How Many Small-Business Websites Are Built with AI? (2026 Study)
The sample
What they run on
* Code without a CMS or builder that uses the modern building blocks AI tools reach for (Tailwind, React, Next.js, Astro) and none of the older ones (jQuery, Bootstrap). Not every such site is AI-made, so this is an upper bound. 0.4% = 5 sites with a clear AI-builder fingerprint plus 5 apps in exactly the shape these builders produce.
Key takeaways
- The trend is clear: AI grows with new companies. Sites from AI site builders are 525% more common among companies founded in Germany in 2026 (2.5%) than among established businesses (0.4%), and AI-style code is 380% more common (15% against 3.1%). Details on new companies are in our study of 1,536 new companies.
- AI-built business sites are still rare. 0.4% of established small-business websites carry the fingerprint of an AI site builder. With this sample size the true share lies between roughly 0.2% and 0.7%.
- Even the generous count stays small. At most 3.1% of the sites are modern code in the style AI tools write, against up to 15% among companies founded in Germany in 2026 (our study of new companies). AI arrives with new sites, not by rebuilding old ones.
- Restaurants lead. 7 of the 10 AI-built sites belong to restaurants, cafés or bars: 1.2% of all food and drink sites, against 0.2% of other businesses.
- WordPress is still the default. 38% of the sites run on it, 45% in Austria. 37% of those use the Elementor page builder.
- The code route is the fast one. Sites built as code were 4.0 s to the main content in our mobile PageSpeed test, WordPress 7.8 s and website builders 6.2 s (medians).
- The web has a lot of dead links. 18% of the websites listed for these businesses didn't respond, and another 2% showed only a placeholder or a domain for sale.
Who we looked at
Most surveys of "AI websites" count what AI tools say about themselves. We wanted the other side: how many ordinary businesses actually run a site an AI built. So we started from OpenStreetMap, the open map that lists shops, practices, offices, restaurants, hotels and trades together with their website. We took every business with a website in the city centres of 16 cities: Hamburg, Munich, Cologne, Frankfurt, Stuttgart, Leipzig, Freiburg, Münster, Nuremberg and Hanover in Germany, Vienna, Graz and Linz in Austria, and Zurich, Basel and Bern in Switzerland. Chains are excluded (anything carrying a brand tag). That gave 35,260 distinct websites.
From each city we drew 200 at random, 3,200 in total, and fetched each homepage once in November 2026. 2,538 of them were real, reachable websites: 1,559 in Germany, 499 in Austria and 480 in Switzerland. These are established businesses that someone has bothered to put on a map, so the sample leans towards shops, restaurants and practices with a street address.
What small-business websites run on
WordPress is still the system of choice, and in WordPress the page builder decides how the site is edited: 37% of the WordPress sites use Elementor, 18% WPBakery and 13% Divi. TYPO3 (6%) is the German-speaking speciality, mostly on larger practices and firms. Among the website builders, Wix (4.1%) and Jimdo (3.3%) lead, followed by Squarespace and Webflow (1.7% each) and the engines behind the STRATO, IONOS and Telekom builders. Swiss businesses use builders most (18%), Austrian ones WordPress most (45%).
A quarter of the sites show no recognisable system at all. Most of them are older hand-built sites: classic HTML, often with jQuery. Only a small part of this group looks like the code AI tools write today.
How many are built with AI?
We looked for the traces AI site builders leave in a homepage: Lovable's script and asset paths, Durable's image CDN, the generator tags of v0, Bolt, Hostinger Horizons and others, and the bare shape of a Vite and React app that these builders ship (an empty root element and a single hashed script). Every hit was checked by hand. Two hits that looked like Base44 turned out to be Wix sites (Wix owns Base44 and lists its domain in a script), and one "app" was a parked domain. We removed all three.
What's left: 5 sites with a clear fingerprint (4 Lovable, 1 Durable) and 5 more apps in exactly the shape these builders produce. 10 of 2,538 is 0.4%. With a sample this size the true share lies between roughly 0.2% and 0.7%.
That is a lower bound. A site that someone wrote with ChatGPT or Claude and uploaded as plain files leaves no trace. The generous upper bound is every site built as code in the modern style AI tools use: 79 sites, or 3.1%. Two thirds of those use a code framework such as Next.js or Astro, which agencies use too, so the real AI share is lower than that.
| Group | AI-builder sites | Modern code (upper bound) | WordPress |
|---|---|---|---|
| Established businesses, DACH (this study) | 0.4% | 3.1% | 38% |
| Companies founded in Germany in 2026 | 2.5% | 15% | 35% |
The comparison is the real finding, and the trend is clear. Among companies founded in 2026, sites from AI site builders are 525% more common than among established businesses: 2.5% against 0.4%, six times the share. Modern code in the style AI tools write is 380% more common: 15% against 3.1%, almost five times. AI doesn't replace existing websites yet. It shows up where a site is built from scratch, and every new company adds to that share. The numbers for new companies come from our study New companies 2026: what their websites run on. The two samples differ (new register entries in three German states, established businesses in 16 cities across Germany, Austria and Switzerland), so read the gap as a clear direction rather than an exact figure.
