What an embedding is
An embedding is a list of numbers that stands for a piece of text. The list is produced by a model trained on a large body of text, and the training arranges things so that texts used in similar ways, about similar things, get similar lists. Each number can be thought of as a coordinate, so the text becomes a point in a space with a great many dimensions rather than the familiar two or three.
The individual coordinates do not mean anything a person could name, such as "price" or "colour". What matters is the arrangement: "waterproof walking boots" and "hiking boots that keep feet dry" land near each other, while "walking boot laces" lands somewhere else, nearer to other accessories. The embedding is the model's compressed sense of what the text is about, expressed as a location.
This is a different job from writing. A language model of the kind explained in how language models generate text produces text; an embedding model produces positions. Many AI tools use both, one to find and group material and the other to write about it.
Why similarity in meaning is useful to marketers
Marketing produces and receives a great deal of text that has to be sorted: keyword lists, search queries, reviews, survey answers, support tickets, product titles, page content. Sorting by exact words misses the point that customers describe the same need in many ways. Sorting by hand is accurate but slow, and it varies between people. Embeddings offer a middle route: a consistent, fast way to put things that mean similar things together, which a person can then check and name.
They also make search more forgiving. A help-centre search built on embeddings can match "my parcel never arrived" to an article titled "Lost or missing deliveries" even though the two share no important words. That same ability is what lets a document assistant find relevant passages, as how grounding and retrieval work explains.
Step by step: from text to groups
- Choose the units. Decide what one item is: a keyword, a sentence from a review, a whole review, a page section. Short items carry less context, so their positions are less certain.
- Embed each item. Each item is passed through the embedding model and comes back as a list of numbers. The same model must be used for everything that will be compared.
- Measure similarity. For any two items, a similarity score is calculated from how close their positions are. A common approach compares the direction of the two lists rather than raw distance, but the principle is the same: closer means more alike.
- Find neighbours. For a given item or query, the tool lists the items with the highest similarity. This is semantic search.
- Form clusters. A clustering method gathers items that sit close together into groups. The method and its settings decide how many groups there are and how tight each is.
- Name and check the groups. A person, sometimes helped by a language model, reads each group, names it and moves items that do not belong. This step is where the groups become useful.
| Step | Setting that matters | Effect of getting it wrong |
|---|---|---|
| Units | How long each item is | Very short items group by surface wording |
| Embedding | Which model, used consistently | Mixing models makes scores meaningless |
| Similarity | The measure and any cut-off | Unrelated items scored as close, or near-duplicates missed |
| Neighbours | How many results to return | Too few misses relevant items; too many adds noise |
| Clusters | Tightness or number of groups | Groups too broad to act on, or too many to read |
| Review | Who checks and against what | Machine groups accepted as decisions |
Where similarity misleads
Similarity of meaning is not the same as sameness for a marketing purpose, and the gaps are predictable.
Same topic, different intent. "Garden room prices" and "garden room planning permission" are about the same thing and sit close together, but one searcher is comparing quotes and the other is checking rules. They differ in search intent and may need different pages. The keyword clustering guide groups keywords by the overlap of their search results precisely because that reflects what a search engine treats as one need.
Different words, same need, and the reverse. Embeddings usually handle paraphrase well, which is their strength, but brand names, product codes, abbreviations and new jargon may be placed poorly if they were rare in the model's training text.
Negation and contrast. "Not suitable for winter" and "suitable for winter" share almost every word and topic, and can sit closer than their meanings warrant. Feedback grouping can put complaints and praise about the same feature together.
Bias of the training text. The model's sense of what is similar comes from the text it learned from. Usage that is common in that text dominates; regional terms, minority dialects and niche trade language can be pulled towards a more common meaning.
Thresholds are choices. Where one cluster ends is set by a parameter, not discovered. Two groups that look distinct at one setting merge at another.
When semantic groups and search-result groups disagree, the disagreement is information. Two queries that mean the same thing but return different results may justify separate pages; two that look different but share results may compete if targeted separately, which is the risk of Keyword CannibalisationTwo or more pages on one site competing for the same query and intent, splitting signals..
A garden room builder compares two ways of grouping keywords
Fenwold Garden Rooms, an invented builder of insulated garden offices, has a list of 600 search queries and uses an embedding-based tool to group them. At the tool's default setting it gets 42 clusters. The team then checks the 15 largest clusters against shared search results, using the method in the keyword clustering guide. 11 of the 15 hold up. In the other 4, meaning and results disagree: "garden office tax" and "garden office business rates" are close in meaning and return largely different results, and "garden room" and "summer house insulated" look different but return mostly the same pages. The team keeps the semantic clusters for topic mapping, splits or merges the 4 for page planning according to the search results, and notes the threshold it used so the grouping can be repeated next quarter.
Where marketers meet embeddings
Most marketers never handle an embedding directly; they meet the results inside other tools.
- Keyword clustering features in SEO tools that group by meaning rather than, or as well as, by results.
- Site search and help-centre search that match questions to articles without shared keywords.
- Document assistants that retrieve passages before writing an answer.
- Feedback and review analysis that tags open text into themes.
- Duplicate and near-duplicate detection across product titles, pages or ad copy.
- Recommendations of related articles or products based on similar descriptions.
Knowing that the same mechanism sits behind all of them explains why they share the same blind spots.
What it means for you
- Use semantic grouping to explore, and search results to decide pages. Meaning is a fast first sort; rankings decide whether one page can serve a group. The AI for keyword research guide applies this to a full keyword research workflow.
- Always review and name clusters. The output is a proposal. Moving the items that do not belong is where the judgement sits.
- Record the settings. Note the tool, model and threshold so a grouping can be repeated and compared.
- Watch the blind spots. Check groups for mixed intent, mixed sentiment, brand terms and codes before acting on them.
- Keep text clear at the source. Headings and product titles that name the thing plainly are easier to place correctly, for search engines, embedding tools and people.
Common misconceptions
- "Close in meaning means same intent." Topic and intent are different; search results are the better test of whether two queries are one need.
- "The clusters are what the data says." They are what the model and the chosen threshold say. Change either and the groups change.
- "Embeddings understand sentiment." They often place opposite opinions about the same feature close together.
- "Every tool groups the same way." Different models and settings produce different groups from the same list.
- "Semantic search does not need good content." It finds the closest text that exists; vague or outdated text is found just as readily.
