A Website with AI in Eight Phases: What a Video Promises and What We Measured

On 9 October 2026 Julian Ivanov published a 33-minute video on his German-language AI automation channel, titled "Building websites with AI that look good and perform (free skill)", in the original: "Websites mit KI bauen, die gut aussehen UND performen (Gratis Skill)". It takes an invented brand, an AI and automation consultancy for skilled trades, from zero to a live site in eight phases: name and logo, keyword research, copy, design direction, build, independent review, launch, Google Search Console. Ivanov packaged the workflow as a skill and gives the instructions away for free. The video names DataForSEO as sponsor, and the launch host is Hostinger with a discount code.

We did not run the skill. We measured the statements around the video: the GitHub numbers, the project's age, the rule catalog, the reachability of the download. The result shows how much of this workflow is already practice on our side, and where the road leads into paid services.

What the skill promises

The agent works one phase at a time and keeps its state in a website-status file inside the project folder, so each phase can run in a fresh chat. Ivanov's own rule is to compact or restart once the context window is half full. Phase 1 settles the name, free domains, logo variants, target group and the goal of the site. Phase 2 collects design references and finds keywords through DataForSEO: real search volumes, current Google results, the live answers of ChatGPT. Phase 3 writes the copy before any design, meta fields included. Phase 4 develops several design directions with the plugin impeccable and one of them gets picked. Phase 5 builds the site. Phase 6 lets an independent reviewer criticize the result, a subagent that did not build the page, and leftover fixes move on sliders. Phase 7 ships robots.txt, sitemap, imprint, privacy page and hosting into the internet. Phase 8 registers the site with Google Search Console. One add-on checks at the end whether a cookie banner or terms of business are needed at all.

Six numbers, checked again

The star count is honest, the project is young: the repository was created on 16 November 2025, roughly eleven months before the measurement. The pattern count is rounded, the README states 59 rules, and those catch exactly the marks of AI frontends you know from the other side. Inter for everything, purple-to-blue gradients, a card inside a card, gray text on colored ground. Nobody outside the author's account can re-check the 25 cents, because DataForSEO bills by consumption. An account there is a decision of its own.

Copy before design is the same idea as our text lint

The core argument of the process: the copy determines the sections and their space, so the copy has to exist before the design. On the design side the video argues that models were trained on the same templates and therefore keep rebuilding the same pages. That is a measurable claim, and impeccable ships it as a rule catalog. We run the equivalent check for our own text archive, most recently on 22 September 2026: 145 German posts, two pass. The most frequent violation was the em dash, 13 occurrences per 1,000 words. Machine-written text betrays itself through structure more than through vocabulary. The video demonstrates exactly that thought for design; on the text side we keep it as a checklist.

Being cited by ChatGPT is measurement work too

Ivanov advises checking after three or four weeks whether Google and ChatGPT show the page. The prerequisites for that were measured for our own presence on 14 September 2026 and are written up in GEO: Why the AI Search Engines Missed Us: crawler permission, a machine-readable summary, structured data, plus a blind spot we found in our own setup. Phase eight of the video fits that pattern. Visibility is a measured value you keep inside a routine.

The part that is commerce

Two observations belong in a level-headed review. The research in the video runs over DataForSEO, which sponsors the channel, with a signup link that carries starting credit. The launch runs over Hostinger with a discount code. Both are decisions with an account and costs attached, and the skill leads straight into them. That is ordinary creator economy, and the vendor choices inside a ready-bought process were still made by somebody else.

The legal add-on at the end, cookie banner usually unnecessary, terms only when you sell directly, is Ivanov's simplification for his example and no legal advice. Whether your own site needs either belongs in front of a lawyer or a trade association. Foreign prompt packs do not reach our machines unexamined either. Our procedure there is called Four files before the installation, and what bundled instructions and plugins may do inside Claude Code has been documented in detail since 5 October right here.

What we keep from it

Three habits of this workflow are worth copying with or without the skill. The state file instead of the hour-long chat: an agent that reads its own progress from a file stays reliable across days. The independent reviewer: our gates are scripts (voice lint, link check, SEO audit) rather than an agent, but the logic behind them, that a second eye sees what the builder stops seeing, holds for both. And search data instead of gut feeling: your own queries arrive free in Search Console, competitor search volumes only exist at data providers. Anyone building sites for customers gets a finished schema from the video. Anyone who runs their own process and gates, our case, needs only the state file from it.

Further reading

• Building websites with AI that look good and perform (free skill), Julian Ivanov, 9 October 2026 (German)
• impeccable on GitHub, 79,822 stars at the measurement date 11 October 2026
• impeccable documentation
• Anthropic's frontend-design skill, the origin project per the README
• DataForSEO, sponsor of the video