"Semantic search" gets used as a synonym for "Google is smarter now." There is a concrete mechanism underneath, and knowing it changes how you write.
The mechanism
An embedding model converts text into a vector — a list of numbers, typically several hundred to a few thousand dimensions. The models are trained so that texts with similar meanings produce vectors that sit near each other in that space, under a distance measure like cosine similarity.
The consequence is that "how much does a crown cost" and "dental crown pricing" land close together despite sharing almost no words, while "crown" in a monarchy context lands somewhere else entirely. Meaning, not string overlap, determines proximity.
Modern search systems use this at multiple points: query understanding, candidate retrieval, and re-ranking. Google's public descriptions of BERT, MUM, and their neural matching work all describe variations on this idea.
What it broke
Keyword density. Repeating a phrase does not move a vector closer to the query's vector in any useful way. It changes the text's statistics without changing its meaning, and embedding models are trained on meaning. The practice was always crude; it is now close to inert.
Exact-match obsession. Optimizing for "dental crown cost" and separately for "how much is a dental crown" as if they were different targets is largely wasted work. They embed to nearly the same place. One well-written section answers both.
Thin pages per keyword variant. The old strategy of one page per phrasing produces a cluster of near-identical documents that a semantic system correctly recognizes as redundant. It also splits whatever authority the topic has earned across several weak pages instead of concentrating it in one strong one.
What it rewards
Actual coverage of the concept. Because embeddings capture meaning, content that genuinely addresses the surrounding concepts — materials, timelines, insurance, alternatives, risks — embeds nearer to the full range of queries in that space. This is not "include related keywords." It is "actually discuss the related things," which produces the related terminology as a byproduct.
Distinct pages for distinct meanings. If two pages embed to nearly the same point, they compete with each other for no gain. The test is not "do these have different keywords" but "do these answer meaningfully different questions." Consolidating near-duplicates is one of the highest-return content operations available to most sites.
Clear, unhedged sentences. Vague writing produces vague embeddings. A sentence that commits to a specific claim occupies a specific location in the space; a sentence that gestures at several possibilities is diffuse.
Passage-level coherence. Retrieval systems embed chunks, not whole documents. A section that stays on one idea produces a clean vector. A section that wanders produces a muddy average that matches nothing well. This is the single most actionable consequence, and it argues for tight, single-purpose sections under descriptive headings.
Where lexical matching still matters
Semantic retrieval did not eliminate exact matching, and most production systems run hybrid retrieval — a lexical index like BM25 alongside a vector index — because each catches what the other misses.
Lexical matching remains important for:
- Named entities. Product names, model numbers, people, drug names, statute citations. If someone searches a specific SKU, you need that exact string on the page. Embeddings are unreliable on rare proper nouns.
- Jargon and acronyms. Field-specific terms that the embedding model saw rarely in training.
- Very short queries where there is not enough text to embed meaningfully.
Which means the practical instruction is not "stop using keywords." It is: use the terms your audience actually uses, once, naturally, in a document that genuinely covers the concept — and stop counting.
A useful test
Take a page and write down, in one sentence, the question it answers. Then take each section and do the same. If any section's sentence is nearly identical to another's, merge them. If a section's sentence is not a real question anyone asks, cut it. If the page's sentence is nearly identical to another page's, consolidate the pages.
That exercise approximates what an embedding-based system does when it evaluates your content, and it is available to you without any tooling at all.