Semantic Search
Semantic search retrieves results by meaning rather than exact keyword matches, converting the query and candidate documents into embeddings (numeric vectors) and ranking by how close those vectors sit in that space, so a page can match a question even when it shares no words with it.
Older keyword search counted term overlap: a page ranked because it repeated the searched words. Semantic search instead measures conceptual similarity, so a query like "shoes that survive trail running" can surface a product described as "rugged off-road sneakers" with no shared terms. The shift happens because both the query and the document are turned into embeddings, vectors that encode meaning, and the system ranks by distance between them. This is why keyword stuffing has lost its force; padding a page with repeated phrases does little to move the vector, while clear, specific writing that states the concept plainly tends to land closer to the queries that matter.
The practical lesson is to write the way a customer would actually ask, and to define the subject in direct language near the top of the content rather than burying it. Specificity helps the embedding carry real meaning: "waterproof to 50 metres" sits closer to a swimmer’s question than "great for water", because it names a concrete attribute the model can locate. Vague, promotional phrasing tends to drift toward the centre of the space, near everything and close to nothing.
Consider a Shopify store selling cast-iron cookware. A shopper asks an assistant, "what pan can I move straight from the hob to the oven without warping?" The product page never uses those words. It does say "fully seasoned skillet, oven-safe to 260 degrees, single-piece construction with no plastic handle". Under keyword search those phrasings might miss each other entirely. Under semantic search they sit close, because oven-safe, single-piece, and no plastic handle collectively encode the concept of moving from hob to oven safely. The store wins the match by describing the attribute honestly, not by guessing the exact query. The same logic applies to reviews: a customer who writes "I left it on a 220 degree oven for an hour and it held its shape" reinforces the concept in language no marketer would script, which is part of why genuine review text is useful raw material for retrieval.
Semantic retrieval also underpins most AI answers. When ChatGPT, Perplexity, or Google AI Overviews gather sources before responding, they typically retrieve by embedding similarity rather than literal keywords, then summarise what they find. Writing that names attributes plainly is therefore easier for these systems to retrieve and quote. A caveat worth stating: similarity is not accuracy. A page can be retrieved for being topically close while still being wrong, which is one reason answer engines lean on corroboration between independent sources, and why consistent, factual product descriptions tend to get cited more often than clever ones.