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.

The dead internet theory says that most of what is online is made by machines for machines, and that the people are a minority on their own network. I can’t measure the whole web. I can measure one operator. This is one person, or one small company, that put 35,611 websites into the top million, every one of them a magazine with today’s date on the front page, a byline, a sensible menu and about a hundred links out, and not one reader. It is the biggest single thing I have found on the web, and it is nothing.

37,154
domains built from one page skeleton, found in a list of 40,174 sites that invite guest posts. 92% of the list.
35,611
of them inside the Tranco top million. One site in every 28 on the list belongs to this operator.
1
analytics account, one storage bucket, one email address and one WhatsApp number behind all of them.
8,232
domains registered in March 2026 alone, the month before I crawled them. 18,689 of 18,735 dated registrations are from 2025 or 2026.

analysis.db cluster_members for cluster 507 and whois_records joined to it; design.db pbn_domains for the Tranco count; tracking_ids, dom_fingerprints and contact_methods for the shared account signals. All of it is laid out in the tables below.

They pass as the web. A keyword classifier that reads only the title and description filed them as entertainment (7,302), technology (6,409), business (3,680) and news (2,930), and 2,697 of them as gambling. Spread across the industries like that, they look like a cross-section of what is online, and in a count of sites they are 35,611 votes for whatever they happen to be doing. A template repeated 35,000 times is indistinguishable, at the site level, from 35,000 sites that agree.

Here are twenty of them, drawn at random from the 37,154 and captured on 29 October 2026. Scroll sideways. The nav bar is the same on every one: Home, All Posts, Business, Esports, Fashion, Featured, More, a search icon and a dark-mode toggle. The lead stories are dated the day of the visit.

vertexspinzone.com homepage, 29 October 2026
vertexspinzone.com · Vertex Spin Zone: Cutting-Edge Tech Insights & Innovations
hyperarenaspace.com homepage, 29 October 2026
hyperarenaspace.com · Hyper Arena Space – Exploring Tech Innovations & Future Trends
pointsocietyblog.uk homepage, 29 October 2026
pointsocietyblog.uk · Point Society Blog: Insights on UK Culture, Politics, and Lifestyle
spreadparlorrise.com homepage, 29 October 2026
spreadparlorrise.com · Spread Parlor Rise: Culinary Adventures and Food Blogging
megapalacenetwork.com homepage, 29 October 2026
megapalacenetwork.com · Megapalace Network: Tech Insights, Innovation & Digital Trends
fortuneworld.us homepage, 29 October 2026
fortuneworld.us · FortuneWorld: Insights on Wealth Building and Financial Freedom
spinxnova.us homepage, 29 October 2026
spinxnova.us · Spinxnova – Exploring Tech Innovations and Digital Trends
slotsinsidertalk.co.uk homepage, 29 October 2026
slotsinsidertalk.co.uk · Slots Insider Talk: UK Casino News, Reviews & Strategies
casinoinfogazette.co.uk homepage, 29 October 2026
casinoinfogazette.co.uk · Casino Info Gazette: UK's Latest Gambling News & Reviews
lunararenalounge.com homepage, 29 October 2026
lunararenalounge.com · Lunar Arena Lounge: Space-Themed Gaming & Sci-Fi Entertainment Hub
luckdaily.de homepage, 29 October 2026
luckdaily.de · Luck Daily: Your Daily Dose of Inspiration and Positive Thinking
dynamiccasinoworld.com homepage, 29 October 2026
dynamiccasinoworld.com · Dynamic Casino World: Latest Gaming News and Casino Reviews
slotsignal1247.de homepage, 29 October 2026
slotsignal1247.de · SlotSignal1247 – Expert Insights on Online Casino Slots & Strategies
infinitecardsworld.com homepage, 29 October 2026
infinitecardsworld.com · Infinite Cards World: Discover Collectible Card Game Insights
truepokerzone.com homepage, 29 October 2026
truepokerzone.com · True Poker Zone: Expert Tips, Strategies & Reviews for Poker Players
dailydigital.uk homepage, 29 October 2026
dailydigital.uk · Daily Digital UK: Tech News, Reviews & Digital Lifestyle Insights
legacyspincentral.com homepage, 29 October 2026
legacyspincentral.com · Legacy Spin Central: Classic Tech Reviews & Retro Computing Insights
pulsegaming997.com homepage, 29 October 2026
pulsegaming997.com · Pulse Gaming 997 – Latest Gaming News, Reviews & Community Updates
clearbet.ca homepage, 29 October 2026
clearbet.ca · ClearBet Canada: Expert Sports Betting Tips and Casino Reviews

Headless Chromium, 1280 by 900 viewport, one capture per homepage. The sample is 20 domains from cluster 507 in analysis.db, seed 20261029 over the sorted member list; 19 are shown because the twentieth, soledrive.it, no longer resolves. The template has been refreshed since the April crawl: the Bootstrap 5 magazine described below has become a dark newspaper layout with a ticker, a Write for us link and a Contact link in the top strip. Whether the new head still hashes to 80abf477 has not been checked. The files are in shots/farm-sample with a manifest.

How it was found

The farm

The farm was found by accident, in a project about something else. In April I had a 43 MB CSV of 40,174 domains, every one of them a site in the Tranco top million with a public “write for us” page, and I wanted to know who owned them. What I typed at the start of that session, on 25 April: “i think there is a way to discern networks here. there are sites that link to each other, a hierarchy of sites. networks of spam websites. that’s what a lot of these are. i want to see which ones are related.” The plan was blunt. Get the HTML for all of them. Check for repeated patterns, same articles, same themes on the back end, same builds. Then a registration check. Then an email search.

The crawl took 44 minutes that morning: 40,174 homepages and 2,097 submission pages, 65 failures, stored raw in a 7 GB SQLite file. The signal pass that evening ran over 42,206 pages in five minutes and computed, among other things, a hash of each page’s head: the tags and attribute names in document order with every value stripped out, so that two sites built from one template hash the same even when their titles, colors and analytics IDs differ. The instruction was “just do all of them, match for similarities in the head.”

The first number looked like a bug. One head hash covered 37,154 of the 40,174 domains, 92% of the list. The first reading, in the session log at 07:01 UTC, was that this was not a network but a hash collision on something common, probably parked domains or a Cloudflare interstitial. A second anomaly in the same batch was waved off the same way: a Matomo analytics site ID of 2 on 37,220 domains, 23 distinct values in 37,255 matches, written down as a regex matching some unrelated library. Both readings were wrong in the same direction. Five sample pages settled it a minute later. The pages were 117 to 425 KB of real templated content, not parking pages, and every one loaded its favicon from the same Contabo object-storage bucket. The bucket was named nextbacklinks. A thousand-page sample came back 1,000 for 1,000. The Matomo tag was real as well: it reports to analytics.tech-vikings.com, site ID 2, on every one of 2,000 farm pages I sampled again for this piece. The operator was counting its own traffic in one account, and the pipeline had recorded that fact on the first pass and filed it under noise.

dom_fingerprints.head_skeleton_hash 80abf477… on 37,154 domains; tracking_ids id_type matomo, id_value 2 on 37,179; pages.db html for the favicon href and the Matomo snippet; session log of 2026-04-25, 07:01 to 07:03 UTC. Matomo was left out of the cluster build as “too noisy” and still is.

The rest of the week added the other signals, and they all landed on the same set of domains.

What 37,154 domains shareDomainsWhere it sits
One head skeleton, hash 80abf477…37,154dom_fingerprints, cluster 507
Favicon from Contabo bucket 83226f39…:nextbacklinks37,154dom_fingerprints.favicon_url, cluster 595
Matomo site ID 2 on analytics.tech-vikings.com37,179tracking_ids
bloggar@gmail.com in the footer7,601contact_methods, cluster 742
links@nextbacklinks.com, behind Cloudflare’s email obfuscation3,771cluster 794
WhatsApp +971 5434499513,771cluster 864
hello@brand.com and +1234567890, template placeholders never filled in3,827clusters 848 and the matching email
Nameservers ns1 and ns2.dyna-ns.net3,444whois_records, cluster 872
Registered in 2025 or 202618,689 of 18,735 datedwhois_records

analysis.db clusters, cluster_members and cluster_profiles for cluster 507 and the clusters that overlap it (7,593 of the 7,601 bloggar domains, 3,768 of the 3,771 nextbacklinks domains and 3,439 of the 3,444 dyna-ns domains are inside the head-skeleton set). Contact extraction ran 2026-04-25 23:38 to 2026-04-26 00:11 UTC; WHOIS 23:39 to 01:29 UTC.

The contacts are the closest thing to a name. The same 3,771 sites carry links@nextbacklinks.com and a WhatsApp number on the United Arab Emirates country code, in the footer of the homepage, not on the submission page. A second address, bloggar@gmail.com, is on 7,601 sites, and 3,827 still show hello@brand.com and +1234567890, which is what the template ships with before anyone edits it. WHOIS returned for 22,292 of the 37,154 domains and 18,735 of those had a registration date: 13,472 in 2026, 5,217 in 2025, and 8,232 in March 2026 alone, the month before I crawled them. Metaregistrar BV registered 7,239 and Dynadot, under four spellings of its own name, 9,895. Every registrant field was blank. So the operator is anonymous. It brands itself nextbacklinks, answers a UAE phone number, stores its assets on a German hosting company’s object storage, buys domains in bursts of thousands, and runs its own analytics server. The country is an inference from the phone number. Nothing in the registration data says where the people are.

contact_methods method_value and source_url; whois_records joined to cluster_members for cluster 507: 14,862 lookups errored, registrar counts and registered_on by month. The 4 May session log quotes 33,912 and 13,211 for 2026 and 2025; that join counted a domain once per cluster it sat in, so the table above is the number to use.

