Semantic Search.

Early search engines worked lexically: they looked for documents in which the words entered appeared. Anyone who typed 'Velo flicken' (Swiss German for 'fix bike') found pages with exactly these words, but not necessarily the page of a workshop offering 'Fahrradreparatur' (bicycle repair). Semantic search solves this problem by modelling meaning: it knows that 'Velo' and 'Fahrrad' mean the same thing and that someone searching for 'Velo flicken Zürich' probably wants a workshop.

Google made this shift in several steps. The Knowledge Graph (2012) maps entities such as people, places and companies and their relationships. Hummingbird (2013) oriented search more towards whole queries rather than individual words. RankBrain (2015) and BERT (2019) brought machine learning and language models into the interpretation of search queries. Today, search systems work with vector embeddings, which represent texts as points in a space of meaning in which similar content lies close together.

This has clear consequences for content. It is not enough to repeat a term often; what matters is that a text treats a topic fully and coherently. A page about heat pumps that sensibly connects terms such as seasonal performance factor, ground probe, subsidies and permits signals expertise, even if the exact search query does not appear word for word.

AI search systems such as AI Overviews, ChatGPT or Perplexity are consistently semantic. They break questions down into sub-questions, find matching passages through similarity of meaning and assemble an answer from them. Clear entities, unambiguous statements and well-delimited sections help to be correctly recognised and cited in the process.