Vector databases store embeddings, which are lists of numbers that describe the meaning of a piece of text, an image or a product. Marketing teams use them to find similar content, match customer questions to the right answer and feed accurate context into AI assistants. The practical value is search by meaning instead of search by exact words.
What is an embedding, in plain terms?
An embedding model reads a sentence and returns a long vector, often several hundred or a few thousand numbers. Two sentences with similar meaning end up close together in that space, even when they share no words. “Cheap running shoes” and “affordable trainers for jogging” land near each other.
Several providers offer embedding models through an API, including OpenAI, Google, Cohere and open-source options such as the sentence-transformers family. The model choice matters less than consistency. If you embed your catalogue with one model, you must embed every query with the same model, or the distances stop meaning anything.
Where do vector databases fit in a marketing stack?
A vector database is the storage and search layer for those numbers. Pinecone, Weaviate, Qdrant and Milvus are dedicated products. PostgreSQL with the pgvector extension, Elasticsearch and OpenSearch also support vector search, which is often enough for a mid-sized site.
The database answers one question very quickly: which stored items are nearest to this new item? Everything useful in marketing is built on top of that single operation.
- Content deduplication: find blog posts that cover the same topic before they compete for the same query.
- Internal linking: suggest the five most related articles for each page.
- Product recommendations: show items that are similar in description, not only in category.
- Support and chat: retrieve the right help article before a chatbot writes an answer.
- Audience research: cluster thousands of reviews or search queries into themes.
How is this different from keyword tools?
Keyword tools count strings. They are still the right source for search volume and competition, because search engines report demand by query. Embeddings do not replace that data; they organise it.
A typical example: a keyword export with 8,000 rows is hard to read. Embedding each query and clustering the vectors produces perhaps 60 groups, each one a candidate page or section. A strategist still names the groups and decides priorities. The machine only does the sorting that used to take days in a spreadsheet.
How Alien Road applies vector databases and embeddings
In a typical engagement the work follows four steps, and each step has a written output the client can check.
Discovery
We start with an inventory: every URL, its title, main heading and body text, plus Search Console queries for the last few months. The question at this stage is simple. Where does the client suffer from overlap, thin coverage or poor linking? Embeddings answer it faster than manual reading, but a person reviews the flagged pairs.
Setup
For most sites we use pgvector inside an existing PostgreSQL instance or a managed service such as Qdrant Cloud. Text is split into chunks of a few hundred words, embedded once, and stored with the URL and language. A small script re-embeds only pages whose content hash has changed, so costs stay predictable.
Measurement
The first report lists near-duplicate page pairs above a similarity threshold that we tune per site, since the right cut-off depends on the model and content style. It also lists orphan pages and suggested internal links. Results are compared before and after changes using Search Console clicks and impressions for the affected URLs.
Reporting
Clients receive a sheet with the decisions: merge, redirect, rewrite or keep. We avoid presenting similarity scores as a ranking factor. They are a diagnostic, and the report says so.
What are the limits?
Embeddings capture meaning roughly, not perfectly. Negation is a known weak spot: “hotels that allow pets” and “hotels that do not allow pets” can sit close together. Short texts such as product titles carry little signal on their own, so it helps to embed the title together with the description and key attributes.
Language is another issue. Multilingual models handle Turkish, German and English in one space, but quality varies by language pair. Test with a sample of known pairs before trusting the output across a full catalogue.
Finally, a vector database is not a source of truth. If the underlying page is wrong, the retrieval step will confidently return wrong information. Content quality comes first.
What does a sensible first project cost in effort?
For a site with a few thousand pages, the technical setup is usually measured in days, not months. Embedding costs through commercial APIs are low for this volume at the time of writing, though prices change, so check the provider’s current page. The larger cost is human review of the clusters and duplicate pairs.
A useful rule: pick one decision the project must inform, such as which 50 posts to consolidate. Projects that try to “add AI search to everything” tend to stall because nobody owns the output.
Common mistakes to avoid
- Mixing embedding models between the stored data and the queries.
- Embedding whole pages as one vector, which blurs several topics into one point.
- Skipping metadata such as language, date and content type, which you will need for filtering.
- Treating the similarity score as an absolute truth instead of a ranked shortlist.
- Forgetting to re-embed after content edits, so the index slowly drifts from the live site.
Where to start this week
Export your URLs and body text, pick one open-source embedding model, and load the vectors into pgvector on a test database. Run a nearest-neighbour query for your ten most important pages and read the results. If the neighbours make sense to a human editor, you have the foundation for duplicate checks, internal linking and a retrieval layer for any assistant you build later.
Keep the first output small and reviewable. One clean list of overlapping posts, acted on, is worth more than a dashboard nobody opens.
Keep reading: AI Optimization · AI Optimization