How an embedding works
A model reads a piece of text — a sentence, a paragraph, a product description — and outputs a long list of numbers that represents what the text is about. Texts with similar meaning produce similar lists, so the machine can measure how close two pieces of writing are without either one containing the other’s words. “Cheap flights to Kathmandu” and “low-cost airfare to Nepal’s capital” land near each other; “flight simulator review” does not.
That closeness is all an embedding gives you. It carries no facts, no dates and no truth, and it cannot tell you whether the text it represents is accurate. It is a position, and everything built on top of it — semantic search, recommendation, retrieval for AI answers — is some form of asking which stored positions sit nearest to the question.
Why embeddings matter
They are the reason exact-match keyword thinking has faded. When a system compares meaning rather than strings, repeating a phrase gains you nothing, while covering the actual concept clearly gains you a lot. Related vocabulary, the questions around a topic and the specifics of your offer all shift the position of your text towards the questions you want to be found for.
They are also the layer underneath most AI retrieval. A passage from your site is embedded, stored, and later compared against a user’s prompt; if it sits close enough, it is fetched and can be quoted. This is what semantic search means in practice, and it is why writing plainly about one subject beats writing loosely around several.
Where embeddings go wrong
The first misunderstanding is treating them as a store of knowledge. They are not a database of facts, and asking an embedding whether something is true is a category error — that is why retrieval systems keep the original text beside the numbers and quote from the text, not from the position.
The second is vagueness. Marketing prose that avoids naming what it sells embeds towards a general fog rather than a specific topic, so it matches nothing sharply. Mixing several unrelated subjects into one passage does the same damage from the other direction: the position lands between the topics and near none of them. Stale content adds a quieter failure — an embedding reflects the text as it was when it was created, so a rewritten page that is never re-processed keeps competing on its old meaning.
How to act on it
For a website, nothing exotic is required. Say plainly what a page is about, in the words your customers use, near the top. Keep one subject per section and name it rather than referring back to “it”. Cover the specifics — the service, the place, the constraint, the price model — because specifics are what pull a passage towards a real question rather than a general theme.
If you are building something of your own, such as an internal assistant over your documents, then embeddings become an operational concern: choose one model and stay with it, since positions from different models are not comparable, re-embed whenever a document changes, and store the source and date beside each entry. That store is a vector database, and its usefulness rises and falls with how clean the text going into it is.
