What Is an Embedding? Definition and Example

An embedding is a list of numbers (a vector) that represents a piece of text so that texts with similar meanings end up close together. OpenAI's documentation defines it as "a vector (list) of floating point numbers" and adds that "the distance between two vectors measures their relatedness" (OpenAI).

On this page

• How does an embedding work?
• What's an example?
• Why does it matter for AI visibility?
• What are the common misconceptions?
• From MentionX data: what we can and can't see
• Worked example: Ledgerline's duplicate-invoice passage
• How are embeddings different from keywords?
• What should marketers actually do?
• Common questions
• Related
• Sources

Checked on 31 October 2026. "Our view" marks MentionX opinion. Ledgerline is a fictional company used for illustration.

How does an embedding work?

An embedding model reads text and outputs a fixed-length vector. OpenAI's text-embedding-3-small returns 1,536 numbers by default, and text-embedding-3-large returns 3,072 (OpenAI). In OpenAI's words, "small distances suggest high relatedness and large distances suggest low relatedness". Its text search example ranks documents by "the cosine similarity between the embedding vectors of the query and each document" (OpenAI).

What's an example?

Fictional example: a user asks, "Which running shoes are good for flat feet?" A retrieval system embeds the question and compares it with embeddings of stored passages. A passage titled "Stability shoes for overpronation" can rank highly even though it never uses the phrase "flat feet", because the meanings sit close together.

Why does it matter for AI visibility?

Many AI search and retrieval-augmented generation systems use embeddings to pick passages before the model writes an answer. Engines rarely say exactly how they retrieve. Our view: passages that answer one question clearly, under a descriptive heading, are easier to match than long pages that mix topics. That's one reason content chunking matters.

What are the common misconceptions?

• "Embeddings are keywords." They aren't. They capture meaning, so synonyms can match.
• "You can optimise embeddings directly." You can't see or edit an engine's embeddings. You can only write clearer passages.
• "All engines use the same model." Models and dimensions differ from system to system.

From MentionX data: what we can and can't see

Method. Read-only aggregate queries over MentionX scans from 22 August to 29 October 2026 (UTC): 2,217 engine runs (2,154 completed) across 18 anonymised projects and six engines.

MentionX records the answers engines give and the URLs they cite. It doesn't see inside an engine's retrieval system, so we can't measure embeddings or say which passage was matched.

What we can see is which kinds of page end up cited. Ordinary articles were cited most often of the page types we could identify, which fits the idea that clear, topic-focused passages are easy to retrieve.

Our view: that's indirect evidence at best. Use it as a reason to write focused sections, not as proof of how any engine's embeddings work.

Limits. A convenience sample of mostly B2B software projects, run through engine APIs, which can behave differently from the consumer apps. Perplexity, Copilot and DeepSeek figures come from a single day (26 October 2026), so treat them as a snapshot, not a trend.

Worked example: Ledgerline's duplicate-invoice passage

Ledgerline is a fictional accounts-payable tool. Figures are illustrative.

Ledgerline's help centre has a page called "Payment exceptions" that mixes duplicate invoices, late fees and currency errors in one long block. A buyer asks an AI assistant, "How do I stop paying the same supplier invoice twice?"

The page contains the answer, but the passage that would be retrieved is about three topics at once, so its meaning is blurred. Ledgerline splits it into three sections. The new one, "How does Ledgerline catch duplicate invoices?", answers in two sentences and then gives the steps. Its meaning now sits much closer to the buyer's question.

How are embeddings different from keywords?

A keyword system looks for the same words in the query and the document. An embedding system compares meaning, so "stop paying an invoice twice" can match a passage about "duplicate payment detection". Many search systems combine both approaches, which is why plain, specific wording still helps: it serves keyword matching and gives the embedding a clear meaning to capture.

What should marketers actually do?

• One idea per section. A passage that answers one question has a clearer meaning than one that covers five.
• Descriptive headings. Headings that name the question give both readers and retrieval systems useful context.
• Use the words buyers use. Embeddings handle synonyms, but the closer your wording is to real questions, the better.
• Avoid synonym stuffing. Piling in near-identical terms muddies meaning rather than sharpening it.

Common questions

Can I check how my page is embedded? Not inside an engine. You can experiment with public embedding models, but results won't match any specific engine.

Do embeddings replace SEO? No. They're one part of how some systems retrieve text. Crawlability, clear content and authority still matter.

Related

• Vector search
• Retrieval-augmented generation
• Content chunking

Sources

• OpenAI, Vector embeddings guide (accessed 31 Oct 2026)