What is Embedding?
An embedding is a long sequence of numbers, usually a few hundred to a few thousand, that a language model produces from a text. The numbers describe not the letters but the meaning: „no VAT on the invoice" and „small business VAT exemption" produce two sequences that sit close together although not a single word matches. A character comparison, which is what the built-in search of most website systems performs, sees no connection between the two at all.
Meaning becomes computable because closeness is measurable. Two sequences are compared by the angle between them, the cosine similarity. A small angle means: the same thing, said differently. Every sequence depends on the model that produced it. Sequences from two different models do not live in the same space and cannot be compared; whoever changes the model recomputes the whole corpus.
In practice, what counts is the unit that gets embedded. A whole page in one sequence blurs, because it averages ten topics; a paragraph that answers one question completely yields a sharp sequence that is found for exactly that question. The groundwork for search by meaning is therefore the same as for citable content: self-contained units of sense, unambiguously headed. The embedding is the sequence of numbers; where it is stored and searched is the job of the vector database.
Why does Embedding matter?
For its April 2026 update the Baymard Institute benchmarked more than 170 shops and apps: 12 per cent of sites have issues when users type the exact product name, 43 per cent when they describe a use case, and 66 per cent on queries that are not about a product at all. The gap sits precisely between wording and meaning, and the embedding is the tool that closes it.
Embedding in practice
- 01The query „do I have to charge VAT" is compared as a sequence of numbers with every paragraph on the site and lands on the page about the small business exemption, although none of the words appear there.
- 02A help centre embeds each answer separately rather than the whole page, so that the question about the notice period hits the one paragraph and not the page with all the contract terms.
- 03After switching the embedding model the entire corpus is recomputed, because the old and the new sequences do not sit in the same space.
