fw, fontcount_by_cms) from domain_css_weight. First-edition comparisons: viz/css_design_report_v2.html against viz/css_design_report.html.
  • Trendy. Sheet-level clamp, container and units: results_v2.json (features). Per-site clamp and container, by CMS, framework and sector: article_results.json (site_features, clamp_by_cms, clamp_by_fw, h1_fractional_px); which stylesheets carry them: clamp_source_probe.py → clamp_source_results.json. Libraries, dark-mode toggles, variable fonts, background saturation, accent hue, palette size: trends_probe2.py → trends_results2.json (15,001-page HTML sample and the palette table). Modern-CSS, motion, image and dark-class shares: viz/design_signals.html section 9 (60,000-sheet and 12,000-page samples, 15 October 2026) and pipeline/checkpoints/signature_probe_results.json.
  • Funnel and method. viz/design_guide.html, sites.db stage columns, design.db renders (viewport, dates), render_runs, excluded_domains by reason.