Vector database for SEO: what it does and when you need one

A vector database turns your pages into numbers and finds near matches. Here is what it does, what it costs, and when a short script beats a store.

Dilshad Akhtar
Dilshad Akhtar
Last updated: 9 November 2026
4 min read
TL;DRAI summary
  • A vector database stores text as numbers and returns the nearest matches by meaning.
  • Most SEO teams can run the same job with a script. The store pays off past a few hundred thousand chunks.
  • Index sections, not pages, and keep prices, stock and legal text out of the index.

Most SEO teams do not need a vector database. They need a short script that turns their own pages into numbers and compares them. That covers clustering, cannibalisation and internal links for sites up to a few hundred thousand chunks. The database earns its keep when the index no longer fits in memory, and not a day before.

What is a vector database?

A vector database stores content as numbers instead of words. You hand it a chunk of text, a model converts that text into a list of numbers, and the database returns the chunks whose numbers sit closest to your query.

That changes the question you ask. A normal database finds rows that match a string. A vector database finds rows that mean something similar, even when they share no words. Pinecone describes the job as indexing and storing embeddings for fast similarity search, with metadata filtering and scale. https://www.pinecone.io/learn/vector-database/

The glossary definition from SEOS Design is plainer still: a system that stores data as vectors and returns results by similarity rather than exact words. https://www.seosdesign.se/en/glossary/vector-database

What an embedding actually is

An embedding is a list of floating point numbers. The distance between two lists measures how related the two texts are. Small distance means high relatedness. https://platform.openai.com/docs/guides/embeddings

Length is set by the model. OpenAI's small model returns 1,536 numbers per input, and the large one returns 3,072. Both accept 8,192 tokens of input at most.

Quality does not scale as fast as price. On the MTEB benchmark, the small model scores 62.3 and the large one 64.6. The older ada model scores 61.0. You pay a multiple for less than two points.

That matters for SEO work. A keyword list and a page body are both short texts. The cheap model is enough for grouping them.

How the search finds a neighbour

A vector database does not compare your query against every row. It builds an index that narrows the search, and the common choice is HNSW, a graph that links each vector to its nearest neighbours.

The trade is memory for speed. Weaviate publishes the bill: each node in the index holds 4 bytes per dimension, which puts a typical vector at 2 to 12 kilobytes. At one million vectors the index wants 2 to 12 gigabytes of memory. https://weaviate.io/developers/weaviate/concepts/vector-index

Read that number before you pick a plan. The disk is cheap and the memory is not.

What a vector database cannot do

Similarity is not truth. Four jobs stay with a normal database or a person.

  • Exact match. Part numbers, model codes and legal phrases need equality, not nearness.
  • Prices and stock. These change hourly, and a snapshot index reports yesterday.
  • Freshness. A page edited this morning is stale until you embed it again.
  • Authority. Distance tells you what is related, never what deserves to rank.

That is why a vector store sits beside your search data. It does not replace it.

Query shape Better tool Why
A product code, a person, an error message Keyword search Exact match has one right answer
Two phrases that mean the same thing Vector search No shared word needed
A price, a stock level, a date Database field The value must be current
"Find pages that overlap with this brief" Vector search Overlap is a meaning question
"Which page should own this query" Vector search, then a person Rank needs a judgement call

Where SEO teams actually use one

Five jobs repay the setup. Search Engine Journal walks through a clustering job of this shape with Vertex AI. https://www.searchenginejournal.com/introduction-to-vector-databases-and-how-to-use-ai-for-seo/533993/

  • Group keywords by meaning, so a cluster is a topic rather than a shared word.
  • Find pages that answer the same question, which is cannibalisation you can measure.
  • Suggest internal links from one page to its nearest neighbours.
  • Match a new brief against pages that already exist, before anyone writes a word.
  • Compare the questions you want to win against the pages you have, and see the gap.

Each one is a batch job. None needs a live query in front of a user, which changes the cost picture.

What it costs

The embedding call is the cheap part. Search Engine Journal quotes Vertex AI at 0.0002 dollars per 1,000 characters. OpenAI estimates 62,500 pages of text per dollar on its small model.

The store is the expensive part. ZenML reports a 64 dollar monthly minimum on Turbopuffer, and p50 latency above 200 milliseconds on cold queries. https://www.zenml.io/blog/vector-databases-for-rag

Latency only matters if a person waits. ZenML puts the target at under 100 milliseconds for interactive apps. Your Tuesday clustering run does not care.

Four questions that decide whether you need one

  1. Do any of your queries need exact match? If most do, keep keyword search first and add vectors behind it.
  2. How many chunks? Count sections, not pages. One post is usually five to ten chunks.
  3. How often do the pages change? Daily edits mean daily re-embedding, so price that in.
  4. Does a person review the output? If yes, a weekly script beats a live service.

Three no answers and a fourth that says "a person reads it" mean you do not need a store yet. Write the script.

What to index and what to leave out

Index the text a reader came for: section headings, body paragraphs, FAQ answers, product descriptions and image alt text. Leave out navigation, footers, cookie banners, prices, stock and legal text.

Two habits keep the result usable. Embed one section per chunk, not one page, so a match points at a paragraph. Store the URL, the page title and the last updated date beside each vector, so you can filter and so you know what is stale.

Our own page is a small example of the other failure. It draws 88 impressions from a single query, at average position 87.3. Close is not the same as right.

Start with a script, not a store

Write the small version first. Pull your pages, split them into sections, embed them, and answer one question with the result. Do it in a notebook and delete it if the answer is dull.

If that answer changes a page, you have a case for a store. If it does not, you saved the subscription.

Want a second pair of eyes on the plan before you build it? book a call.

Ready to Build Your Dream Website?

Let's discuss your project and create something amazing together.