Definition
An embedding is a list of numbers that represents the meaning of a piece of text, placed so that texts with similar meanings sit close together. Unlike a keyword index, which matches strings, an embedding lets a system match ideas. Retrieval in AI search runs on this comparison.
At a glance
- CategoriesTechnical
- Related fieldsAI Search, Technical
- DifficultyAdvanced
At a glance
- An embedding turns text into coordinates, and closeness between coordinates stands in for similar meaning.
- The same method handles images, audio and product records, not only prose.
- Two different models give different coordinates for the same sentence.
- Nothing in an embedding is human-readable, which is why retrieval failures are hard to debug.
Turning a sentence into coordinates
The text is first split into tokens, which is where tokenization does its job, and the model reads those tokens together rather than one at a time. That is the point: it is how a model learns that “bank” in a sentence about rivers is not the “bank” in a sentence about mortgages, an idea the BERT paper made standard.
The output is a fixed-length list of numbers. Every passage on your site reduces to one, gets stored, and is compared against the list produced from a user’s question. OpenAI’s embeddings guide describes the same setup: embed the documents once, embed the query at search time, return the nearest.
Nearest is measured mathematically, not semantically. The system has no idea what your page says.
Why your synonym problem went away and a new one arrived
Keyword matching punished you for using different words to your customers. Embeddings mostly fix that: a page about “trainers” can be retrieved for a question about “running shoes” without the word appearing.
The new problem is the opposite. Because matching runs on overall meaning, a page vaguely about the right topic scores reasonably on everything and wins nothing. Specificity separates passages now, not vocabulary coverage.
There is a second catch. An embedding is made per chunk, so a chunk mixing three subjects lands between all three and near none. Mixed sections are the quiet reason good content fails retrieval.
Writing text that lands in the right neighbourhood
You never touch an embedding. You control the text that becomes one.
- One subject per section. A section covering sizing, returns and materials embeds as an average of three things and matches none well.
- Say the thing plainly in the opening line. The clearer a chunk’s topic, the tighter its coordinates.
- Keep your buyers’ words in the text. Synonyms are handled well, but the exact phrase still helps, and hybrid systems run a keyword pass alongside.
- Give each section a heading that names its subject. Headings often anchor where a chunk begins.
Frequently asked questions
-
Is an embedding the same as vector search?
No. An embedding is the representation. Vector search is the act of comparing embeddings to find the closest matches. You create embeddings once and store them; you run vector search every time somebody asks a question. One is the data, the other is the lookup.
-
Do I need to create embeddings for my own website?
Not for AI search visibility. Assistants embed your content themselves when they retrieve it. You would create your own to build something, such as a site search or a support bot over your docs. For getting cited, your job stops at writing clear text.
-
Why does a page about the right topic still miss?
Usually chunking. Semantic search matches a chunk, not a page, and a chunk that wanders across subjects has blurred coordinates. Split it, give each part a heading and an opening claim, and the same words start matching.
-
Is understanding embeddings actually useful for a marketer?
Up to a point, and past it this is a distraction. Knowing that matching happens on meaning, per chunk, changes how you structure a page. Knowing the dimensionality of a particular model changes nothing you do. Learn the shape of it and stop there.