AI optimization for websites in South Tyrol: what it means and what it doesn't

A craftsman in the Puster Valley asked me last year what he was doing wrong. His website was new, fast, well designed. On Google he ranked on page one for his village name. And yet: when someone asked ChatGPT who restores windows in the valley, his name did not come up. A company in Bruneck did — with a site that was older and slower.

This has nothing to do with design and little to do with speed. It has to do with what a machine finds on a page when it reads it — and what it doesn't.

What AI optimization actually means

The term is being sold on every corner right now, so let's start with the distinction.

Search engine optimization gets your page high up in a list of results. The person does the clicking.

AI optimization makes sure a language model can state the facts about your business correctly when asked. There is no list, no click, no second chance. Either the machine has the facts, or it doesn't.

The difference sounds academic but is practical. To make a list, it's enough that your page is roughly on topic. To build an answer, a machine needs verifiable specifics: what exactly you do, where, for whom, since when, in which languages, under what conditions. If that isn't stated explicitly somewhere, it gets guessed — or left out.

The four things it almost always comes down to

We have looked at a fair number of South Tyrolean websites through this lens over the past months. The findings repeat.

First: the facts exist only in running text. "Family-run for over 30 years" reads fine to a human. A machine needs a founding year in a place where it expects one — otherwise the answer will later say "for several decades", or nothing at all. Structured data exists precisely for this: the same information again, in a form that requires no interpretation.

Second: the details contradict each other. The phone number on the website is formatted differently than the one in the Google business profile; the address in the legal notice differs from the directory listing. To a person these are trivia. To a machine they are evidence that one of the sources is wrong — and it cannot know which.

Third: the opening hours live inside an image. Sounds trivial, happens constantly. What sits on the page as a graphic does not exist for a language model.

Fourth: there is no clear statement of what you don't do. This gets overlooked. If a machine cannot tell where your service ends, it will recommend you for enquiries you have to turn down. That costs you time and the asker trust.

Retrofitted or built in

All four points can be fixed after the fact. The question is not whether it is possible, but how often you will have to do it again.

Retrofitting means: someone goes through the pages, enters the markup, aligns the directories. After that the state is good — until the opening hours change, a service is added, or the site is rebuilt. Then it starts over, usually in several places at once, and usually nobody notices because nothing breaks. It simply becomes quietly inaccurate again.

Built in means: the system generates the machine-readable form from the same data it builds the page from. Whoever changes the opening hours in the editing area also changes what a machine reads — without a second step, because both come from the same source.

At Berger+Team we build both: the btlabs Core platform, where this runs by default, and the work on websites built differently. The honest sentence is this: for a small, well-kept site, retrofitting is perfectly fine. Beyond a certain number of pages, languages and changes per year it becomes a task nobody reliably completes — and then the construction method is the real answer.

Why South Tyrol is a special case

Three languages here are not a bonus feature, they are daily life. And that is exactly where a problem arises that Hamburg or Vienna doesn't have.

Most multilingual websites are really monolingual websites with a translation bolted on. The German version is complete, the Italian one slightly shorter, the English one an excerpt. Visitors barely notice. To a machine these are three differently well-informed sources about the same business — and which one it draws on also depends on the language of the question.

Someone asking in Italian who gets a thin answer, while the German version would have had everything, loses you an enquiry at the language border. In a region where roughly a quarter of the population speaks Italian, that is not an edge case.

The point isn't that every language needs the same words. It's that every language needs the same facts.

This too is ultimately a question of construction. Where each language version is a separate page someone has to maintain, the third one falls behind — that is not carelessness but workload, repeating with every change. Where languages sit as equals in the foundation, the core data is the same in every language — and a coverage indicator shows where text is still missing.

What AI optimization does not deliver

A lot is being promised right now, so here is the honest part.

Nobody can guarantee that ChatGPT will recommend you. There is no submission, no ranking, no lever to pull. Language models are trained at different times, access the web differently, and weight sources differently. Anyone promising you a placement is selling something they do not control.

What you can influence is the starting position: that the facts about your business are present, unambiguous, free of contradiction, and equally complete in all three languages. That is the work. Whether it translates into a recommendation also depends on things outside your website — above all on whether others write about you.

Second, it takes time. Before a model knows about a change without web search, months can pass; systems with web search pick it up sooner. Anyone promising fast results either means something else or hasn't measured.

Where to start

Not with a tool, but with an inventory. Three questions are enough to begin:

1. Are the hard facts machine-readable anywhere? Address, opening hours, services, founding year — not just as text, but marked up.
2. Do all sources say the same thing? Website, Google profile, directories, social profiles. Literally the same, not roughly.
3. Are the second and third languages equivalent? Not in length — in facts.

If you can answer yes to all three without checking, you have the foundation. Then comes the harder question of what points to you from outside — but that is another article.

How to check whether it works we have described elsewhere. And why a machine reads literally rather than by meaning is here.

The Puster Valley craftsman's page now states what he does — in a place where a machine goes looking. No dial puts him into the answers; what is now in place is the precondition.

For all three questions, one more is worth asking: does someone have to keep this up by hand, or does the system handle it? The first answer is work, the second is a setting — and the difference shows not today, but the fourth time round.

The questions that come up most often in first conversations — cost, timing, guarantees — we have collected here: Frequently asked questions about AI optimization. Where this leads when the interface itself disappears: when the interface disappears.