A website from a developer, not just from ChatGPT

Does Google penalize AI generated websites?

No, Google penalizes thin content, regardless of who wrote it. The core update of March 2026 targeted mass produced pages without value of their own; affected sites lost 50 to 80 percent of their organic traffic within two weeks. The benchmark was intent and result, not the production method.

That is an important distinction, because the widespread short version, that AI text gets penalized, is wrong and points in the wrong direction. Thin text written by hand is treated exactly the same. Conversely, generated text that contains its own measurements, its own experience and verifiable information holds up. The difference is the substance, not the tool.

How do you recognize a generated website?

By a bundle of design features that generators reproduce over and over. The most frequent are: Inter as the only typeface, a gradient from indigo to violet, three or four cards of equal size in a row, glassmorphism as the base note, floating color shapes behind the title area, a large centered icon above the heading, neon borders on cards in permanently dark layouts, and in the header a rounded mark to the left of a bold name.

The most reliable single marker, though, is not a design element but the text: vague statements that do not say what is being sold. “We deliver tailored solutions for your digital success” could sit on any competitor’s site. That is exactly the test: any sentence that could sit unchanged on somebody else’s website is filler.

Do your customers even notice?

Some do, and it is the ones that matter. Studies on detecting generated design report hit rates of 75 to 85 percent for trained observers and 55 to 60 percent for average users. So the damage does not arise across the board but with the observers who judge professionally.

For an IT service provider that is the actual target group; for a trades business it matters less. Even so, the connection holds there indirectly: whatever is recognized as generic feels interchangeable, and interchangeable means comparable on price. A website that looks like ten others no longer argues on quality.

What did this cost me personally?

Feedback on my own draft and a rebuild of the header. In an internal tool I had built a header that followed exactly that pattern: rounded mark on the left, bold name beside it. The feedback was short: that looks like AI output. It was justified, and it was uncomfortable, because it came from somebody I wanted to sell AI systems to afterwards.

Out of that came eight binding build rules that apply to this website and to client projects. Four of them are prohibitions: no gradient as a brand element, no card grid as the default answer, no glassmorphism with floating color shapes and neon borders, no stock photos. The fifth concerns the header: name first, mark behind it. The remaining three demand something: every heading names something that applies only to this company, motion exists only with a function and with an alternative for reduced motion, and every service page shows substance: price range, process, duration, technology in plain language. That is why this website uses Manrope instead of Inter and flat color instead of a gradient.

Where does AI genuinely help in building a website?

With everything that is drafting, translation and checking, which is a large part of the work. I use language models daily: for first text drafts that are then reworked, for researching legal bases including the citation, for test cases, for implementing recurring patterns in code and for spotting contradictions between two documents.

What they do not replace: the decision about which figure is correct. For the market research behind this website I opened 24 provider sites individually and took the prices from the pages themselves, not from a summary. A model would have named plausible amounts, and some of them would have been wrong. The difference between usable and worthless sits exactly at this point.

What can a generator really not do?

Four things that have nothing to do with text quality. First: take responsibility. If the privacy policy states a legal basis incorrectly, the operator is liable, not the tool. Second: make decisions against the average. A model produces what was frequent in its data, and that is by definition what everybody else has too.

Third: know the limits of your own offering. My service pages state what is not included; boundaries like that come out of experience with projects that went wrong. Fourth: the operation afterwards. A website is not finished on the day it is published: it needs updates, measurement and somebody who is reachable. What that costs is covered in What a website in Carinthia costs in 2026.

How do I check my own work?

With automated tests that stop a change before it goes live. Axe tests run against the accessibility criteria on every page, once in light mode and once in dark mode, currently 136 runs across 68 pages. On top of that come tests for load behavior, forms, legal texts and the structure of the structured data.

For the blog there is an editorial test of its own: every article has to be between 900 and 2,500 words, name at least one figure of its own and at least two sources with a valid address, carry a creation and a modification date, and open under every subheading with a complete answer sentence. If one point fails, the test run fails and the article does not go live. That is the technical implementation of an editorial rule, and the reason why no thin article can sit here.

What does that mean for your decision?

That the question is not with or without AI but with or without someone responsible. A website built with tool support by somebody who checks the figures, knows the legal bases and takes on the operation beats a hand typed page without checking, and an unsupervised generated one by a wide margin.

In practice, for choosing a provider, that means: have them show you what on the site is verifiable. Are prices there? Are limits there? Is it clear who is liable? If those three things are missing, it is secondary whether the text came from a human or a model: either way it says nothing.

How do you start?

With a look at your existing site and the question of which sentence on it applies only to you. If you find none, you have found the actual problem, and it can be solved without a rebuild: add figures, name limits, describe processes.

What the technology delivers is told to you by the website check in under a minute: title, description, structured data, mandatory disclosures, load time. What a rebuild costs and how long it takes is on the websites page, with no inquiry needed.