Definition
An embedding is a list of numbers that stands for a piece of text. It is produced by a model trained on a large body of text so that texts used in similar ways, about similar things, receive similar lists. Each number works like a coordinate, which places the text as a point in a space with a great many dimensions. Individual coordinates mean nothing a person could name; only the arrangement matters, with closeness standing for similarity of meaning.
The guide to how embeddings group similar text draws the key distinction: a language model produces text, while an embedding model produces positions. Many AI tools use both, one to find and group material and the other to write about it.
In practice: retrieval, clustering and similarity search
Three uses account for most of what marketers meet.
- Retrieval. A document assistant embeds every passage in its library and, for each question, fetches the passages whose embeddings sit nearest the question's. This is the indexing step behind grounded answers.
- Clustering. Keywords, reviews, survey answers or support tickets are embedded and gathered into groups of near neighbours, which a person then reads, names and corrects.
- Similarity search and duplicates. Finding the help article closest to a customer's wording, or pages and product titles that say nearly the same thing.
The working rules are few. Compare only embeddings made by the same model. Decide what one item is before embedding it, because very short items carry little context. And treat cluster boundaries as settings rather than discoveries: two groups that look distinct at one threshold merge at another.
Where embedding similarity and search results disagree
The guide to keyword clustering lists embedding similarity as one way to group queries by meaning: fast and cheap, and prone to grouping queries a search engine treats differently or separating ones it treats as the same. Its method therefore uses a semantic pass to propose groups and a check of results-page overlap to confirm or split them, with intent as a hard boundary.
The disagreements are predictable. Queries on the same topic with different purposes, such as prices against planning rules, sit close together. Negation is weak, so "suitable for winter" and "not suitable for winter" can look nearly identical. Brand names, product codes and new jargon may be placed poorly if they were rare in the model's training text. The AI marketing hub places embeddings among the mechanics worth understanding precisely because their output looks more certain than it is.
A stationery shop checks its embedding clusters against search results
Nibquarry Stationery embedded 1,800 search queries from its Search Console data and grouped them into 140 clusters. One cluster joined "fountain pen ink" and "ink for calligraphy", which sat very close together, and the shop's first plan was a single combined ink page. Before mapping, the marketer compared the top 10 results for each query: they shared only 2 URLs. The fountain pen results were bottle ink shops and colour comparisons; the calligraphy results were dip-pen inks, tutorials and starter kits. The shop kept two pages, one per need, and split 11 other clusters the same way after the same check. Merging them would have put two different needs on one page.
Related terms
A Language ModelAn AI system trained on large amounts of text that writes by predicting which piece of text is likely to come next. writes text, while an embedding model places it; tools often pair the two. GroundingMaking a language model answer from specified material, such as documents retrieved for the question, rather than from what it absorbed in training. commonly relies on embeddings to find the passages a model answers from. Search IntentThe goal behind a search query: to learn, to find a site, to compare or to act. is the distinction embeddings handle least reliably, which is why intent labels act as a hard boundary when clustering keywords. Keyword CannibalisationTwo or more pages on one site competing for the same query and intent, splitting signals. is the risk on the other side: if near-identical queries are wrongly split onto separate pages, those pages compete for the same results.
Common confusions
- Embeddings understand meaning. They encode patterns of usage; closeness reflects how texts are typically used, including the biases of the training text.
- Close embeddings mean one page. Search engines decide what counts as one need; the results page is the check.
- Any two embeddings can be compared. Only embeddings from the same model share a space.
- Clusters are found, not chosen. The method and its threshold decide how many groups appear.
- Embedding and embedding a report are related. The term here is the AI representation of text, not placing a dashboard inside another page.
