Embedding

An embedding is a list of numbers that represents the meaning of a piece of text, positioned so that passages with similar meaning sit close together in the same mathematical space. Retrieval systems compare embeddings to find passages that answer a question, even when those passages share no words with it.

The term belongs in a publisher's vocabulary for one reason: it explains why matching moved from words to meaning, and therefore why plain, self-contained writing beats repeated phrasing.

Why do AI systems represent text as numbers?

Because numbers can be compared for closeness and words cannot. String matching can only tell you whether two texts share a term; a numeric representation of meaning lets a system ask how similar two passages are, and rank by the answer. That is the operation retrieval needs.

The practical consequence is the one every publisher notices eventually: a page can be retrieved for a question it never uses the wording of. "Does my policy cover a rental car" and "am I insured driving a car I do not own" are different strings and near-identical meanings. Under meaning-based matching, one passage can serve both.

How does your page become an embedding?

In pieces, not whole. A retrieval pipeline splits documents into passages, converts each passage into an embedding, and stores them for lookup. At answer time the incoming question is converted the same way and compared against the stored passages, and the closest ones are handed to the model that writes the answer. The lookup step itself — how a system finds the nearest stored embeddings — is vector search; this entry is about what the stored numbers represent.

Two things follow, and both are about the split rather than the numbers. The unit that gets represented is the passage, so a passage carrying two mixed ideas produces one blurred representation that matches both weakly. And the passage travels alone, so anything it depends on — a definition three paragraphs up, a pronoun with no antecedent inside it — is not there when it arrives. The splitting step itself is chunking, and the full pipeline from crawl to quoted answer is how AI search works.

What does the embedding layer change about how you write?

It changes what "matching" rewards. The table below is the substitution, stated as writing decisions:

None of this is a trick, and that is the point. The writing that performs well under meaning-based retrieval is writing that says what it means in the smallest self-contained unit possible — which is also what a human skimming a page wants. Our content contract encodes exactly that: one idea per section, 2–4 sentence paragraphs, the answer in the first sentence, applied across 400+ published pages on 3 production builds [our data].

What can you not control about embeddings?

Effectively all of the machinery. No major engine publishes the embedding model it uses, how many dimensions its vectors carry, how passages are split before encoding, or what similarity threshold decides retrieval. Technical treatments such as iPullRank's AI Search Manual describe the retrieval-and-relevance shape of these systems, not the parameters inside any specific one — because those are undisclosed.

Treat any tactic that claims to optimize the numbers as unproven. What is measured, and the strongest controlled result the field has, is passage content: the Princeton GEO study (KDD 2024) tested content-level edits — adding citations, quotations and statistics to the source text — and reported visibility gains of up to 40% in its benchmark. That is a benchmark result on the paper's own corpus rather than a field guarantee, and its levers all live inside the retrieved passage. Nobody can promise a citation from any of it.

How do embeddings handle your brand and product names?

Poorly, if the page never says what the name refers to. An embedding represents whatever meaning the text expresses, and a product name on its own expresses very little — it is a string the model has no reason to associate with a category, an audience, or a problem. Two companies with similar names get similar representations if neither page explains what it does.

The fix is not repetition, it is apposition: state the category next to the name, in the same passage, in the words a reader would use. "Magnitude, an agency that builds and operates authority sites" is a sentence a retrieval system can represent. "Magnitude is redefining what's possible" is not — it describes no category, so it sits near nothing useful.

This is also why entity work and embeddings are different jobs that look alike. Embeddings capture what a passage means; identity resolution — deciding that this "Magnitude" is that organization — is a separate mechanism with its own signals. Writing the category next to the name serves both, which is the rare case where one sentence does two jobs.

Do keywords still matter if retrieval is semantic?

They matter as clarity, not as counts. A passage still has to name its subject — a system cannot represent a meaning the text never states — so the terms your readers use should appear where the passage establishes what it is about. What stops paying is repetition: saying the phrase five more times does not move a meaning-based match, and it makes the passage worse to read. The fuller comparison between keyword lookup and meaning-based lookup belongs to the vector search entry in this glossary.

Retrieval is also only half the story: whether the answer model actually uses a retrieved passage, and attributes it, is a separate step, described in RAG for marketers. The writing system built on all of it — passage-first pages with their evidence tiers attached — is our generative engine optimization guide.