Most on-site search still works like a librarian who only checks the spine of the book for an exact word match. Vector search works more like a colleague who understands what you actually meant. If you are weighing up an upgrade to your search and recommendations stack, it helps to understand what vectors are, where they help, and — just as importantly — where the traditional approach is still fine.
This is a primer for commerce teams, not engineers. We will skip the linear algebra and focus on the practical decisions you need to make.
The problem with keyword search
Classic search engines match the words a shopper types against the words in your product data. This is fast and predictable, but it breaks in familiar ways:
- A search for “warm jacket” misses a product tagged “insulated coat” because none of the words overlap.
- “Trainers” returns nothing on a catalogue that only uses “sneakers”.
- A typo such as “hoddie” falls flat unless someone has configured a synonym or fuzzy rule.
- Long, natural queries like “shoes for standing all day” confuse a system built for one or two keywords.
Teams patch these gaps with synonym lists, redirects and manual rules. That works until the list grows to thousands of entries that nobody fully trusts. There is always another phrasing you did not anticipate. We dig into this failure mode in How to Fix Zero-Result Searches.
What a vector actually is
A vector (also called an embedding) is a list of numbers that represents the meaning of a piece of text, an image, or both. A machine learning model reads your product description and outputs, say, 768 numbers that capture its semantic fingerprint.
The crucial property is this: things that mean similar things end up with similar numbers. “Insulated coat” and “warm jacket” land close together in this numeric space, even though they share no words. “Coat” and “kettle” land far apart.
You do not need to understand the numbers themselves. The only intuition that matters: every product and every search query becomes a point in a high-dimensional space, and proximity equals relatedness.
Similarity, not exact matching
When a shopper searches, their query is turned into a vector too. The engine then finds the product vectors closest to the query vector. That is semantic search: matching by meaning rather than by overlapping characters. For a deeper treatment of the shopper experience this enables, see Semantic Search for eCommerce.
Where vector search earns its keep
In our experience, the payoff is largest in a few specific situations:
- Large or messy catalogues. When product data is inconsistent — different teams, suppliers or languages describing similar items in different ways — vectors paper over the gaps automatically.
- Natural-language and conversational queries. As shoppers increasingly type full questions (“something to wear hiking in the rain”), keyword systems struggle and embeddings shine.
- Discovery-led categories. Fashion, home, gifting and hobbies, where shoppers browse by intent rather than knowing the exact product name.
- Cross-lingual search. A single multilingual model can match a Dutch query to an English product description, which is invaluable for cross-border retail. We cover this in Multilingual and Cross-Border Search.
Where keyword search is still the right tool
Vectors are not a universal upgrade. Keep traditional matching — usually alongside vectors — when:
- Shoppers search by exact identifiers: SKUs, model numbers, ISBNs, part codes. Embeddings are poor at precise strings.
- You sell in a narrow, well-defined catalogue where a tidy synonym list already does the job.
- Filters and facets do most of the navigation work, as with many B2B or spare-parts businesses.
The honest answer for most retailers is hybrid search: combine keyword precision with vector recall, then blend the scores. This is the architecture we recommend most often.
How a vector search system fits together
There are four moving parts worth knowing about:
- An embedding model turns text (and optionally images) into vectors. You can use a hosted model or run your own.
- A vector index stores those vectors and finds nearest neighbours quickly. Many search platforms now bundle this.
- A query pipeline embeds the incoming query, retrieves candidates, and merges keyword and vector results.
- A ranking layer re-orders the candidates using business rules, popularity, margin and availability.
That last layer matters more than teams expect. Relevance is necessary but not sufficient — you still need merchandising control, which we explore in AI Merchandising Control.
Costs and trade-offs to plan for
Vector search is not free. Budget for the realities:
- Re-embedding. Every time you change product descriptions or swap the model, vectors must be regenerated. Plan a refresh pipeline.
- Latency and infrastructure. Nearest-neighbour search over millions of items needs the right index and tuning. For typical mid-market catalogues this is manageable; at very large scale it needs real attention.
- Explainability. When a keyword match returns the wrong thing, you can see why. When a vector match does, the reason is buried in numbers. Good tooling and logging are essential for debugging.
- The cold-start gap. Brand-new products with thin descriptions produce weak vectors. Enrich your product data first.
A common pitfall is treating vector search as a switch you flip. In practice it is a tuning exercise: you will iterate on the blend of keyword and vector scores, the ranking rules, and the product data feeding the model.
A pragmatic evaluation checklist
Before committing, work through these questions with your team:
- What proportion of current searches return zero or poor results? That is your headroom.
- How clean and complete is your product data? Vectors amplify good data and expose bad data.
- Do you need multilingual coverage now or within a year?
- Can your platform support hybrid ranking, or is it vector-only?
- How will you measure success — search conversion rate, zero-result rate, click depth — before and after?
If you cannot answer the measurement question, start there. You can read our view on the metrics that matter in Site Search Is Your Highest-Converting Channel.
The bottom line
Vector search is a meaning-based upgrade to on-site search that closes the gap between what shoppers type and how your catalogue is written. It is most valuable for large, varied or discovery-led catalogues and for natural-language and multilingual queries — and it works best in a hybrid setup with keyword matching and a strong ranking layer, not as a wholesale replacement.
If you are deciding whether vector search is worth the investment for your catalogue, our AI Search and Recommendations team can help you scope a pilot and measure it properly. Get in touch and we will give you a straight read on the likely return.