Vector embedding
A vector embedding is a list of numbers that represents the meaning of a piece of text, an image, or a product record, positioned so that semantically similar items sit close together in the same space. Embeddings are what make semantic retrieval possible: a query and a product description can be compared numerically even when they share no words.
What an embedding actually is
A model reads a span of text and outputs a fixed-length array of floating-point numbers, typically several hundred to a few thousand of them. The numbers have no individually interpretable meaning. What matters is relative position: text about corrosion-resistant fasteners lands near text about 316 stainless hex bolts, and far from text about invoicing.
Similarity is then arithmetic, usually cosine similarity between two vectors. That is the entire mechanism behind "it understood what I meant" in semantic search.
Why embeddings punish thin records
An embedding can only encode what the text says. A record reading VLV-0500-BR-NPT — Ball Valve embeds to a nearly generic point in valve space. It has no bore size, no material, no connection type, no pressure rating, so it sits at roughly equal distance from every query about valves and wins none of them.
A record that spells out brass body, 1/2 inch nominal, NPT threaded, 600 WOG, full port, lands somewhere specific. It will be retrieved for the queries that describe it and correctly ignored for the ones that do not. Precision in the source text becomes precision in the vector.
This is the concrete, mechanical reason fill rate shows up in AI visibility. It is not a proxy metric. It is the input.
Where embeddings are weak
Two failure modes matter for technical catalogs.
Numbers and units embed badly. Embedding models capture meaning, not magnitude. "22 kAIC" and "10 kAIC" sit close together in vector space because they are the same kind of statement, which means pure vector retrieval will happily return a part that fails the numeric constraint. Serious product search combines vector retrieval with structured filters on typed numeric fields, a hybrid approach — and the structured filter is only possible if the value exists as a number in a field, not as text in a paragraph.
Part numbers embed badly too. HB-150-50K-U carries no semantic content. Exact and fuzzy lexical matching handles identifiers; embeddings do not. Any catalog search that dropped keyword matching entirely in favor of vectors got worse at the single most common B2B query type.
Embeddings of what, exactly
Multiple embeddings of the same SKU usually coexist: your own site search index, an AI engine's index of your public page, a marketplace's index of your feed, and a buyer's internal procurement tool indexing whatever you syndicated to them.
Each is built from a different text. If your feed carries a truncated description while your PDP carries the full spec block, those two embeddings represent different products in practice. Keeping one enriched golden record upstream of every channel is what stops those representations from drifting apart.
Frequently asked questions
What is a vector embedding in simple terms?
A way of turning text into coordinates. Similar meanings get nearby coordinates, so a system can measure how related two pieces of text are by measuring the distance between their number lists, even when the wording is completely different.
Do embeddings understand numbers and specifications?
Poorly. Embedding models encode meaning rather than magnitude, so "22 kAIC" and "10 kAIC" land close together. Numeric constraints need structured filters on typed fields; vector similarity alone will return parts that fail the spec.
How do embeddings relate to semantic search and RAG?
Embeddings are the representation, semantic search is the retrieval step that compares them, and RAG is that retrieval feeding a language model which then writes an answer. All three depend on the same thing: whether your product text contains the fact in the first place.
Can we control how AI engines embed our products?
Not directly. You control the text they embed. Complete, specific, unit-bearing product descriptions produce distinctive vectors; sparse boilerplate produces generic ones that never win a query.