What the sites are for is written on them. Each homepage is a fake magazine on one Bootstrap 5 template served from a /static/blogapp/ folder, 17 stylesheets in the head, byte-identical from host to host. A data-font-scheme attribute on the html tag rotates through a short list of Google font pairings, 19 values in a 2,000-page sample, so neighbouring sites look different at a glance. The names are generated: slotvision876.de, spininsight360.us, notebuzzechotrail.com, turbogamesworld.com, with 12,315 on .com, 9,011 on .co.uk, 5,068 on .us, 3,912 on .de, 2,618 on .ca and 963 on .xyz. The articles are fresh. In 300 sampled homepages I counted 7,693 bylines dated April 2026, the month of the crawl, under names like Johnny Washington. Every site has a “Write For Us, Guest Post Submission” page and none of them quotes a price; the content profiler found 0 of 37,154 with one. The product is the outbound link. The 37,154 homepages carry 3,641,416 links to other hosts, 100 each, nine in ten marked dofollow, and 41 of them point at another farm site. They point at 9,874 hosts, and the anchor text says what was bought: 97 in every 100 anchors are gambling terms, in a dozen languages, and more than a third of those are for casinos outside a national self-exclusion register or licence, GamStop in Britain, Spelpaus in Sweden, ROFUS in Denmark, CRUKS in the Netherlands, AAMS in Italy. Every other vertical together, adult, crypto, loans, pharmacy, essay mills, is under half of one per cent. The gaming trade press is in the top ten: gameshub.com is linked from 14,232 of the sites as “online casino”, gamingamerica.com from 10,696 as “online casinos” and “best offshore casinos”, casinobeats.com from 10,004 as “no kyc casinos”, gamblinginsider.com from 9,756. But the biggest targets of all are British organisational domains: oncallfire.uk from 22,742 sites, regionalnetworks.org.uk from 17,186, taysideaviation.co.uk from 13,233, falconryforum.co.uk from 11,078, every one of them as “non gamstop casinos” or “casinos not on gamstop”, the phrase for casinos outside the UK’s self-exclusion scheme. Two of the biggest targets wear charity names and are not charities any more. autismspeaks.ca, on 9,409 of the sites, carries the name of a Canadian autism charity; by July 2025 it showed a hosting placeholder, by January 2026 it was “Best Online Casinos in Canada 2026”, and today it redirects to gamingamerica.com’s Canadian casino listings. The farm’s anchor on it is “best real money online casinos in canada”. roadpeace.org, on 10,428, is still the National Charity for Road Crash Victims, and the farm links to its homepage as “online casino”, “casino” and “non gamstop casinos”, all dofollow. Its WordPress uploads folder for February 2026 holds 888casino logos and a file named 30bet-non-GamStop-casinos, and /visit/harry-casino/, /visit/madcasino/ and /visit/wino-casino/ redirect out of the site. Somebody put casino pages on a charity’s website, and the farm is pointing at them. The charity blocks my requests, so that part is from the archive. So “private blog network” is a loose fit. It is a link-selling farm with all of its linking pointed outward. Its domains rank between 240,054 and 997,789 on Tranco, none in the top 100,000 and every one inside the million, which is the list everyone samples from when they want to describe the web.

analysis.db farm_links and farm_targets, built 29 October 2026 from the April homepage HTML of all 37,154 cluster-507 domains in pages.db (scripts/extract_farm_links.py); byline date strings from 300 random homepages; opportunities.rank and the TLD split of cluster_members; content_profiles.is_paid and price_amount; anchors from eight farm homepages in pages.db that link to either host; Wayback captures of roadpeace.org (24 July 2026 homepage, 27 and 28 February 2026 uploads and /visit/ redirects) and autismspeaks.ca (22 July 2025, 21 January 2026, 7 October 2026), and a direct fetch of both on 29 October 2026. The sector counts near the top of the page are a keyword classifier run on the same domains’ titles and descriptions.

Whether it works is a different question, and I could check part of it. The anchors are real searches. “Non gamstop casinos” is typed 33,100 times a month in Britain, “non gamstop casino” 14,800 times, “best non gamstop casinos” 12,100, and where Google sells an ad against the phrase the click costs 41 dollars. That is the money the farm is chasing. An independent backlink index, DataForSEO’s, sees the farm the way Google would. The top anchors it records for falconryforum.co.uk are “non gamstop casinos”, “non gamstop casino”, “casino not on gamstop” and “casinos not on gamstop”, and its monthly history shows the switch being thrown: falconryforum.co.uk had between 188 and 320 referring domains every month from January 2024 to February 2026, then 1,608 in March 2026, 2,283 in April and 3,892 by September. oncallfire.uk went from 243 in December 2025 to 1,830 in January and 12,926 in September. roadpeace.org, the charity, went from 1,532 in January 2026 to 11,176 in April. The farm is not the only supplier, either: of the thousand domains sending falconryforum.co.uk the most links, 676 are farm sites and the rest are other sites of the same kind, a bike shop, a pottery, a car hire, all first seen this spring. The intent is rank, in the one market where ranking pays. The outcome, as of 30 October 2026, is nothing. Google UK returns a single organic result for “non gamstop casinos”, a hacked page on india.com, and between one and four for the variants, and none of the farm’s eight biggest targets appears for any of them. Google has one page of oncallfire.uk left in its index, titled “Employer advice | Best Non GamStop Casinos”, one of falconryforum.co.uk, a handful of casino pages each for summerhall.co.uk and unbound.co.uk, and none at all for regionalnetworks.org.uk or bereavementcommission.org.uk. None of the eight ranks for a single tracked keyword, while gamstop.co.uk ranks for 488, and the one target that does rank, esports-news.co.uk, sits third and fourth for two of the phrases. The farm sites themselves are indexed, nine to 87 pages each, and rank for nothing. So the farm is pointed at a search result Google has emptied, and the pages it points at have been dropped. Whether the buyers know that, or whether the seller cares, I cannot tell from here.

DataForSEO, 30 October 2026, run from scripts/dfs_verify in backlink-opportunities, raw responses in results.jsonl, 1.45 dollars of API credit: Google Ads search volume for the 30 most common anchors, United Kingdom; live Google organic results, depth 100, for 14 phrases in the UK, Sweden, the Netherlands and Germany, each result matched against the 9,874 farm targets and 37,154 farm domains; Labs ranked keywords and rank overview, UK; site: searches for the eight biggest targets and ten random farm domains; backlinks summary, anchors, monthly history and referring domains. Controls for the empty rank data: gamstop.co.uk, esports-news.co.uk and casinomeister.com all return keywords through the same call. Two site: searches, roadpeace.org and taysideaviation.co.uk, returned a search-engine error and are not counted. A browser cross-check of the Google result page was stopped by Google’s bot check and not attempted again.

I had assumed the targets were a mix, some customers’ own sites and some innocent sites hacked to carry the links. The second part is mostly wrong. I took the 764 targets that a thousand or more farm sites point at, which carry 1,842,965 of the links, asked each registry when the name was created, asked the Internet Archive what it had looked like before, fetched the ones that would answer and followed the redirects. Of the 600 names with a registry date, 163 were created in 2025 or 2026. oncallfire.uk, the farm’s biggest target, was registered on 21 August 2025 on a name the archive had been capturing since 2018. regionalnetworks.org.uk, taysideaviation.co.uk and kysty.co.uk are the same, new registrations of old names. Another 172 keep an old creation date but now carry a casino title: fourpure.com, a brewery’s old address, is “Best Non GamStop Casinos UK”, cca-glasgow.com, registered in 1997, is “UK’s Best Online Casinos Not on GamStop”, and aboutcookies.org.uk compares casino sites. These were not stolen. They sit at the same registrars as the clear re-registrations, Namecheap, Regery and SafeNames, which is what a name bought at an expiry auction looks like; a hijacked one would still be at the registrar it was taken from. A further 79 old names simply redirect: autismspeaks.ca to gamingamerica.com, deutsche-seniorenliga.de to pokerstrategy.com, heartandlung.ca to gameplayspace.com, itself an old name turned casino site. Add them up and 434 of the 764 targets, 57 percent by count and 58 percent by links, are dead organisations’ names bought by someone else and put to work. Thirty-four are living publishers that rent out a gambling section, readwrite.com, videogamer.com, cardplayer.com, cointelegraph.com, indiatimes.com, and the farm links into those sections by path. Eleven were gambling publishers to begin with. Two hundred and eighty-four I cannot see, behind Cloudflare or without a registry record. That leaves the hacks. I can confirm one, RoadPeace, a charity that still runs its own site and has carried casino redirects inside its WordPress since February. That is one target in 764 and about six links in a thousand. One more, citytosea.org.uk, is live, clean, archived clean in September, and linked by 5,368 farm sites with the anchor “UK online casino”, and I do not know why. The farm is not the only one feeding these names, either. Of the 500 outside domains most often sending links to the same eight targets, 335 share a single page skeleton of their own, an older network of mostly .co.uk sites first seen in September 2025, four months before this farm began linking. Two factories, one set of customers. So the farm points mostly at graves. The names of charities, theatres, breweries and flying schools are the point: an old name with a history of links is worth more to a casino than a new one, and when the organisation dies the name comes up for sale. The farm is the second layer. The first is the market in expired domains.

Registry RDAP lookups for 660 of the 764 targets linked by 1,000 or more farm sites, 30 October 2026, at the registries directly (Nominet, Verisign, PIR, DENIC, SIDN, EURid, CIRA and others), a creation date for 600; Internet Archive latest capture for 382 and first-capture year where known; a live refetch of the 228 unwalled hosts counting casino anchors and where they sit; redirect destinations from the 29 October fetch; farm_links paths. Tables rdap, wb_latest, casino_links and target_provenance in dfs.db, scripts target_provenance.py, classify_provenance.py and casino_links_check.py in backlink-opportunities. A name bought at an expiry auction keeps its creation date, so the registry alone cannot tell a purchase from a theft; the registrar is the tell, and it is a tell, not a proof. The 139 Cloudflare-walled old names could hide more cases like RoadPeace; if every one did, hacks would be 18 percent, and nothing else points that way. A cloaking test, fetching the live targets as Googlebot, was refused by the tool’s permission check and not run. The second network: DataForSEO referring domains for the eight top targets, 5,000 each, 2.40 dollars of credit; the homepages of the 500 most shared outside senders fetched and their head skeleton hashed, 69 failed to answer.

I argued with the finding twice before I believed it. On 29 April the draft story said Google had failed to catch a network a hobbyist could detect in a weekend, and I asked “what makes me think that?” We had no evidence of years of operation; the registration dates said the opposite, and Tranco membership says nothing about Google rankings. On 4 May, after a different count came back, I asked for the code behind a claim that domains had been deindexed, and then “so are we even sure they’re a pbn?” What is measured: four infrastructure fingerprints on 37,154 domains and the same contact details on 3,771 of them. What is inferred: that the links are paid for. I never bought one. The verdict “pbn” in the database is a working label that means do not pitch these, not a finding about what Google did or didn’t do.

Session logs of 2026-04-29 01:50 UTC and 2026-05-04 19:50 to 19:56 UTC; cluster_profiles.verdict and implication for cluster 507.

The other networks, by the same method

Once a head hash counts as an ownership signal, every third-party ID on a page is one. A Google Analytics property, a Tag Manager container, an AdSense publisher ID and a hosted-font kit are each tied to one account, and one account on fifteen domains means one organisation pays for fifteen domains. The cluster build groups domains by the value of each signal and keeps every group of two or more: 483 clusters across 16 signal types in the 40,174. A profiler then read each cluster and gave it a verdict: 302 single operator, 101 shared template, 47 legitimate publisher, 29 pbn and 4 parking. The pbn verdict covers 37,469 domains, 37,154 of them the farm. The vendor IDs are the smallest clusters and the strongest ones.

