The AI website tech stack: which tools you actually need

"Which tools do I need for an AI website?" usually gets answered with a list of AI tools. That's the easy part. Claude, ChatGPT and the AI site builders all write a decent homepage in minutes now. What makes it a website a business can run is everything around that homepage: where it lives, how a contact request reaches you, and who changes the opening hours in March.

This guide goes through the stack layer by layer, says which layers you can skip, and links our deeper guide for each one. It also uses our own data: a fresh count of how many small-business websites are already built with AI, and what the sites we rebuilt in 2026 actually contained.

Key takeaways

• AI-built sites are still rare. Of 2,538 established small-business websites in Germany, Austria and Switzerland we checked in November 2026, 0.4% were clearly or very likely made with an AI site builder, and at most 3.1% looked like code an AI tool would write.
• New companies are ahead. Among companies founded in Germany in 2026, 36% of the new websites run without any CMS or website builder, and up to 15% look AI-written.
• The site is the small part. Among the sites we rebuilt in 2026, one in five had forms, one in five a cookie consent tool and almost a quarter Google Tag Manager. Each of those is a layer someone has to own.
• Every add-on costs speed. In our PageSpeed data, sites with a tag manager or tracking pixel needed a median 11.0 seconds to show their main content on a phone, against 6.4 seconds without.
• Pick the editing layer before you build. Whoever changes the site after launch decides the rest of the stack.

The AI website tech stack at a glance

1. The facts: the layer no tool replaces

The clearest predictor of a good AI-built site isn't the model. It's whether the request contained a fact the model couldn't have guessed: the Saturday opening hours, the three services that actually pay, the reason customers pick you over the shop next door. Write those into a short brief, and in Claude Code into a CLAUDE.md file in the project, so every session starts from them. Our guide to building a website with Claude Code has a template.

2. Design: drafts, skills or a designer's file

Left alone, every AI model falls back on the same look: a centred hero, a soft gradient, three cards. Three ways out. Draft visually in Claude Design or Figma and build from the export. Load a design skill that gives Claude stronger defaults; we compared them in the best Claude skills for web design and wrote a separate guide to the impeccable skill. Or bring a designer's file and have the AI only build it.

3. The model and the builder

There are two families of builder. Coding agents such as Claude Code work in a folder of files you own and can move anywhere. AI site builders such as Lovable, v0 or Bolt give you a chat and a preview in the browser, and usually a hosted app at the end. For Claude Code, the model that matters is Opus 5.5 for design and building and Sonnet 5.5 for small edits; the details are in which Claude model is best for web design. If you're still choosing between the two big assistants, see Claude vs ChatGPT for websites. Claude Pro, which includes Claude Code, costs $20 a month, or $17 a month billed yearly.

4. The code format: readable HTML beats a clever app

This is the layer that gets decided by accident. Many AI site builders produce a single-page React app: the HTML the server sends is an empty container, and the content only appears once JavaScript has run. Google can render that, but not every crawler does, and every visitor waits for the script first. For a business site with ten pages and a contact form, ask for plain HTML and CSS, or a static generator such as Astro that outputs plain HTML. Our guide on static website SEO explains why that's the safer format for search.

It is also where AI-built sites give themselves away. In our new count, 0.4% (10 sites) of established small-business sites were made with an AI site builder, and 6 of those 10 were exactly this kind of app. Sites that Claude or ChatGPT wrote as plain files leave no fingerprint at all, which is why the real share lies somewhere up to 3.1%. The full numbers are in our study of AI-built small-business websites.

5. Hosting

Static files are cheap to host. The catch is the fine print: GitHub Pages says it isn't meant for running an online business, and Vercel's free Hobby plan is for non-commercial use only. Cloudflare Pages and Netlify have free plans that allow a business site. Our guide on where to host a website you built with Claude compares the four and covers what hosting it yourself means over a year.

6. Forms and email: the layer that earns the money

A static site has no server to send an email, so a contact form needs a back end. Netlify has a form feature built in; elsewhere you add a form service or a platform that collects messages for you. Test it with a real message before launch, and again after every redesign. A form that silently stops working is the most expensive bug a small-business site can have, because nobody notices it.

7. Editing afterwards: decide this before you build

The first change request arrives within weeks. If you're the only editor and happy in a terminal, Claude Code stays your editor. If a colleague or a client needs to change things, you need a CMS or a visual editor, and some of them need the site built in a particular way. That's why this layer comes before the build, not after. Which CMS fits a website built with Claude Code compares Decap, TinaCMS, CloudCannon, headless options and managed platforms.

What small-business sites actually need

Stacks are easy to over-build. Here is what the sites we rebuilt in 2026 actually used, as detected by our copier. It's a good check list for which layers your stack must carry.

ShiftPress data, 2026: features our copier detected on the sites it copied, share of sites. A site can have several.

8. Measurement and consent: every tool costs speed

Each script you add is one more thing loading before your content. In our PageSpeed data on more than 130 small-business sites, the median site needed 7.1 seconds on a phone to show its main content, where Google calls 2.5 seconds good. Sites with a tag manager or tracking pixel were at 11.0 seconds, sites with a cookie consent tool at 8.7 seconds against 6.3 without. That's a pattern, not proof, but it's a strong argument for adding only the tools you actually read. How to build a fast website with AI has the checklist and a set of rules to give Claude.

9. Search and AI search

The basics are the same for an AI-built site: one clear title and description per page, a sitemap, Search Console, and structured data that says who you are. AI search adds two habits: answer the question in the first lines of a page, and keep the facts (prices, areas, opening hours) in plain text rather than images. Our technical SEO checklist and the explainer on llms.txt cover the details.

10. Domain, email and legal pages

Moving a domain to a new host is where email breaks, because the mail settings live in the same DNS zone. Read how to change DNS without breaking email before you switch. In Germany and Austria a business website also needs a legal notice (Impressum), and a privacy policy is expected across the EU and in Switzerland. An AI can draft both, but it can't know your register number or your data processors, so check every line. No site yet? You can describe your business and get a first version that you then change by chat.

Three stacks that work

Prices as listed by the vendors in November 2026. TinaCMS has a free plan for two users; CloudCannon Standard is $55 a month, or $49 billed yearly.

Frequently asked questions

The bottom line

The AI part of the stack is the part that got easy. Spend your choices on the layers that decide whether the site works for a business: plain, readable code, a form that reaches you, an editing path for whoever changes it, and as few extra scripts as you can live with.