Why restaurants?
7 of the 10 AI-built sites belong to restaurants, cafés and bars, from an Italian restaurant in Bern to a sushi bar in Austria. That's 1.2% of the 605 food and drink sites in the sample, against 0.2% of the other 1,933 businesses. Our reading: a restaurant site is mostly one page with photos, a menu and opening hours, exactly what an AI builder produces well from a short prompt, and restaurants change owners and concepts often, so their sites are rebuilt more often too.
Is the AI format faster?
We ran Google PageSpeed Insights (mobile, the same test as our free website check) on a stratified sample of 423 sites from the study, nine in ten of them in Germany. The time to the main content (Largest Contentful Paint) splits clearly by how the site is built:
| How the site is built | Median time to main content | Median page weight | Median PageSpeed score | Main content within 2.5 s |
|---|---|---|---|---|
| Built as code (no CMS or builder) | 4.0 s | 1.25 MB | 78 | 37% |
| Other systems (TYPO3, Contao, shops …) | 5.5 s | 1.61 MB | 72 | 21% |
| Website builder (Wix, Jimdo …) | 6.2 s | 2.28 MB | 72 | 10% |
| WordPress | 7.8 s | 2.63 MB | 67 | 9% |
Sites built as code are lighter and show their content sooner. That matches what our website carbon report found for new companies. It is a pattern, not proof: a hand-coded site is often simply a smaller site. But it shows that the format AI tools write is not the problem. The add-ons are: in our separate data on sites we rebuilt, those with a tag manager or tracking pixel needed a median 11.0 seconds to show their main content. Our guide on how to build a fast website with AI turns that into a checklist. The quickest version of all this is to let ShiftPress build the first draft and spend your time on the facts only you know.
What this means for your business
1. You are not late. If you are thinking about building your site with AI, almost none of your competitors have done it yet. Four in ten still run WordPress, many with a page builder that makes every change a small project.
2. The tool matters less than what comes after. An AI builder gets you a first version fast. Then the site needs a host that allows business use, a form that reaches you, an editor for whoever changes it, and someone who keeps it fast. Our guides cover each step: the AI website tech stack, where to host a site built with Claude and a CMS for a site built with Claude Code.
3. Ask for plain HTML. 6 of the 10 AI-built sites we found send an empty page that only fills once JavaScript runs. For a business that wants to be found, plain HTML is the safer format, and it's the one that came out fastest.
Already built your site with an AI?
Upload it to ShiftPress as a ZIP: a Claude Design export or plain .html pages, up to 30 per file.
Every page becomes a real page you can edit, and from then on you change it by describing what you want. Free to
start, no credit card. Your own domain comes with a paid plan.
How we collected this
- Source of the businesses: OpenStreetMap, queried via the public Overpass API on 1 November 2026. Objects with a website and a shop, craft, office, healthcare, hotel or selected amenity tag (restaurants, cafés, bars, practices, driving schools and similar) inside a city-centre bounding box. Objects with a brand tag (chains) were excluded. Social media profiles and directory listings were dropped. One row per domain.
- Sample: 200 websites per city drawn at random (fixed seed), 3,200 in total. Each homepage was fetched once, following redirects.
- Excluded: 591 sites that didn't respond (offline, refused or timed out), 61 placeholders or domains for sale, 8 bot-protection pages, 2 links that led to a social profile.
- Recognising the system: fingerprints in the homepage code and response headers (paths such as
/wp-content/, generator tags, builder scripts and CDNs, framework markers). Every AI-builder hit was checked by hand; three false positives were removed. - Speed: Google PageSpeed Insights API, mobile strategy, on a stratified sample of the analysed sites (all AI-builder sites, and random samples of WordPress, builder, code-built and other sites). Most measured sites are German (381 of 423), so the speed figures describe Germany best.
- Limits: city centres only, so rural businesses are under-represented. OpenStreetMap lists what volunteers mapped, which favours shops and restaurants. Sites built with AI and uploaded as plain files can't be detected, so the AI share is a lower bound. Small groups (10 AI sites) allow no comparison between countries.
- Data protection: we publish only aggregated figures. No list of websites. Business names in the text are not given; the examples describe the type of business only.