Vendor ID sharedDomainsWho it turned out to be
GA4 G-7PC84FGF0J, UA-61683903-1 and GTM-KPMZ4JM, the same 15 sites on all three15Australian Community Media: canberratimes.com.au, newcastleherald.com.au, illawarramercury.com.au, bendigoadvertiser.com.au and 11 more regional papers. Nine different editors’ addresses at austcommunitymedia.com sit on up to 15 of the same sites.
GTM-NFD2QGD and GTM-TLFP4R, the same 11 sites on both11MediaNews Group, formerly Digital First Media: eastbaytimes.com, sun-sentinel.com, dailypress.com, chicoer.com, marinij.com, santacruzsentinel.com and five more.
GTM-57MCTQ810Wise Geeks: wisegeek.com, delightedcooking.com, thehealthboard.com, smartcapitalmind.com and six more Q&A sites. hello@wisegeeks.co is on the same 10.
GTM-TCW5HW9Minute Media’s FanSided network: apptrigger.com, culturess.com, dodgersway.com, nflspinzone.com, yanksgoyard.com and four more.
GA4 G-MHPD2T7WYT5joybuy.com, .co.uk, .de, .fr and .nl, one retailer’s country sites.
AdSense ca-pub-13326139855682295Five social-bookmarking spam sites: freewebmarks.com, ubookmarking.com, ybookmarking.com, social-bookmarkingsites.com, newsocialbookmarkingsite.com. One payee.
AdSense ca-pub-57342927112666675cellularinsider.com, citizenside.com, emojiguide.com, mobileburn.com, touristsecrets.com. One payee, no shared name anywhere on the pages.
GTM-T339N6S4Hachette UK: hachette.co.uk, hodder.co.uk, littlebrown.co.uk, orionbooks.co.uk, all on WordPress VIP.
GTM-TPR8JL3Urban One: bossip.com, globalgrind.com, tvone.tv.
UA-22379302-1 and GTM-NJ3CZ792LinkNow Media running linknow.com and linknowmedia.com, confirmed by an identical theme and plugin set.

clusters and cluster_profiles where signal is gtm, ga4, ga_ua or adsense; tracking_ids for the domain lists. In the 40,174 there are 27 Tag Manager clusters covering 117 domains, 9 Universal Analytics clusters covering 50, 7 GA4 clusters covering 30, 13 AdSense clusters covering 36 and one Facebook pixel on 5. The farm’s Matomo ID is the largest vendor-ID group of all and is excluded from the clusters by design.

The same signal reads two ways, and the verdict table is where the reading happens. The farm’s Matomo ID sits on 37,179 domains with one email address between them. Australian Community Media’s three Google IDs sit on 15 domains with nine named editors between them. One is a person pretending to be thousands of publishers; the other is a company that owns fifteen newspapers and runs one analytics account, which is what companies do. A shared ID proves an account, not an intent. The smaller farms show the other side of the same coin: three more head-skeleton templates on 107, 99 and 70 sites each, and 7 UK blog domains that redirect straight to casinorealm.nz, out of 57 external redirects in the whole list.

cluster_profiles for clusters 506, 484 and 487; redirects where is_external = 1, grouped by final_host.

A second crawl, 226,643 top-million homepages rendered in a browser for another project, has a vendor ID the guest-post list never saw: the kit ID in a hosted-font stylesheet URL. Adobe Fonts serves each account from use.typekit.net/<kit>.css, Monotype from fast.fonts.net, Hoefler from cloud.typography.com, and two sites loading the same kit were set up by the same person or vendor. In the 226,643 rendered sites, 7,749 load Adobe Fonts through 8,065 kit IDs, and 414 of those IDs appear on two or more sites. Monotype is on 103 sites with 101 IDs, two of them shared. Hoefler is on 97 with 95, two shared. The shared ones name their owner.

Font kitSitesWho
Typekit vrp2tco50Tourism boards on one destination-marketing vendor: neworleans.com, visitdenver.com, austintexas.org, cotswolds.com, discovernorthernireland.com, anchorage.net.
Typekit tsu4adm40Scripps local TV stations: 10news.com, abc15.com, denver7.com, fox13now.com.
Typekit ynn5aia33US school districts on one vendor: ocps.net, scps.k12.fl.us, pasadenaisd.org, mukilteoschools.org.
Typekit ocw2hya20South African betting sites on one platform: betfred.co.za, bettabets.co.za, boltbet.co.za, afribet.bet.
Typekit xod7oqu and mnx6iby, the same 18 sites on both18Vail Resorts: breckenridge.com, beavercreek.com, kirkwood.com, epicpass.com.
Typekit gpc7ewp17Car dealerships on one dealer-site vendor: autonationchevroletdoral.com, bergstromauto.com, covertford.com.
Typekit iup5rhb15Direct-to-consumer brands sharing an e-commerce agency: akubra.com.au, islandcreekoysters.com, homefieldapparel.com.
Monotype 56c28163-b5b8-47c5-99be-39d4ad6b964d5Kingston, Southwark, Trafford and Comhairle nan Eilean Siar councils and the East London NHS trust: one UK public-sector web vendor.
Hoefler 73464562babyruth.com and butterfinger.com, both Ferrara Candy.

stylesheet_links.href in design.db, distinct domains per kit, re-counted directly on 28 October; the totals per service from the 15 October signature probe (pipeline/checkpoints/signature_probe_results.json, font_kits). ynn5aia is from network.db font_kits, built 16 October from the rendered HTML of 226,630 homepages, which also adds Hoefler 6104956 on four NiSource gas utilities and a Monotype kit on four Epson country sites.

None of these is a link farm. They are the ordinary shape of the web’s ownership: a vendor builds forty television station sites or fifty tourism boards from one template and one font account, and the template votes forty or fifty times in any count of sites. The farm is the same mechanism at a scale where it stopped looking like a vendor and started looking like the web.

What it looks like from the top

Nobody reads these sites. That is not a figure of speech; it is the design. The articles are dated the day you arrive because they are written for a crawler that rewards freshness, not for a person who would notice that yesterday’s front page is gone. The hundred links on each homepage are for the customers who paid for them. The analytics account is for the operator, counting the crawlers. The sites do not link to each other because there is no reader to pass between them. Every part of the page faces a machine. The template is a magazine because a magazine is what a machine has been taught a trustworthy site looks like.

Put that against the top million. One list, meant to be a census of the web that matters, and one in every 28 entries on it is this operator. The rest of the list is not clean either, as the smaller farms, the parked pages and the placeholders in the same crawl show, but those are hundreds of domains each. This is tens of thousands, built in bursts of eight thousand a month, and it ranks because ranking is what it is for. What I can measure is the shape: one skeleton, one bucket, one tracker, one phone number, no readers and no cross-links. What I infer is the business: links sold by the piece to anyone with a budget and a page to point at, even when the page sits on a hacked charity. What I cannot see is who, or where, or how many more of these there are below the million, where no list reaches.

The dead internet, as it is usually told, is bots talking to bots on social platforms. This is the quieter version. It is a publishing industry with no audience, running at a scale that makes it 3.6% of the only map we have, and it was invisible until one number came back too neat to be true.

35,611 of 1,000,000 is 3.56%, or one domain in 28. The rest of this section is the tables above, read once more.

WordPress is on 37,933 of the 158,294 clean sites, 24.0%, which makes it the only CMS that matters in this corpus: Shopify is next at 2.7%. And WordPress themes are the clearest picture of template culture on the web, because the theme’s slug is in the URL of every stylesheet it loads (/wp-content/themes/<slug>/), so you can count them without guessing.

30,700
clean sites expose a theme slug in their stylesheet URLs.
17,513
distinct theme slugs across those sites.
87.6%
of slugs (15,345) exist on exactly one site.
1,707
slugs are on two to five sites: a studio’s house theme reused across its clients.
1,948
slugs look like child themes: a commercial parent, customised by someone.
12,004
sites use only slugs that appear nowhere else.

stylesheet_links hrefs matching /wp-content/themes/, clean corpus (theme_probe.py). On all 226,643 rendered sites the signals companion counted 18,571 slugs, 16,170 on one site.

The popular twenty

Theme slugClean sites
hello-elementor2,214
astra1,341
generatepress981
newspaper940
divi674
kadence612
hello-theme-child-master398
divi-child318
astra-child299
salient280
avada278
bb-theme262
newspaper-child232
flatsome208
hello-elementor-child204
blocksy200
avada-child-theme193
oceanwp190
bricks189
generatepress_child166

Sites per slug, clean corpus. Child slugs (divi-child, astra-child, hello-elementor-child and so on) are counted separately from their parents.

Two things are true at once. The top of the list is six commercial themes, and the first one, Hello Elementor, is a blank canvas for a page builder: it exists to get out of Elementor’s way. Astra, GeneratePress, Kadence and Blocksy are the lightweight starter themes that get customised in a settings panel; Newspaper (tagDiv), Divi, Salient, Avada and Flatsome are the ThemeForest heavyweights, each a product with its own page builder and its own typography defaults. Add the child themes to their parents and the six families (Hello, Astra, GeneratePress, Newspaper, Divi, Kadence) account for roughly 6,500 of the 30,700 sites with a slug.

And 87.6% of theme slugs exist on exactly one site. That is custom builds, renamed child themes, agency starter themes with the client’s name on them. The web has a very large bespoke cohort on WordPress. It just doesn’t look bespoke, because the page builder on top of the bespoke theme ships the same typography defaults as everyone else, and the block library underneath it ships the same CSS to every site. The bespoke part is the folder name.

What WordPress does to a site’s CSS

GroupSitesOne font familyClamp() anywhereAvg stylesheetsAvg CSS
No CMS detected107,55849.3%19.0%16.2583 KB
WordPress37,93334.4%42.8%34.1953 KB
Shopify4,27835.2%92.3%75.71,088 KB
Squarespace1,06939.1%25.5%27.81,363 KB
Wix6265.0%91.4%32.01,172 KB

domain_css_weight, domain_tags and blob_features joined per site, clean corpus (article_probe.py).

A WordPress site loads twice the stylesheets of a site with no CMS (34 against 16) and carries 370 KB more CSS. It is also twice as likely to have clamp() in its CSS, and that is not WordPress designers writing fluid type: the block-library stylesheet that WordPress core ships puts clamp() on 7,246 clean sites, and 5,058 of them have it nowhere else. Themes account for another 3,324 and plugins 1,775. A WordPress site uses a single font family 34.4% of the time, against 49.3% for the no-CMS group, because the theme sets one face, the page builder another, and the plugins inherit a third.

