The State of AI Web Design in 2026: What the Public Data Shows

Everyone who looks at AI-built websites in 2026 has an impression: they all look the same, they are all purple, they all use Inter. Impressions are easy. Data is harder, and the best public data available says something more interesting than the impression.

This page used to report a study of our own: "1,000 AI-built sites" scanned between January and March 2026, with averages, distributions and trends to one decimal place. We hold no dataset for it, so it has been removed in full. What replaces it is the largest public measurement of the web there is, the HTTP Archive's Web Almanac, whose 2025 edition was published on 15 January 2026, plus a few other dated sources. Where the data stops, this page says so.

The source: the HTTP Archive and its Web Almanac

The HTTP Archive crawls millions of sites every month and publishes the raw data; the Web Almanac is its yearly analysis, written and reviewed by named contributors. The 2025 edition has a chapter on generative AI (Chapter 4, by Christian Liebel, Yash Vekaria and Jonathan Pagel) and one on fonts (Chapter 1). Most of the figures below come from those two chapters.

Finding 1: AI-built sites are now visible in the crawl

The generative AI chapter counted the subdomains of "vibe coding" platforms in the HTTP Archive dataset:

Two cautions the authors make themselves: the counts only cover sites on the platforms' own subdomains (anything on a custom domain is invisible here), and only pages that made it into the dataset. Lovable's own figure, quoted in the chapter, is "100,000 new projects" a day. The gap between the two numbers is a reminder of how much generated output never becomes a public, crawled website.

Finding 2: the "AI purple" is real on the web, and mostly a Tailwind effect

The chapter took on the most repeated claim about AI design, that generated sites are purple, head-on. It tracked Tailwind's indigo-500 (#6366f1) and two other colors commonly called "AI colors," #8b5cf6 (violet) and #a855f7 (purple), in the CSS of root pages, quarterly from Q3 2022 to Q4 2025:

Read the two rows together. Across the whole web, these colors rose sharply. Among Tailwind sites, their share stayed flat. The authors' conclusion: the trend "may be less a result of AI 'choosing' purple and more a reflection of the general rise in Tailwind's popularity." They also cite Adam Wathan, Tailwind's creator, who has said the Tailwind team used deep purples heavily in its documentation and examples in the early 2020s, around the time models were trained on that material.

The method has limits, which the authors state: they read CSS variables only, so hard-coded hex values and older preprocessors are invisible, and a one-point change in lightness would escape detection. Their figures are a conservative floor.

Finding 3: no measurable surge in gradients, shadows or fonts

The same chapter: "We also tested for surges in gradients, shadows, and specific fonts, but found no statistically significant increases."

That is the most important sentence in this article. The sameness of AI-built pages is easy to see on any single page and on any showcase of generated sites. At the scale of the whole web, three years after ChatGPT, it does not show up as a statistically significant increase in those features. Two explanations fit: generated sites are still a small share of the crawl, and the features they share (Tailwind's palette, soft shadows, gradients, a neutral sans-serif) were already common on hand-built sites.

Finding 4: Inter is rising, from a small base

The fonts chapter gives the shares for the most-declared families: Roboto on about 10 to 10.7% of pages, Poppins 5.8 to 6.0%, Open Sans 5.0 to 5.6%, Montserrat around 3.3 to 3.9%. And: "One notable climber is Inter, which is now on roughly 1.4-1.5% of pages," a rise the authors put down to "adoption by many modern frameworks and designers."

So Inter is a fingerprint of a particular stack and era, not of the web as a whole. On a page that also carries Tailwind's palette and stock components, it is one more default. On its own, it is a popular, well-made typeface.

Finding 5: Tailwind's footprint depends on who measures

The Almanac's color analysis rests on Tailwind's growth, and Tailwind's share of the web varies with the method. W3Techs, which surveys the top sites by its own method, reports Tailwind on 0.3% of all websites and on 1.8% of the websites whose CSS framework it can identify (W3Techs, October 2026). The HTTP Archive identifies Tailwind with its own fork of Wappalyzer. Different methods, different denominators: compare trends within a source, not numbers across sources.

Finding 6: sites are starting to address AI agents directly

Two smaller signals from the same chapter, both measured in 2025:

• llms.txt, a proposed file that offers a site's content to language models, is valid on about 2.1% of pages (2.13% desktop, 2.1% mobile).
• AI crawlers in robots.txt: mentions of GPTBot rose from about 2.6% of robots.txt files in 2024 to about 4.5% in 2025, and are far more common on popular sites (20.9% of the top 1,000). Nearly all such directives disallow access.

What the data does not tell you

• Which tool built which site. Platform subdomains are the only reliable marker at crawl scale, and most serious projects move to their own domain.
• How a generated page looks as a whole. The Almanac counts features one at a time. The recognizable "AI look" is a combination: a centered hero, three equal cards, a blue or indigo accent, a neutral sans-serif, fade-up motion. No public dataset measures the combination.
• Whether a page converts. Nothing above measures what visitors do.

That last gap is where tools like Sailop's free scan sit: it reports, for one page, which of eight common signatures are present. Three saved results from 13 October 2026 show the spread: tailwindcss.com carries seven of the eight, flowbite.com four, stripe.com one. That is three pages, not a survey, and the scan measures defaults, not quality.

What to do with this

• Do not design to a statistic. The sameness that matters is on your page and your competitors' pages, not in a web-wide average. Look at the five sites a customer would compare you with.
• Choose your accent. The Almanac shows how far Tailwind's palette has spread with Tailwind itself. If your accent is a stock swatch, you share it with every other site that never changed it. See how to pick an accent color that isn't Tailwind blue.
• Choose your typefaces. Inter is good and increasingly common; a pairing you chose for the content is better. See the typography problem.
• Break the default structure. The three-card row and the centered hero are what make generated pages recognizable at a glance. See the card-grid problem.
• Scan before you ship. Run the free scan on the page you are about to publish, and fix what it finds.

For what "AI slop" means and where the word comes from, see the AI slop guide. For the patterns themselves, see 90+ AI design patterns to avoid.

Sources

• HTTP Archive, Web Almanac 2025, Chapter 4: Generative AI, by Christian Liebel, Yash Vekaria and Jonathan Pagel, published 15 January 2026.
• HTTP Archive, Web Almanac 2025, Chapter 1: Fonts, published 15 January 2026.
• W3Techs, Usage statistics of Tailwind, read 17 October 2026.
• Sailop's saved scans of tailwindcss.com, flowbite.com and stripe.com, 13 October 2026.

Revision note

Rewritten on 17 October 2026. The earlier version, "Data from Scanning 1,000 Sites," presented a study we hold no data for: a sample of 1,000 AI-built sites, 200 human-built ones, score distributions, pattern frequencies, trends from a 2025 sample and framework breakdowns. Every one of those figures has been removed. The article now reports the HTTP Archive's published measurements, which in places contradict the old figures: the web-wide rise of "AI colors" tracks Tailwind's growth, and there is no statistically significant surge in gradients, shadows or fonts.