Grounding.

A language model produces text from statistical patterns learnt during training. It has no concept of truth and no access to events after its training cut-off. Grounding closes that gap: before answering, the system retrieves matching documents, places them into the context window and instructs the model to answer from those alone. The result is a statement that can be traced back to a specific source. Grounding is therefore the technical reason AI answers can carry source links at all, and the reason a website can be quoted rather than merely summarised from memory.

In Google's ecosystem the feature is called Grounding with Google Search. The model decides per request whether to search, pulls passages from the Google index and returns grounding metadata alongside the answer: source URLs plus the mapping of which passage supports which statement. ChatGPT with web search, Perplexity and Claude follow the same basic pattern, only with different retrieval sources. For websites this leads to an uncomfortably clear rule. If you are not inside the source that is being grounded from, you cannot be cited from it either.

Grounding also explains why small craft details suddenly carry weight. What gets retrieved and placed into context are passages, not whole pages. A statement spread across three paragraphs and diluted with marketing language survives that extraction badly. A statement that stands complete in one sentence, with subject, service and figure together, is carried over cleanly. This is why a line such as «we look after local SEO from CHF 300 per month» performs better in AI answers than a price graphic with no accompanying text underneath it.

One case from practice. A medical practice wanted to know why an assistant kept quoting the wrong opening hours. The cause was not the model but the sources it grounded from: the website, the Google Business Profile and two directories carried three different versions. After all four were brought into line and the hours were marked up as structured data, the answer was correct within a few weeks. Grounding makes inconsistencies visible that classic SEO left unpunished for years. As long as only humans read the website, an outdated entry in some directory went unnoticed by nearly everyone.