Who runs WordPress

SectorWordPress share
food_drink43.4%
news_media41.7%
nonprofit40.0%
parenting_family36.5%
religion33.9%
arts_design29.8%
all clean sites24.0%
retail14.0%
fashion_beauty12.3%
dating12.2%
social_media6.6%

Share of each sector’s clean sites with WordPress as the detected CMS; sectors from domain_classification (compute_v2.py, sector_table).

Restaurants, newspapers, churches and charities: the organisations that need a website and don’t have a developer. Retail and fashion sit at the bottom because they’re on Shopify. The newspaper theme market is visible in one row: Newspaper by tagDiv is the fourth most common theme in the corpus, and news sites load Google Fonts, use three or more families and run 14px body text more than any other sector, which is exactly what that theme does out of the box.

Who makes the themes

The theme’s own style.css carries an author line. In a 250,000-stylesheet sample scanned for the signals companion, 1,596 theme headers were readable. The authors: Elegant Themes (Divi) 102, the WordPress team 48, Elementor 45, WP-Script 26, tagDiv 22, Themesia 22, Underscores.me 20, Jegtheme 15, Beaver Builder 13, MVP Themes 13, ThemeNectar 12. Two of those are tells. “Underscores.me” as the author means a developer started from the blank starter theme and never changed the header: a bespoke build with a fingerprint. And WP-Script’s 26 are retrotube and ultimatube, adult tube themes, which is how the theme slug alone identified 271 adult sites the keyword classifier had missed. Those sites are in the excluded set, not in anything above.

Author counts from pipeline/checkpoints/signature_probe_results.json (blob_scan, 250,000 of 1,856,009 blobs, all rendered sites), as reported in viz/design_signals.html section 5.

Every year a top web trends article comes out. It’s mostly bullshit. But it points to something more interesting in design and perhaps everywhere. We want to know what to do and it’d be nice if it were easy. The design trend article is an extension of that, we don’t want to rebuild the wheel ourselves, we want some guy at Figma to give us some pointers. And that’s fine.

So, instead of giving you an article that says NeoBrutalist backgrounds with animated post-structuralist references is the direction the web is going and taking screengrabs from stuff I like on the Webby Awards, I’m going to try to define the standard web experience.

When I started this project, I was trying to build a flexible starter theme for my custom WordPress sites. I wanted to see what the numbers say about a starting point. The process of designing something then becomes a question of how and why we move away from the center. And the questions came at me in rapid fire.

It got technical, so apologies in advance. The jump links at the top are there for skimming.

Over the years, peeking behind the curtains on the sites I like, I found no real consensus on how to make a site. I did develop my own preferences, but by the numbers, what I found is that the adoption of systems, and how easy they are to adopt, is what determines what the web looks like.

A note on the numbers before I start. Everything below comes from 158,294 websites: the homepages in the Tranco top million that were in English, were actually sites, rendered in a real browser, and were not part of one link farm or an adult or gambling site. I call that the clean corpus. Where a table says “sites” it means a share of those 158,294, or of the subset that had the element in question (not every homepage has an H1). Where a table says “stylesheets” it means a share of the 1,734,591 unique CSS files those sites load, counted once each no matter how many sites share them. Those are different populations and they tell different stories, so I say which one every time. How the corpus was built is its own piece.

TRENDY

This is a bit of a reaction to the obnoxious annual listicle of design trends. There are questions here about how ideas diffuse across cultural artifacts in the age of the internet that I’m not qualified to answer. What I can say is that out of the items I can measure looking at the markup and stylesheets of these websites, it’s hard to say that these “trends” are significant. This is just some shit a designer thinks is cool and has seen a couple of times. And that’s fine. The headline is obnoxious and to me that’s less fine.

Now the list. Figma’s “web design trends 2026” names thirteen. I can only measure five of those with the data that I have with any degree of confidence. So here they are:

1Figma trend · measurable

3D

One homepage in 200 has anything 3D in it

three.js 0.4% of pages, WebGL contexts 0.1%, embedded 3D viewers 0.1%. About one homepage in 200 has anything three-dimensional in its code.

2Figma trend · measurable

Vibrant color palettes

One site in eight, and it is probably blue

Take each site’s dominant background color: 75.2% of 155,615 sites are neutral, 12.8% are tinted, and 12.0% have a saturated mid-tone background. Nearly every site (95.9%) has at least one saturated color somewhere, because a blue link counts, and of those accent colors blue, azure and cyan are 58.5%. The vibrant-palette web is one site in eight, and even then it’s probably blue.

3Figma trend · measurable

Bold typography

Bold is the plurality, not a trend

44.8% of H1s are 700 or 800*; 25.8% are 56px or larger*; 10.5% are uppercase*. The light headline (400) is on 21.8%*. Bold is the plurality, not a trend.

4Figma trend · measurable

Dark mode

One site in eight is dark by default; a toggle is on one in twenty

12.2% of sites render a dark body by default* (gaming 45%*, entertainment 32%). A dark-mode toggle is on 5.4% of pages, a color-scheme meta tag on 1.2%, and prefers-color-scheme appears in 1.4% of stylesheets. Dark-by-default sites are more than twice as common as sites with a toggle.

5Figma trend · measurable

Motion and micro-interactions

A quarter of stylesheets animate something; the libraries are on 2 to 3% of pages

@keyframes in 23.4% of sheets, animate.css on 8,286 clean sites, scroll-animation libraries on about 3% of pages, GSAP 2%, Lottie 1.2%. prefers-reduced-motion in 5.0% of sheets. Transitions on hover are presumably near-universal but I didn’t count them.

On motion: canvas, WebGL, 3D and animation libraries

From the corrected trends probe on 15,001 clean homepages:

Library or technique (HTML scan) Pages Share
cookie or consent banner 4,490 29.9%
Swiper or Slick carousel 2,783 18.6%
mega menu 1,843 12.3%
variable font axis range 879 5.9%
dark-mode toggle 817 5.4%
AOS or ScrollReveal (scroll animation) 432 2.9%
GSAP 305 2.0%
Lottie 185 1.2%
three.js 56 0.4%
WebGL canvas context 19 0.1%
model-viewer, Sketchfab, Matterport, Spline 12 0.1%
Framer Motion 8 0.1%

15,001 clean homepages, first 400 KB of rendered HTML, tightened patterns (trends_probe2.py).

The most common “motion” on the web is a carousel. Carousels are relatively easy to build by hand and they are easy to make as plugins because of a handful of great JavaScript libraries. What I see in the numbers is that design friction is tied to what we do or don’t see on the web. People do hard things less.

Scroll-triggered animation libraries, GSAP and Lottie together are on about 6% of homepages. Three.js is on four in a thousand. Admittedly these are harder libraries to master and require more assets.

The cookie banner is at 30%, which is six times more common than every animation library combined. This is brought on by necessity, everyone is scared to get sued, people make plugins for that.

Frameworks, themes, stacks, templates

Before we hop into the pixels of it all, I want to put it out there: I like the internet. I like its scaffolding. I’m just trying to wrap my head around the numbers here, and this is the part that explains why things are the way they are.

WordPress comes with most hosting setups as a one-click install. Bootstrap and Tailwind used to be drag and drop, or just pasting a CDN link into your HTML. Now you can spin up either in the terminal with a single line you copy to your clipboard with the click of a button.

IT’S EASY TO START. That’s the point. And then you’re locked into all of their choices. Which is fine…but not really.

Stack Clean sites Share of 158,294
Bootstrap (winning CSS framework) 79,482 50.2%
No framework detected 55,257 34.9%
Tailwind 23,084 14.6%
Sites carrying both Tailwind and Bootstrap tags 73,656 46.5%
WordPress 37,933 24.0%
No CMS detected 107,558 67.9%
Shopify 4,278 2.7%
Drupal 3,966 2.5%
Squarespace 1,069 0.7%
Webflow 752 0.5%
Wix 626 0.4%

domain_tags, one winning framework and CMS per site.

Half the clean web is on Bootstrap. Wix, Squarespace and Webflow together are 1.5% of the Tranco top million, which tells you the builders you see advertised are not where the popular web is built; it’s built on WordPress and on hand-rolled Bootstrap. Both of which are FREE. The 73,656 sites tagged with both Tailwind and Bootstrap are a plugin or a component library bringing the second framework along for the ride.

WordPress is a whole other rabbit hole, but as the most popular CMS very popular themes and plugins also shift the definition of the standar web look in the data. I get into what the Tranco database project has to say about WordPress here.

So what does the standard web look like? Mostly it looks like Bootstrap or Tailwind, Google’s font library, and maybe a few of the popular JavaScript libraries. A lot of what a client thinks of as design is inherited choices from open-source projects like these. In the sense that these frameworks make a lot of the choices, we can think of the web as mostly undesigned.

76
stylesheets loaded by the average Shopify store.
5%
of Wix sites use a single font family. Wix sets a different font on every widget.
13,071
sites carry Bootstrap’s body text color #212529; 5,802 carry its primary blue #0d6efd.
16,742
sites carry the browser’s unstyled link blue #0000ee. One site in ten has a link nobody styled.
45,442
sites have a body background, body text and H1 color combination that appears nowhere else.
23
color combinations cover 28,966 sites: white, black, #333, #212529 and their mixes. Framework defaults.

domain_css_weight, computed_styles and the palette table, clean corpus. The color “system” is each site’s body background, body text and H1 color taken together.

It gets messy

I did a quick count of how many stylesheets a site actually has going at once. On the homepages of these 158,294 sites the browser is running 22.4 stylesheets on average, 703 KB of CSS. That’s the average; it’s a lot worse on the platforms.

Platform Sites Stylesheets per site CSS per site
No CMS detected 106,102 16.2 583 KB
WordPress 37,922 34.1 953 KB
Shopify 4,278 75.7 1,088 KB
Wix 626 32.0 1,172 KB
Squarespace 1,069 27.8 1,363 KB

domain_css_weight joined to domain_tags, clean corpus; external and inline sheets together.

Twenty-two stylesheets is a lot of surface area for rules to conflict, and it’s exactly how you get 17 breakpoints on one page. Not all of them are min, not all of them are max; the default for any one stylesheet can be the opposite of the one next to it, and you can have conflicting rules inside a single sheet. A Shopify store loads 76 of them. My own site loads 11, five files and six inline blocks, and I thought that was a lot.

