Why Everything Looks the Same
. Study 002.
Every few years somebody writes that the web used to look like something and now every page looks like the same page. I wanted to know how much of that survives a measurement, so I measured 40 homepages in a browser, and then the same twelve pages as they looked in 2001 and in 2011.
The complaint is usually about four things: the hero section at the top, the typography, the colour, and the layout underneath. Those are all measurable, so this study measures them instead of arguing about them. What follows is what the numbers support, and, further down, the places where they do not support the complaint at all.
How it was measured
Each page was loaded once in a headless Chrome at 1440 by 900 points, left alone for six seconds, and then asked about its own computed styles: the background colour behind the page, the largest piece of text in the first screen and how it is set, the first thing that looks like a button, the width the content is capped at, whether a header sticks to the top, and how many rows of three or more equally wide boxes exist. Nothing was clicked, no cookie banner was dismissed, and no page was scrolled.
The sample is forty homepages I picked by hand, ten from each of four kinds. Picking them by hand is a limit worth naming: a different forty would give different numbers, and this is not a random sample of the web. All forty loaded, so the figures below rest on 40 pages.
Product and SaaS: airtable, asana, figma, intercom, linear, loom, notion, slack, stripe, vercel.
Design studios: activetheory, area17, basic, buzzworthy, dogstudio, hellomonday, instrument, locomotive, pentagram, resn.
Shops and consumer brands: aesop, allbirds, casper, everlane, glossier, ikea, muji, nike, patagonia, warbyparker.
Personal sites and blogs: adactio, brunosimon, daringfireball, jvns, kottke, lynnandtonic, sarahdrasner, simonwillison, tholman, waxy.
Two things about the machine are worth naming. It sits in Berlin, so a few of the pages served their German version, which changes the words in the specimens below and nothing about the geometry. And it reports a desktop browser at one window size, so every number here describes the wide view of a page, not the phone.
To see whether the measurement is stable, I ran the whole sample twice, about twenty minutes apart. The background colour came out identical on 40 of 40 pages and the headline size on 40 of 40. Nothing in those two measurements moved, which says more about the measurement than about the pages: a second visit from the same machine sees the same layout.
What forty homepages have in common
The first thing the numbers show is not a style. It is a shape.
The colour, as measured
Every stripe below is one page's measured background, sorted from dark to light. It is not a diagram of the data. It is the data.
40 homepages, background colour as computed by the browser, 24 of them above 85 per cent lightness.
The button, as measured
The first button-shaped thing in the first screen, drawn at the radius it actually has.
The headline, as measured
The smallest, the median and the largest display-sized headline in the sample, each set at the size and weight its page uses. The median page uses 7 distinct text sizes and 2 font families in the whole document, and the opening screen of the median page is exactly one window tall.
Where the complaint does not hold
Five traits recur often enough to call them a house style: a near-white background, a centred headline, a content column capped between 1000 and 1500 points, a header that sticks, and a row of three equal cards. None of the 40 pages carries all five. 9 carry four of them. The interesting part is who they are.
Product and SaaS. 8 of 10 on a near-white background, 8 of 10 with a centred headline, 7 of 10 carrying four or more of the five traits.
Design studios. 7 of 10 on a near-white background, 6 of 10 with a centred headline, 1 of 10 carrying four or more of the five traits.
Shops and consumer brands. 7 of 10 on a near-white background, 4 of 10 with a centred headline, 1 of 10 carrying four or more of the five traits.
Personal sites and blogs. 2 of 10 on a near-white background, 2 of 10 with a centred headline, 0 of 10 carrying four or more of the five traits.
The absences say as much as the averages. 12 of the 40 pages never cap their content width at all, 20 have nothing button-shaped in the first screen, and on 1 there is no display-sized text in the first screen to measure, because the opening screen is a canvas. All of those sit in the two groups that are not selling software.
So the sentence that everything looks the same is too wide. The traits cluster: Product and SaaS carry them most often, Personal sites and blogs least. The pages that keep doing something else are the ones with the least to lose when a visitor leaves, and what they do instead is visible in the same numbers: darker backgrounds, headlines that start at the left margin, their own typefaces, and no row of three cards.
