Text that sounds less like AI
Seven of thirteen generated websites in our test invited visitors in for coffee, including sites that had nothing to do with coffee. What makes a text sound like AI, what Swedish websites with good copy do instead, and the rules ANVL checks in code because a model will not follow them.
You can usually tell. A generated text has a sound to it before you have read a full sentence: a warm invitation that promises nothing, a phrase about moments to remember, a dash in every other sentence. None of it is wrong, exactly. It is just that nobody who knew the business would have written it.
We went looking for where that sound comes from in ANVL, and what to do instead.
The coffee problem
The site generator's rule file gave one example of a good call to action: Come by the workshop. We have coffee. The model took it as a recipe. Coffee turned up on 7 of 13 sites in our test set, including sites with no connection to coffee at all. The same tone came from three other examples, and from an onboarding voice sample, come and see us in…, that travels with every generation.
An example in a prompt is not a hint. It is the thing most likely to be copied.
The examples now say what happens next instead: what the visitor sends, what they get back, and when. The list of words to avoid got two new categories: false closeness, like "come by", "see you soon" or "… is waiting for you", and mood phrases, like "under one roof", "we take care of the rest" or "actually", when there are no facts behind them.
What a person writes instead
A list of what not to write only gets you so far. So we read six Swedish websites with good copy and pulled out nine patterns for what to write. Say what happens after the click. Measure instead of praising. Include one detail that is true only of this business. Make buttons a verb and an object.
The patterns are given as shapes, not as sentences to copy, and every fact in the text must come from what you gave us. A detail that is not in your material does not get invented to make the sentence work.
Two rules the model would not keep
In testing, two things were ignored no matter how clearly the rule was written: a ceiling on dashes, and using one form of address throughout. Swedish has a formal and an informal you, and a generated text would switch between them from one paragraph to the next.
So those two are measured in code. The form of address is your brand's when it is set, otherwise whichever one dominates the material, and quotes do not count. What the check finds goes to the quality pass that already runs, so it costs no extra call. Your own edits are left alone: there, the text is yours.
The same rules on every path
Email drafts had no word list at all, and one-pagers and document decks got it only afterwards. Now it is part of generation everywhere. An email draft gets a follow-up pass if the check finds something, and the correction is only used if it is better.
We ran the old and the new rules on the same four briefs. The coffee and the formulas were gone, and fewer details were made up.
Why it matters
If you are evaluating ANVL: every tool can write a paragraph. The question is whether your customers can tell it was not you. That is not solved by a cleverer model. It is solved by deciding what good text does, checking for it, and refusing to fill gaps with things that sound nice.