Modern CSS in general

I empathize with the design pros evangelizing new ways to make things responsive or faster. In the numbers, though, that change is coming slowly, if it’s coming at all.

flexbox53.4%all browsers since 2015
gap26.3%all browsers since 2021 (for flexbox)
grid15.5%all browsers since 2017
logical properties11.1%all browsers since 2021
aspect-ratio8.8%all browsers since 2021
:has()7.3%all browsers since 2023
@layer1.9%all browsers since 2022
color-mix()1.5%all browsers since 2023
@container1.3%all browsers since 2023
oklch()0.5%all browsers since 2023
subgrid0.5%all browsers since 2023
view transitions0.1%all browsers since 2025

Share of stylesheets using each feature, from a 60,000-sheet sample scanned for the signals companion. The year is when the feature became available in every major browser (Chrome, Firefox and Safari), the point after which a site could use it without a fallback.

px in sheets75.9%
rem in sheets23.0%
sites where rem beats px13.3%
sheets with custom properties30%

Units from the 1,734,591 unique sheets the clean sites load; rem beats px on 20,859 of 156,880 sites with CSS; custom properties from the 60,000-sheet sample, median seven variables per sheet that has any.

What a standard looks like

The shame of dropping into a rabbit hole, sometimes, is that you come out with very little to show for it. So for me I wanted to come out with a standard, a baseline: a set of CSS rules that make the averagiest site there ever was. The black t-shirt and jeans of the mid-to-late aughts.

So, according to the numbers, it goes something like this:

  • Body. 16px Open Sans on a 24px line-height, black on white. Everything below assumes that 16px base, which is also what the numbers say.
  • General headings. Roboto, weight 700, line-height 1.2, sentence case. Nothing uppercase.
  • Letter-spacing. None. Normal letter-spacing on every element, nothing tightened, nothing opened up.
  • H1. 48px, or 3rem, on a 1.2 line-height.
  • H2 and H3. In the numbers these two share a font size: both land on 24px most often. So I bumped the H2 up to 32px (2rem) and kept the H3 at 24px (1.5rem), assuming a body, or base, font size of 16px. That gives 48 / 32 / 24 / 16, a straight 1.5 step.
  • Button. Square corners, no border, no shadow, 16px, weight 400, sentence case.
  • Breakpoint. One, at 768px, min-width.
THE STANDARD PAGEModal value of every property, clean corpus

This is what most of the web looks like

A second-level heading at thirty-two pixels

Sixteen pixel Open Sans on a twenty-four pixel line-height, weight four hundred, no letter-spacing, sentence case. Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

Read more

A third-level heading at twenty-four pixels

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

H1 Roboto · 48px · 700 · line-height 1.2
H2 Roboto · 32px · 700 · line-height 1.2
H3 Roboto · 24px · 700 · line-height 1.25
Body Open Sans · 16px · 400 · line-height 24px
Button Arial · 16px · 400 · radius 0 · no border · no shadow · sentence case · padding 4px 12px · background #efefef, the browser’s own button grey
Colors body background #ffffff (73,744 sites) · body text #000000 (46,090) · H1 shown in black, though the single most common computed H1 color is white (36,878), because so many H1s sit on a dark hero banner

Each property is the most common value on its own (top font, most common pixel size, top weight, median line-height, most common color). Nobody’s site is exactly this; a lot of sites are within a step of it.

The median site is a little quieter than this: its H2 is 1.79 times the paragraph (about 29px) and its H3 1.33 times (about 21px).

The aggregated numbers are not a design choice that any single person made. It’s the intersection of the defaults, and I think that makes it the perfect starting point for a website. It’s serviceable. If you copied and pasted this CSS into your stylesheet, no one would bat an eye at the site. And as a baseline, it lets you change single elements and see the effect they have on the look and feel of the site.

body {
  font: 400 16px/24px "Open Sans", sans-serif;
  color: #000;
  background: #fff;
}
h1, h2, h3 {
  font-family: Roboto, sans-serif;
  font-weight: 700;
  line-height: 1.2;
  letter-spacing: normal;
  text-transform: none;
}
h1 { font-size: 48px; }   /* 3rem   */
h2 { font-size: 32px; }   /* 2rem   */
h3 { font-size: 24px; }   /* 1.5rem */
button, .btn {
  font: 400 16px/1.5 Arial, sans-serif;
  padding: 4px 12px;
  background: #efefef;
  color: #000;
  border: 0;
  border-radius: 0;
  box-shadow: none;
}
@media (min-width: 768px) {
  /* the one breakpoint everyone agrees on */
}

The standard stylesheet: every value is the most common one in the clean corpus except the H2 size, which is the second most common so that it differs from the H3.

Putting it all together by industry

As this project developed I kept adding more passes over the markup and stylesheets. At some point I realized that I was curious about what the sites actually were, and how all of my design questions would change by just adding an industry label.

At first I ran a script against each site’s title and meta description, scoring them against keyword lists, which left over half of the sites with a “general” industry label. Then we did two model runs through Ollama’s hosted API, first Gemma, then DeepSeek. That gave us 28 industry slices.

And because of that, we were able to extrapolate styling patterns across technology, news, gaming, fashion and beauty, and food and drink.

Technology

32,032 sites

Technology (32,032 sites, a fifth of the corpus) is the baseline the others deviate from: Inter is its H1 (7.8%), half of tech sites use one font, 5.8% uppercase their H1, 14.2% run dark, the Tailwind share is 16.8%, and pages are short (6,163px) and light (579 KB of CSS). What tech has instead is Next.js: 3,025 of the corpus’s 12,000 Next.js sites, though finance uses it at a higher rate (1,287 of 8,512 sites).

Ship the next thing faster

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation.

Read more
H1 Inter · 48px · 700
Body Inter · 16px · 400 · line-height 24px
Button Arial · 16px · 400 · radius 0px
dark body14.2%
bold H149.0%
uppercase H15.8%
uppercase button6.5%
square button44.5%
one font50.6%
Google Fonts42.5%
breakpoints / site17.5

Inter leads both H1 and body; 23.4% of tech H1s carry negative letter-spacing, the highest tightening rate of any sector, so the H1 here has letter-spacing of -0.02em.

News and media

7,651 sites

News and media (7,651 sites). The most homogeneous sector on the web. Roboto is the H1 on 18.7% of news sites, triple its share anywhere else, with Open Sans at 10.2%; 61.5% load Google Fonts, the highest of any sector; 41.7% run WordPress, the second highest; and news has the smallest body text, with 25.6% of homepages at 14px against 13% in technology. Homepages average 8,679px tall, versus 6,163px in tech. News sites are second only to food in fonts used (27.5% use three or more) because the masthead, the headlines and the body are set separately. This is the newspaper theme market (Newspaper by tagDiv is the fourth most common theme in the whole corpus) doing exactly what it was sold to do.

Council passes budget after midnight vote

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation.

Read more
H1 Roboto · 28px · 700
Body Open Sans · 16px · 400 · line-height 24px
Button Arial · 16px · 400 · radius 0px
dark body4.7%
bold H154.1%
uppercase H17.1%
uppercase button11.8%
square button57.8%
one font32.5%
Google Fonts61.5%
breakpoints / site24.1

The smallest headlines of the five: the H1 mode is 28px against 48px in tech, and a quarter of news homepages set body text at 14px.

Gaming

7,456 sites

Gaming (7,456 sites). The dark web. 45.1% of gaming homepages render a dark body, against 12.2% overall and 2.7% in fashion. Gaming also has the fewest breakpoints per site (16.2 distinct widths on average, lowest of all sectors), one of the lowest Bootstrap shares (41.1%, with entertainment at 42.8%), and a 56.7% bold H1 rate, the highest. Whatever they’re built on, gaming sites carry fewer layers of CSS, and they look it: dark, heavy headings, fewer breakpoints.

Patch 4.2 is live: new maps, new ranked season

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation.

Read more
H1 Open Sans · 32px · 700
Body Times · 16px · 400 · line-height normal
Button Arial · 16px · 400 · radius 0px
dark body45.1%
bold H156.7%
uppercase H113.7%
uppercase button9.9%
square button44.2%
one font51.7%
Google Fonts45.4%
breakpoints / site16.2

Shown dark: 45.1% of gaming homepages render a dark body, the sector’s signature, though white is still the single most common background. Most common body font is Times, the browser default, on 503 sites; Roboto is second on 445.

Fashion and beauty

3,713 sites

Fashion and beauty (3,713 sites). The sector that looks most like a template, in a good way. 70.9% of fashion buttons are square, the highest of any sector. Only 31.1% of fashion H1s are bold, the lowest anywhere; the light-weight fashion headline is real. Fashion runs Bootstrap at 66.8%, the top of the table, averages the most breakpoints per site (30.3), and has the highest clamp() share (52.2%) and container-query share (41.1%), none of which is from designers writing modern CSS. It’s the Shopify and WooCommerce stack: e-commerce platforms ship clamp and container queries in their checkout and theme CSS, so the most conservative-looking sector on the web is the one with the most “modern” code in it.

The new season, in twelve silhouettes

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation.

Read more
H1 Montserrat (tied with Poppins) · 40px · 400
Body Poppins · 16px · 400 · line-height 24px
Button Arial · 16px · 400 · radius 0px
dark body2.7%
bold H131.1%
uppercase H117.2%
uppercase button15.4%
square button70.9%
one font38.4%
Google Fonts57.5%
breakpoints / site30.3

Montserrat and Poppins tie for the H1 at 127 sites each; the H1 weight mode is 400, the only featured sector where the regular-weight headline wins.

Food and drink

3,832 sites

Food and drink (3,832 sites). The loud one. Food has the highest uppercase rate for both H1s (22.6%) and buttons (21.4%), double the web, the largest body text (34.7% of food homepages run body text at 17px or more, against 13.6% in tech), the lowest single-font rate (only 26% use one family; 29.2% use three or more), the highest Tailwind share (23.8%) and the highest WordPress share (43.4%). Restaurants and food blogs are where display type, big text and multiple fonts actually live.

Brunch, served all day

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation.

Read more
H1 -apple-system · 48px · 700
Body -apple-system · 16px · 400 · line-height 24px
Button Arial · 16px · 400 · radius 0px
dark body5.1%
bold H147.2%
uppercase H122.6%
uppercase button21.4%
square button61.4%
one font26.0%
Google Fonts44.3%
breakpoints / site25.1

Both H1 and body most often resolve to the system stack (-apple-system), so this card is set in your operating system’s font, as 569 food sites are.

Each card is set in the sector’s most common H1 font, size and weight, its most common body font, size and line-height, and its most common button font, size, weight and radius (computed_styles, clean corpus). Headline copy is invented. The small numbers are the sector’s share of sites with that trait. Scroll sideways.

FONTS

I’ve never made a font. I’ve manipulated them in Photoshop and Illustrator, but making one from scratch is a real talent. So when Google or Adobe just collect them by the dozens and make them easy to include in your site, it’s a no-brainer. So, when talking about our English-language Tranco list that doesn’t include porn or gambling, the most popular fonts are five Google fonts and the operating system’s default font.

First font family on the body Sites
Open Sans 11,825
Roboto 9,995
Inter 9,063
-apple-system (the system stack) 7,986
Times (unstyled browser default) 7,812
Arial 6,467
Poppins 5,770
Montserrat 5,071
Helvetica Neue 4,639
Lato 4,612

Computed body font on 157,837 clean sites that had a body style captured. First family in the resolved font stack.