The same twelve pages, twenty five years apart
A snapshot of today cannot show convergence on its own, so I ran the same measurement over twelve homepages that existed in 2001 and still exist: Apple, Amazon, the BBC, the New York Times, CNN, Microsoft, eBay, Craigslist, IMDb, NASA, MIT and the Guardian. The old versions come from the Internet Archive, rendered with their own stylesheets and images, and the archive's toolbar is excluded from every measurement. All 36 requests came back.
The first difference is not a style at all. In 2001 the browser could find display-sized text in the opening screen of 2 of the 12 pages, because the rest of what looked like a headline was a picture of one. In 2011 that was 8, today it is 11. Where it is measurable, the median headline grew from 14 points to 16 and then to 36.
The structure moved in one direction. Pages that cap their content width in the stylesheet: 1 in 2001, 1 in 2011, 9 today. Rows of three or more equally wide boxes: 0, then 0, then 6. Headers that stick to the top: 0, then 0, then 1.
One number runs the other way, and it is the one most people would guess wrong. The median page in this group used 2 distinct text sizes in 2001 and 8 today, with the same 2 font families. Pages have more typographic levels now than they used to, and still feel more alike, because the levels sit in the same arrangement.
Two limits belong here rather than in a footnote. The archive renders old pages with whatever it managed to capture, so a missing stylesheet can make a page look plainer than it was. And the single page that reports a capped width in 2001 is one I could not reproduce by hand, so read that column as at most one.
What the numbers do not show
They do not show quality. A centred headline on white is not a failure, and a dark left aligned one is not automatically braver. They do not show intent either: the same shape can come from a template, from a design system, or from a deliberate decision that happens to land in the same place.
They also cannot answer the question about generated sites. To say anything about what AI tools did to the look of the web, I would need a sample of pages whose build date and tooling I actually know, and that sample does not exist here. Anyone claiming a number for it right now, including me, is guessing.
The convergence that shows up in these numbers is not a fashion. It is the shape of a decision that nobody has to defend: a capped column behaves at every window size, a sticky header is always reachable, a row of three cards fits whatever you have to say, and white is what a screenshot looks good on. Each choice is defensible on its own. Together they are a form that fits everything and belongs to nobody.
What the measurements changed in my own work is smaller than the argument. The sites in this sample that look like themselves do not do it with effects. They do it by deciding one thing early, usually the type or the colour, and then letting the rest follow from it. That costs no frames, needs no framework, and is the part a generator cannot do for you, because it is the part where somebody has to choose.
What this was built from
The measuring tool and the raw numbers are next to this page, so the run can be repeated or contradicted: messen.js drives the browser, im-browser.js is what runs inside each page, and the results are messwerte.json and messwerte2.json for the two runs over today's sample. The archived pages are in archivwerte2.json, which is the run every number above comes from; archivwerte.json is an earlier attempt that requested the raw archived files without their stylesheets, kept here because it shows how much that choice changes the result. The two sample lists are seiten.json and archiv2.json.
Sources
- John Allsopp: A Dao of Web Design. A List Apart, 7 April 2000. Retrieved 16.10.2026.
- Ethan Marcotte: Responsive Web Design. A List Apart, 25 May 2010. Retrieved 16.10.2026.
- Bootstrap: About, on its origin at Twitter and its first public release on 19 August 2011. Retrieved 16.10.2026.
- W3Techs: Usage statistics of Bootstrap, reporting Bootstrap on 14.6 per cent of all websites. Retrieved 16.10.2026.
- Google: Material Design 3, the design system as published. Retrieved 16.10.2026.
- Tailwind CSS: Styling with utility classes. Retrieved 16.10.2026.
- Internet Archive: Wayback Machine, source of the 2001 and 2011 snapshots, requested in raw mode so the archive's own toolbar is not part of the measurement. Retrieved 16.10.2026.