Vector databases store embeddings, which are lists of numbers that represent the meaning of text, images or products, and find items with similar meaning quickly. In marketing they power semantic site search, content recommendations, duplicate detection, internal linking suggestions and retrieval for AI assistants. Moving from a pilot to production mostly means getting data preparation, model choice, evaluation and costs right.
An embedding model, such as OpenAI’s text-embedding models, Google’s Gemini embeddings, Cohere Embed or open models from the sentence-transformers family, turns a piece of text into a vector of a few hundred to a few thousand numbers. Texts with similar meaning end up close together in that space, so “cheap flights to Antalya” and “low-cost tickets Antalya” sit near each other even though they share few words. A vector database indexes those vectors and answers “what is closest to this?” in milliseconds.
Do you need a dedicated vector database?
Not always. Options fall into three groups.
- Extensions to databases you already run: pgvector for PostgreSQL, vector search in MySQL HeatWave, MongoDB Atlas Vector Search, Redis vector search.
- Search engines with vector support: Elasticsearch, OpenSearch, and hosted search services such as Algolia.
- Dedicated vector databases: Pinecone, Weaviate, Qdrant, Milvus and Chroma.
For a pilot with tens or hundreds of thousands of items, an extension like pgvector is usually enough and avoids a new system. Dedicated databases make sense at large scale, with heavy filtering needs, or when a team wants managed hosting with vector-specific features.
Which marketing use cases are worth a pilot?
Semantic site search
Visitors type what they mean, not your product names. Embedding-based search finds relevant items for vague or misspelled queries. Combining it with keyword search, known as hybrid search, usually beats either alone.
Related content and products
“You may also like” blocks based on meaning rather than shared tags. Useful for blogs with thousands of posts where manual tagging is inconsistent.
Duplicate and cannibalisation detection
Embedding every page reveals clusters of near-identical articles targeting the same topic. Those are candidates for merging or redirecting.
Internal linking suggestions
For each page, the nearest semantic neighbours are natural link targets. An editor still chooses which links fit.
Retrieval for AI assistants
Retrieval-augmented generation depends on finding the right passages; the vector index is that retrieval layer.
Pilot checklist
- Define one use case and one success measure, such as search exit rate or click-through on related items.
- Clean the source data: remove boilerplate, navigation text, cookie notices and duplicates before embedding.
- Decide the unit: whole pages, sections or paragraphs. Long pages often need splitting into chunks.
- Choose an embedding model and test it in your languages; multilingual performance varies.
- Store metadata with each vector: URL, language, date, category, product availability.
- Build a test set of 50 to 100 real queries with the results a human expert expects.
- Compare vector, keyword and hybrid search on that test set before launch.
What goes wrong between pilot and production?
Pilots often run on a clean snapshot. Production has live data, and that is where problems appear.
- Stale vectors: content changes but embeddings do not. Re-embed on publish and on update.
- Model changes: vectors from different embedding models are not comparable. Switching models means re-embedding everything.
- Filtering: users expect results limited to their language, region or in-stock items; filters must be part of the query, not applied afterwards.
- Cost growth: embedding is cheap per item but adds up with frequent re-indexing; storage and query costs scale with volume.
- Multilingual mixing: a Turkish query matching English pages may be technically right but unhelpful; filter by language unless cross-language results are intended.
Production checklist
- Automate embedding on content create, update and delete events.
- Version the embedding model and store the version with each vector.
- Monitor query latency and result quality with the test set on a schedule.
- Log zero-result and low-similarity queries; they reveal missing content.
- Set access control if the index contains anything not public.
- Document the rebuild procedure so a model switch is a planned task, not an emergency.
How Alien Road applies vector databases and embeddings
In a typical engagement our AI optimisation team uses embeddings first for content analysis, then for user-facing features. Discovery begins with a crawl of the client’s site and an inventory of content and product data, including languages and update frequency.
Setup usually starts on the client’s existing database with a vector extension, which keeps infrastructure simple. We clean and chunk content, embed it, and run the first analyses: duplicate clusters, orphan pages without internal links, and topics with thin coverage. These findings feed SEO and content plans directly.
For user-facing search or recommendations, we build the query test set with the client’s team and compare approaches before switching anything on. Measurement tracks search exit rates, clicks on recommendations and conversions from search users. Monthly reports include the top zero-result queries as a content to-do list.
How do you evaluate result quality without a data science team?
Keep it simple and repeatable. For each test query, list the three to five results an expert would expect. Run the query, and record how many of those appear in the top five results. Average across queries, and compare approaches on that single number. Rerun after every change to chunking, model or ranking.
Add a sanity check by reading results. Numbers can hide odd behaviour, such as one very long page appearing for every query because its embedding sits near the centre of everything.
How much does a vector setup cost?
For a site with a few thousand pages, costs are modest. Embedding the full content once through a commercial API usually costs little, and storage in an existing PostgreSQL database with pgvector adds almost nothing. Costs grow with volume, re-indexing frequency and query traffic, and managed vector databases charge by storage, compute or queries depending on the plan. Estimate with your own numbers: items, average length, update rate and expected searches per day, then check current pricing pages.
Where to start this week
Export the titles and first paragraphs of your 500 most visited pages. Embed them with any mainstream embedding model, using a short script or a notebook, and find the pairs with the highest similarity. Read the top twenty pairs. If several are near-duplicate articles competing for the same search query, you have a concrete content cleanup task and a working proof of what embeddings can show.
Keep reading: AI Optimization · AI Optimization