AI website search uses machine learning — usually embeddings, and often a language model — to interpret what a visitor means rather than matching the characters they typed. It makes descriptive, misspelled and conversational queries work, and can answer factual questions directly instead of returning a list of links.
The term covers several distinct mechanisms, and vendors rarely separate them. Semantic retrieval embeds queries and content into a shared vector space so meaning-similar items are retrieved. Query understanding uses models to correct, expand and classify what was typed. Answer generation grounds a language model in retrieved documents to reply in prose. Learned ranking uses behavioural data to reorder results.
A product can do any one of these and be described as AI search. Which ones it does determines what it fixes.
Queries that previously returned nothing start working: descriptions instead of names, problems instead of products, questions instead of keywords. The search box stops being a filter over titles and starts behaving like a way to ask the site something.
Missing or wrong content. If a policy is not published, or a product lacks the attributes a shopper filters on, no model will retrieve it. AI search amplifies the quality of the underlying content — in both directions.