Vector Embedding.

An embedding model reads a passage of text and returns a list of numbers, typically several hundred to a few thousand values. That sequence is the vector. It is meaningless to a human, but it encodes the meaning of the passage in a way that lets distances be calculated. Two texts about shop opening hours land close together; a text about engine oil lands far away, even when they share not a single word. Semantic similarity thereby becomes a measurable geometric quantity.

Search in AI systems rests on exactly this. When a question is asked, it too is translated into a vector and compared with the vectors of stored passages. What comes back are the nearest sections, which then feed into the answer as context. That explains one of the biggest differences from classic SEO: exact keyword matching is no longer necessary. A passage can serve as evidence even though it nowhere contains the wording of the question, as long as it answers that question completely.

Two practical rules follow for the writing itself. First, a section should deal with exactly one topic. If it mixes three thoughts, the resulting vector sits somewhere between all three and matches no question particularly well. Second, the context belongs inside the section. A paragraph that says only 'from CHF 300 per month' loses its reference once it is extracted; 'we look after Local SEO from CHF 300 per month' stays understandable on its own and therefore stays findable.

A worked case: one page answered a question about support response times thoroughly but was never cited by assistants. The reason was structure. The figure sat in the middle of a long stretch of running text about collaboration, alongside material on onboarding and invoicing. Once that was split into three short single-topic sections, each with its own heading, the support passage became graspable as a vector of its own and started appearing in answers. Nothing about the content had changed — only the portioning.