#1Google Fonts

Open Sans

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

11,825 sites set this on the body
#2Google Fonts

Roboto

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

9,995 sites set this on the body
#3Google Fonts

Inter

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

9,063 sites set this on the body
#4System stack: whatever the OS has

-apple-system

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

7,986 sites set this on the body
#5Browser default, never set

Times

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

7,812 sites set this on the body
#6System font

Arial

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

6,467 sites set this on the body
#7Google Fonts

Poppins

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

5,770 sites set this on the body
#8Google Fonts

Montserrat

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

5,071 sites set this on the body
#9System font (Apple)

Helvetica Neue

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

4,639 sites set this on the body
#10Google Fonts

Lato

Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

4,612 sites set this on the body

The ten most common body font families, each card set in the family itself (Google Fonts loaded for this page; system faces render with whatever your machine has). Scroll sideways.

On headings the order shuffles but the cast doesn’t change: Roboto, Open Sans, Poppins, Inter, Montserrat, in that order, lead the H1 list, followed by the system stack and Arial. Nothing else reaches 3,000 sites.

Free and easy wins the web. So much so that 7,812 sites never set a body font at all. Someone shipped a homepage with the browser default. Awesome, why not?

Where the fonts come from:

Font source Sites Share of 158,294
Loads a Google Fonts stylesheet 75,495 47.7%
Loads an Adobe Fonts (Typekit) kit 8,119 5.1%
Loads Bunny Fonts 1,048 0.7%
Body set in a system or generic face (Arial, Helvetica, system-ui, Times, serif…) 45,932 29.1% of 157,837 bodies
Every captured role in a system face and no Google Fonts link 20,426 12.9% of 157,838

stylesheet_links hrefs and computed_styles, clean corpus.

Just under half the web loads Google Fonts. About one site in eight runs entirely on fonts the visitor already has, and those are mostly the Bootstrap and Tailwind default stacks (-apple-system, BlinkMacSystemFont, “Segoe UI”…), not a considered choice of Helvetica.

How many fonts does a site use?

Across seven roles (body, paragraph, H1, H2, H3, button, nav), 44.7% of sites use exactly one family. Another 37.2% use two. Only 3.6% use four or more.

Distinct font families across 7 roles Sites Share
1 70,545 44.7%
2 58,775 37.2%
3 22,836 14.5%
4 or more 5,682 3.6%

157,838 clean sites with computed styles.

Most common pairings

#1H1 / body pairing

Roboto

Verdana. Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

824 sites pair these
#2H1 / body pairing

Montserrat

Open Sans. Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

490 sites pair these
#3H1 / body pairing

Poppins

Open Sans. Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

260 sites pair these
#4H1 / body pairing

Roboto

Open Sans. Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

253 sites pair these
#5H1 / body pairing

Linux Libertine

sans-serif. Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

246 sites pair these
#6H1 / body pairing

Poppins

-apple-system. Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

243 sites pair these
#7H1 / body pairing

Montserrat

Roboto. Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

239 sites pair these
#8H1 / body pairing

Poppins

Roboto. Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

223 sites pair these
#9H1 / body pairing

Raleway

Open Sans. Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

216 sites pair these
#10H1 / body pairing

Roboto Condensed

Roboto. Lorem ipsum dolor sit amet, consectetur adipiscing elit. Sed do eiusmod tempor incididunt ut labore et dolore magna aliqua, quis nostrud exercitation ullamco.

211 sites pair these

The ten most common H1-over-body pairings on the clean corpus where the two families differ. Heading set in the H1 family, paragraph in the body family. Linux Libertine over a generic sans is MediaWiki’s default skin (243 of the 246 sites with that pairing are MediaWiki installs or Wikipedia mirrors); it renders here only if you have the font.

Roboto over Verdana

The most common H1-over-body pairing is Roboto over Verdana, on 824 sites. It is not a choice anyone made: 97% of those sites run tagDiv’s Newspaper WordPress theme, which ships with Verdana body text and Roboto headings, and two thirds of Newspaper sites never change them. The theme is well designed and only costs 60 dollars. Envato claims the theme has been sold 160K times.

The Newspaper theme demo: a news magazine homepage with Roboto headlines over Verdana body text
tagDiv’s Newspaper theme, the one behind 97% of Roboto-over-Verdana sites. It costs $60 on Envato.

You can see this pattern repeat itself. Third place, Linux Libertine over sans-serif, is MediaWiki’s Vector skin. Like WordPress, you will find MediaWiki among the one-click install options on most hosting services (it used to be called Installatron on cPanel). So even the idea of pairs as a way to distinguish yourself has, for the most part, been left to other designers up the food chain. Long gone are the days of people changing the CSS on their MySpace.

Headings and body, role by role

The same measurements for each role, so you can compare them. The H2 and H3 both land on 24px most often; what separates them on the median site is weight and the ratio to the paragraph, not a step on a scale.

Role Sites Top font Most common size Median size ÷ paragraph Top weight Median line-height Uppercase Tightened letter-spacing
H1 113,043 Roboto 48px 2.43 700 1.20 10.5% 15.4%
H2 127,888 Roboto 24px 1.79 700 1.20 12.7% 12.0%
H3 111,776 Roboto 24px 1.33 700 1.25 11.4% 9.2%
body 157,837 Open Sans 16px 1.00 400 1.50 0.0% 1.2%

computed_styles, clean corpus, one element per role per site. The tabs below carry the full size bands, fonts, weights, line-height and letter-spacing for each role.

H1113,043 sites
SizeSitesShare
under 16px
4,743
4.2%
16 to 19px
6,051
5.4%
20 to 23px
5,448
4.8%
24 to 31px
15,843
14.0%
32 to 39px
19,774
17.5%
40 to 55px
32,018
28.3%
56px and up
29,166
25.8%

The single most common values are 48px (7,888 sites), 40px (7,600), 32px (7,230), 36px (5,678) and 60px (5,319). So if you want one number, 40 to 48px is the center of gravity, but half the web is outside that band. The 16,242 sites with an H1 under 24px are using the tag as a label, not a headline. That’s a real pattern and it’s not in any trends article.

A better way to think about it is the ratio of the H1 to the paragraph text, because that’s what you actually perceive.

H1 size divided by paragraph size Sites Share
under 1 (H1 smaller than body copy) 5,191 4.8%
1 to 1.5 12,667 11.6%
1.5 to 2 14,522 13.3%
2 to 2.5 23,117 21.2%
2.5 to 3 17,337 15.9%
3 to 4 21,675 19.9%
4 and up 14,284 13.1%

108,793 clean sites with both an H1 and a paragraph in pixels. Median 2.43, middle half between 1.79 and 3.21.

The median H1 is 2.4 times the paragraph. If your body is 16px that’s about 39px, which matches the size table. The spread is wide enough that “a modular scale of 1.25” or whatever is clearly not what the web does. One site in eight runs the H1 at four times the body or more. One in twenty runs it smaller than the body.

Top families

Roboto 6,392
Open Sans 5,839
Poppins 5,434
Inter 5,356
Montserrat 5,144
-apple-system 3,188

Weights

700 45,054 39.9%
400 24,591 21.8%
500 14,139 12.5%
600 12,613 11.2%
800 5,576 4.9%

Weight is the one place headings do agree. The H1 is bold (700) on 45,054 sites and regular (400) on 24,591; 500, 600 and 800 split the rest. The light editorial headline, 300 or below, is too rare to make the top five. Bold or heavier (700 or 800) is 44.8% of H1s. The uppercase H1 is on 10.5% of sites, the uppercase H2 on 12.7%.

The rest

Most common size 48px (7,888 sites)
Most common size / weight / line-height 32px · 700 · normal (1,972 sites)
Median line-height ratio 1.2 (default "normal" on 8.8%)
Letter-spacing normal 75.4% · tightened 15.4% · opened 9.2%
Uppercase 10.5%
Median size ÷ paragraph 2.43 (108,793 sites with both)

Letter-spacing on the H1 is mostly untouched.

H1 letter-spacing Sites Share
normal or 0 85,107 75.3%
negative (tightened) 17,369 15.4%
positive (opened up) 10,347 9.2%

113,043 clean H1s. The most common non-zero values are -1px (2,013 sites) and 1px (1,675); as a proportion of the font size the most common is -0.02em.

Three quarters of H1s have the default letter-spacing. The tightly spaced display headline that every SaaS template ships, -0.02em or so, is on one site in seven. If you learned “always tighten large type” as a rule, you learned it from the 15%.

H2127,888 sites

The H2 is where the hierarchy starts to blur. The most common size is 24px (11,372 sites), then 32, 30, 36 and 40. A quarter of H2s sit in the 24 to 31px band, but 40px and up is another 27%, so plenty of sites run an H2 at H1 size. The median H2 is 1.79 times the paragraph, about 29px on a 16px page. Bold (700) on 37.7%, regular on 21%. Uppercase on 12.7%, which is more than the H1.

SizeSitesShare
under 16px
7,129
5.6%
16 to 19px
13,666
10.7%
20 to 23px
14,095
11.0%
24 to 31px
33,465
26.2%
32 to 39px
25,278
19.8%
40 to 55px
25,449
19.9%
56px and up
8,806
6.9%

Top families

Roboto 6,666
Open Sans 6,278
Poppins 6,272
Montserrat 6,186
Inter 6,076
-apple-system 3,720

Weights

700 48,185 37.7%
400 26,874 21.0%
500 18,159 14.2%
600 18,091 14.1%
300 5,870 4.6%

The rest

Most common size 24px (11,372 sites)
Most common size / weight / line-height 24px · 700 · normal (1,748 sites)
Median line-height ratio 1.2 (default "normal" on 7.4%)
Letter-spacing normal 77.9% · tightened 12.1% · opened 10.1%
Uppercase 12.7%
Median size ÷ paragraph 1.79 (123,651 sites with both)
H3111,776 sites

The H3 is the role with the least agreement, and it leans small. 24px is the most common value (13,141 sites), but 20, 18 and 16 follow, and 16 to 19px is the band on 26% of sites. A quarter of H3s are body size and separated from the paragraph by weight alone. The median H3 is 1.33 times the paragraph, about 21px on a 16px page. Bold (700) on 36.2%, with 600 (18.6%) closer behind than on the H2. This is why the standard page bumps the H2 to 32px: in the raw numbers H2 and H3 share a mode, and something has to give.

