Semantic search interprets the meaning of a query rather than matching its characters, so results can be relevant without sharing any words with what was typed.
It is typically implemented with embeddings and vector retrieval, sometimes with a language model performing query rewriting or classification first. The visible difference for a user is that descriptive, conversational and problem-shaped queries begin to work.
Semantic search is not strictly better than keyword search; it is better at different things. It excels at paraphrase, description and vague intent, and underperforms on SKUs, part numbers, exact titles and negation. Users notice the second category immediately when it breaks.
The mature position, now standard in serious systems, is hybrid retrieval: run both, fuse the result sets, and let a re-ranking stage settle the order.