SizeSitesShare
under 16px
12,146
10.9%
16 to 19px
29,038
26.0%
20 to 23px
20,417
18.3%
24 to 31px
31,352
28.0%
32 to 39px
10,507
9.4%
40 to 55px
6,511
5.8%
56px and up
1,805
1.6%

Top families

Roboto 6,512
Inter 6,145
Open Sans 5,773
Poppins 5,723
Montserrat 5,374
Lato 2,916

Weights

700 40,411 36.2%
400 22,750 20.4%
600 20,788 18.6%
500 17,611 15.8%
300 3,798 3.4%

The rest

Most common size 24px (13,141 sites)
Most common size / weight / line-height 28px · 500 · 33.6px (1,305 sites)
Median line-height ratio 1.25 (default "normal" on 7.0%)
Letter-spacing normal 81.0% · tightened 9.2% · opened 9.7%
Uppercase 11.4%
Median size ÷ paragraph 1.33 (108,748 sites with both)
Body157,837 sites

Body text is the one role with a settled answer: 16px on 84,214 sites (53%), weight 400 on 92%, line-height 1.5. A third of sites still run under 16px, almost all of them at 14 or 15px, and only 3% go 20px or bigger. If every other number in this essay is a distribution, this one is a point.

SizeSitesShare
under 16px
50,789
32.2%
16 to 19px
101,953
64.6%
20 to 23px
4,444
2.8%
other (24px and up)
651
0.4%

Body text is the one place the web has fully converged, and the number is boring: 16px, weight 400, 24px line-height. That exact triple is on 37,125 sites. The next most common is 16px with the browser’s default normal line-height on 17,473, and then 14px on 20px (7,561). Paragraph sizes: 16px on 38,402, 14px on 21,326, 18px on 16,250, 20px on 10,086. Eighteen and twenty are more common on food and nonprofit sites (more on that below), but 16 is the web.

Top families

Open Sans 11,825
Roboto 9,993
Inter 9,062
-apple-system 7,986
Times 7,812
Arial 6,467

Weights

400 145,245 92.0%
300 6,762 4.3%
500 4,393 2.8%
600 342 0.2%
200 328 0.2%

The rest

Most common size 16px (84,214 sites)
Most common size / weight / line-height 16px · 400 · 24px (37,125 sites)
Median line-height ratio 1.5 (default "normal" on 18.4%)
Letter-spacing normal 95.4% · tightened 1.2% · opened 3.3%
Uppercase 0.0%

What unit is that 16px set in? Of the 831,292 unique stylesheets that declare a font-size at all, 75.2% use px somewhere, 29.5% em, 29.0% rem, 2.2% clamp() and 2.1% vw. Per site, taking whichever unit a site’s font-size declarations use most: px on 101,651 sites (65.2% of 155,878), rem on 34,856 (22.4%), em on 14,357 (9.2%). A fluid unit (clamp or vw) is the dominant font-size unit on 585 sites. The rem-based type scale that every modern tutorial starts with is on about one site in five.

moved from Trendy

Fluid type and clamp()

The question of whether anyone uses clamp() is born out of the slight discomfort of having to learn a new style. The answer to the question is much more interesting. 2% of the 1.7 million stylesheets powering the 158K corpus use clamp(). 27% of websites have clamp() on their site. It’s coming from larger organisations’ assets: YouTube’s embedded player, the WordPress block library, app build bundles. Six files explain most of it.

Sheets say no, sites say yes

clamp() and container queriesCountShare
clamp() in unique stylesheets34,5372.0%
clamp() on sites42,85227.3%
container queries in unique stylesheets10,4410.6%
container queries on sites18,25111.6%

Sheets: 1,734,591 unique stylesheets by content hash. Sites: 156,880 clean sites with any CSS captured, 1,414 of the 158,294 had none. A feature that lives in one popular shared sheet scores low on the first and high on the second.

Who put it there

Of the 42,852 sites with clamp() somewhere, most got it from one of six files they never opened. One row per file, host dropped and versions merged, so WordPress’s block library counts once across every site that serves its own copy.

Stylesheet with clamp() in itSitesShare of clamp() sites
inline <style> blocks the site, or its builder9,87423.0%
YouTube embed, player stylesheet video embed7,80718.2%
WordPress block library WordPress core6,48815.1%
Shopify portable-wallets Shopify checkout3,6148.4%
Shopify checkout-web Shopify checkout2,7516.4%
Vimeo embed, player stylesheet video embed2,2515.3%
Next.js build CSS app bundle1,3703.2%
Shopify checkout button Shopify checkout7871.8%
LiteSpeed cache bundle WP cache7031.6%
Kadence theme global.css WP theme5501.3%

42,852 clean sites with clamp() in any stylesheet; a site can appear in several rows. After row ten nothing reaches more than 1.1% of clamp() sites; the tail is WordPress plugins, app bundles and generic file names. The two YouTube player versions alone are the only clamp() on 5,587 sites, which makes an embedded video the single biggest reason a site “uses clamp()”. The inline row is not all hand-written: Wix and most page builders inline their platform CSS, and the Next.js and LiteSpeed rows are bundles, so the share a site’s own designer wrote is somewhere under the 12.9% in row 3 of the ladder below.

The platform talks, the headline doesn’t

Split by platform, the first two numbers swing with whoever ships the CSS. The third one, the share of homepages whose H1 computed to a fractional pixel size like 43.2px, barely moves.

clamp() anywhere clamp() on font-size H1 at a fractional px
No CMS detected75,368 H1s
19.0%
8.0%
13.5%
WordPress37,922 sites
42.8%
10.2%
11.1%
Shopify4,278 stores
92.3%
84.7%
11.6%
Wix626 sites
91.4%
1.4%
11.0%
Webflow752 sites
26.2%
12.8%
18.2%

Clean corpus. clamp() anywhere and on font-size: share of each platform’s sites with CSS captured. Fractional H1: share of each platform’s sites with an H1 captured; 14,623 of 113,043 overall (12.9%). A fractional pixel size also comes from a plain rem or em value, so the green bar measures headings sized in anything but hard pixels, not strictly fluid ones. Shopify puts clamp() on font-size on 84.7% of stores and its homepage H1s are no more fluid than anyone else’s.

The same question, seven answers

Does anyone use clamp()?SitesShare
1in unique stylesheets34,5372.0% of sheets
2on sites, anywhere42,85227.3%
3not from a named platform or embed sheet20,18212.9%
4on font-size, at least once16,94910.9%
5on font-size, five or more times6,6824.3%
6as the dominant font-size unit3080.2%
7H1 at a fractional px (any relative unit)14,62312.9% of H1s

Rows 2 and 3: of 156,880 clean sites with CSS. Rows 4 to 6: of 155,878 with a font-size declaration. Row 7: of 113,043 with an H1. Row 3 drops WordPress core, theme and plugin files, Shopify’s cloud sheets and the Wix, Squarespace and Webflow static hosts, but keeps inline styles and same-origin bundles, so it is a ceiling for hand-written clamp(), not a count.

The silver lining here is that if you make something modern and useful, people use it and you end up having an outsized influence. Not counting inline, the top six clamp() stylesheets reach 20,900 sites, when counting for overlap.

The less gracious reading is that very few independent designers are using clamp(), and that “design progress” is just moving the goalposts by elite tech design cadres. This allows them to maintain their position at the top of the hierarchy based on how well they know the new rules. A design bureaucrat, if you will.

But that is not the argument I’m making at the moment because, obviously, screens have gotten big as hell and there should be a good way to deal with that in browsers.

Embedding this loads YouTube’s player stylesheet, the file in row two of the table above. This page uses clamp() now.

BUTTONS

Buttons are some of my favorite things to design. With pseudo-classes and basic animations you can get a lot of personality out of plain CSS, and even more if you’ve got some JavaScript under your belt. So the button can tell us a lot about the creativity of the internet.

Buttons were a little tricky to count. In my experience most buttons are actually styled hyperlinks, and I didn’t want header, nav, footer or cookie-bar buttons in the numbers; a newsletter submit and an Accept All are buttons, but nobody designed them for this page. So a button here is anything with a background color and some padding, or a border and some padding, and I took the first visible one between the header and the footer on every site, re-rendered from the stored markup and stylesheets. That found a button on 94,424 of 157,549 sites. The other 63,125 had nothing that qualified: 17,714 had nothing button-like on the page at all, 13,401 had only bare text links wearing a button class, and 15,943 had buttons only in the header, footer or cookie bar. The numbers suggest the button is the most diverse element on the page: the 94,424 buttons fall into 81,979 distinct styles, and 78,417 of those styles exist on exactly one site.

Here are the most repeated buttons found on the corpus between the header and the footer. The most common one is the one nobody styled.

×
1,788 sites · the browser's unstyled button, Arial 13.333px, 2px outset
Load More
391 sites · a forum software's Load More
323 sites · a bookmarking-site tile, no text
Search
268 sites · the unstyled button at 16px
Search
107 sites · a WordPress theme's search submit
Ignore
104 sites · a notification bar's Ignore
Reject non-essential
94 sites · white with a grey border, 6px 12px, 14px, 4px radius
Search
86 sites · a 61×61 dark search square on local TV news sites
New posts
83 sites · a forum's New posts
Learn More
70 sites · a 24px-radius outline pill on newspaper sites

The ten most repeated button styles, by exact computed fill, text color, radius, size, weight, padding, border, case, shadow and letter-spacing (font family not part of the key), each re-rendered with those values on its group’s most common page background. The browser default is the only one above half a percent of the 94,424.

94,424
sites with a designed button between the header and the footer, out of 157,549 rendered (59.9%). The probe looked at 4,144,862 button-like elements to pick them.
21.0%
of those buttons have no fill or a fill identical to the surface they sit on: 12,301 outlines (border, padding, transparent) and 7,560 fills the same color as the nearest opaque background. 10,429 match the page background exactly.
87.8%
have a :hover rule of their own; 56.8% a :focus rule, 53.5% an :active rule. 69.4% have a transition, 0.3s being the favorite. Only 7.6% move on hover and 1.0% run a keyframe animation.
21.6%
are uppercase (20,383). The rest keep the case they were typed in. Letter-spacing is untouched on 81.0%.

Designed-button pass over the clean corpus, 94,424 buttons; hover, focus and transition figures from a stylesheet walk that tests each pseudo-class rule’s selector against the chosen button.

Color

Button color Buttons Share of 82,123 fills
blue (200° to 250°)17,71921.6%
white, exactly12,43615.1%
grey and other neutrals10,88513.3%
red (345° to 15°)8,59710.5%
orange (15° to 45°)8,22310.0%
green (70° to 170°)8,18510.0%
teal and cyan (170° to 200°)6,3097.7%
yellow (45° to 70°)2,8123.4%
black, exactly2,5953.2%
purple, magenta and pink (290° to 345°)2,3932.9%
indigo and violet (250° to 290°)1,9692.4%

Fill color converted to HSL; anything under 15% saturation that is not pure white or black is a neutral. The 12,301 outline buttons have no fill to classify. Blue is the color of the web’s button; add teal and it is 29.3% of fills, more than red, orange and yellow together.

Padding and font size

Most common padding Buttons Share
1px 6px (the browser default)5,4285.7%
6px 12px (Bootstrap 3 and 4)3,9284.2%
12px 24px3,2483.4%
8px 16px3,2253.4%
10px 20px2,3832.5%
0px 20px1,2761.4%
Font size Buttons Share
under 12px2,5622.7%
12 to 15px39,63242.0%
16 to 24px50,76653.8%
25px and up1,4641.6%

94,424 designed buttons. Padding as computed; the median is 9.9px vertical and 20px horizontal, and the median button is 45px tall. The single most common font size is 16px (29.0%), then 14px (19.0%) and 18px (10.5%).

So the button of the web is blue or white, 16px, sentence case, about 10px by 20px of padding, with a hover rule that swaps its background color over 0.3 seconds. One in five has no fill at all, or a fill you can’t see. The personality I was hoping to find is there, but it’s rare: fewer than one in twelve buttons moves when you hover it, one in a hundred animates, and the single most repeated button on the web is the one the browser drew because nobody got to it.

WHITE SPACE

I work with architects, and white space is an appealing concept to them. It gets batted around, but usually they’re talking about larger divs having top and bottom padding or margin. We’re going to concern ourselves with text spacing.

Line height

It’s the one spacing value everybody has an opinion on.

Role Median line-height ÷ font-size Middle half Sites measured
body 1.50 1.43 to 1.55 128,826
paragraph 1.50 1.40 to 1.60 135,288
H1 1.20 1.07 to 1.25 102,114
H2 1.20 1.10 to 1.33 118,271
button 1.44 1.15 to 1.50 109,515

Clean sites where both values resolved to pixels. The body line-height is the browser default normal on another 28,974 sites (18.4% of bodies), which works out to about 1.2 in most fonts.

So the body is 1.5 and the headings are 1.2, give or take. Both numbers are exactly what Bootstrap ships (line-height: 1.5 on the body, 1.2 on headings), and the distribution shows it: 64.5% of bodies sit between 1.4 and 1.6, and the single most common H1 ratio is exactly 1.2 on 24,060 sites, with 1.0 (16,325) and 1.1 (15,827) behind it. On paragraphs the band is wider: 20% run at 1.6 to 1.8, which is the “airy blog” setting, and 7.6% of paragraphs run under 1.2, which is cramped.

Letter spacing

Almost nobody touches it, and the higher up the page you go the more they do.

Role Normal or 0 Tightened Opened up Sites
H1 75.3% 15.4% 9.2% 113,043
H2 77.8% 12.0% 10.0% 127,888
H3 80.9% 9.2% 9.7% 111,776
paragraph 87.3% 3.5% 9.1% 146,715
body 95.4% 1.2% 3.3% 157,837
nav 93.8% 1.6% 4.6% 112,734
button 91.1% 1.2% 7.6% 138,363

computed_styles letter_spacing per role, clean sites with the element captured. Tightened is any negative pixel value, opened up any positive one.

Letter-spacing is a headline habit. One H1 in seven is tightened, usually by about -0.02em, and the rate falls with every level until body text, where 95% of sites leave it alone. When it is opened up it’s by 1px or less. The nav, which the all-caps, spaced-out portfolio nav would lead you to expect is the element with the most letter-spacing on the page, is one of the least: 93.8% at normal, and only 1.7% of navs are uppercase at all.

BREAKPOINTS

I love it when I’m talking to a prospective client and they raise their eyebrow and ask me if the site is going to be responsive. It’s a can of worms they’re about to open, and it usually ends with their eyes glazed over. For me the easiest way to show it is to change the width of a browser in front of them. I think the backbone of responsive design, at least in a way that is easily explainable, is the breakpoint.

Wide: the hero runs two columns
Wide: the hero runs two columns
Medium: the frame grid goes four across
Medium: the frame grid goes four across
Narrow: two across, then everything stacks
Narrow: two across, then everything stacks

Resizing a browser window across the breakpoints of one site. Scroll sideways for the three widths. I made this little page to show you the way breakpoints work. The movie is Le Cercle Rouge; watch the Melville stuff, this and Army of Shadows are so cool.

I’ve been working on websites since 2014 or so, and the evangelizing class of web design was pushing for mobile-first design even then. Back then I was just picking up Bootstrap templates or modifying the CSS of a WordPress site in the customizer or a child theme’s stylesheet. When you work like that, the defaults are already set and you’ll end up seeing a lot of conflicting style rules. Many of them were born from the fact that multiple stylesheets were being resolved for a single page.

Mobile sites have the easier layouts. Most of it is stacking elements on top of each other. If you are using a grid or display flex, you’re down to two or three columns at most. It took me years to figure out why it was better to have media queries for the wider views and leave the default rules for mobile.

Still, 12 years out, I run into issues caused by other people’s stylesheets and breakpoints. It gets real messy real fast. So then the question becomes: what are the most common breakpoints? If I were to write a complete stylesheet from scratch, where should my breakpoints be?

Breakpoint Sites that have it anywhere in their CSS Share
768px (min) or 767px (max) 129,300 87.0%
768px min-width 115,646 77.8%
767px max-width 100,454 67.6%
1200px min-width 82,964 55.8%
992px min-width 72,653 48.9%
600px max-width 68,258 45.9%
480px max-width 58,904 39.6%
1024px (either direction) 83,909 56.5%

148,616 clean sites whose stylesheets contain at least one width media query, out of 156,880 with any CSS captured. Counting occurrences instead of sites, 768px min-width appears 3,479,053 times and 767px max-width 2,917,948 times across the 1,734,591 unique sheets. One gap: Tailwind 4 writes its queries in range syntax (width >= 48rem), which the parser does not count, so the Tailwind cluster is undercounted by however many sites are on version 4.

768 is the iPad’s portrait width from 2010. It’s Bootstrap’s medium breakpoint, and it turns out to be Tailwind’s too, and it is the only breakpoint the two share: Bootstrap runs 576, 768, 992, 1200 and 1400; Tailwind runs 640, 768, 1024, 1280 and 1536 (written in rem, and 48rem is 768px).

That’s why it’s in 87% of sites’ CSS and nothing else comes close: it’s the one width where the two biggest frameworks and the iPad all coincide. The next cluster, 992 and 1200, is Bootstrap’s grid. 640, 1024 and 1280 are Tailwind’s. 1440 shows up in about one site in five and 1920 in one in ten, but neither comes close to the top of the table. Nothing above 1280 is a convention. The web still stops thinking about layout at 1200 pixels.

Then there’s the distinction between min and max, and that’s really about the default. min-width: 768px means these rules apply at 768 and wider, so tablet and up, and everything outside the query is the default, which is mobile. max-width: 767px is the same boundary from the other side: desktop is the default and the rules subtract as the screen shrinks, hide the sidebar, collapse the nav, un-float the columns. Why it matters: with min-width the phone, the weakest device, applies the least CSS and nothing ever has to undo anything. With max-width the phone applies the most CSS and most of it is overrides. And at a width nobody tested, a mobile-first site falls back to the simple layout and a desktop-first site falls back to the complex one. That’s also why 767 and 768 are one breakpoint in the table above: mobile-first authors write 768 min, desktop-first authors write 767 max.

How many breakpoints does a site have? More than you’d think.

Distinct breakpoint widths per site Sites Share
1 to 2 5,687 3.8%
3 to 4 6,109 4.1%
5 to 6 10,984 7.4%
7 to 10 21,905 14.7%
11 to 20 43,368 29.2%
21 or more 60,563 40.8%

148,616 clean sites with width media queries. Median 17 distinct widths. Every stylesheet a site loads is counted, including the slider plugin and the cookie banner.

It’s the clearest measure in this whole study of how little of a site’s CSS the site’s “designer” wrote. A lot of it is just handed down by a built-in framework or a vendor’s specific stylesheet. In this sense responsiveness looks and behaves like an obligation put upon website owners, and we just take the easiest way out to solve the problem, time and again. So a site becomes a patchwork of stylesheets, and the designer’s job is to tape the seams, so to speak.

Mobile-first versus desktop-first, measured by whether a site’s media queries are mostly min-width or mostly max-width:

Direction of a site’s width queries Sites Share
mostly min-width (mobile-first) 50,884 34.2%
mostly max-width (desktop-first) 50,444 33.9%
mixed, neither dominates 34,017 22.9%
max-width only 7,237 4.9%
min-width only 6,034 4.1%

148,616 clean sites. “Mostly” means one direction outnumbers the other by at least 1.5 to 1. Counting raw queries, min-width wins 72,555 sites to 74,112 for max-width, a dead heat.

We’re 15 years out from Web 2.0 and the mobile-first design evangelism, and in that time no real consensus has been reached. Mostly min-width and mostly max-width are about the same. That 23% with both used evenly, those are my favorite. The duct tape warriors. The only thing we can say is that 91% of sites have a mix of breakpoints, and that’s alright. Like I said before, it’s messy out there.

CONCLUSION

I started off this piece being a bit of an asshole to the Figma writer churning out trends pieces for the year. I think it’s wonderful that we have some new interesting sites out there that are getting recognized. So, to reiterate, the collection of all this data isn’t about saying “nuh-uh” to the idea that something is growing in popularity.

There’s this eye roll from deep in my soul that says the internet is stagnant in its design, and that there is a monopoly on the way it looks, and what it is comes from institutions that need ubiquity, like YouTube. And I think that’s a standard thing, if not a hack thing, to say: the internet is boring. It looks the same.

It’s not wrong though; the data supports it. It is in fact very homogenized, because that’s the easiest way to build sites at the speed and scale necessary for running businesses. In doing the research, we can point to interesting open source projects, like Tailwind, and large orgs, like YouTube, building out infrastructure to make themselves easy to apply. These types of projects have a fundamentally deeper effect on what the web looks like at the baseline than the trends article.

Right now the web page is a base 16px font, and we have rules about the relationship between headings, font sizing, letter-spacing and line-height based on the data. Small changes across these attributes distinguish the work. We can map our choices against the industry standard cards we created earlier.

For me, I’m imagining rolling my own underscores theme with the standard page’s measurements, modernized: rem for everything up to 768px, and clamp() for everything above it. Best practices for clamp() have yet to be calculated. Maybe next week.

On the horizon, the most interesting question I’m left with is what LLMs are going to do to the standard page. They’re easier to interact with than frameworks, content management systems and the rest. The barrier to entry to making websites has just dropped significantly. Part of me is excited. I hope more people will stretch in their design choices now that it’s so much easier to get the basics down.

But that sick little voice is telling me that next year, when I run this project back, we’re just going to see a new footprint: the Claude or ChatGPT website footprint. And that’s fine.