Help people find what they mean, not just what they type.
A search box that returns nothing for "shoes for wet weather", or recommendations that suggest the item someone just bought, quietly cost you revenue. We build semantic search and recommendation systems that pair AI understanding with business rules and honest measurement, and we build them AI-first.
Signs your search and recommendations are costing you
Most of these show up in analytics you already collect. If several apply, better relevance is probably worth more than another new feature.
A high zero-result rate
Queries that return nothing are often synonyms, misspellings or descriptions of a need rather than a product name.
Lots of immediate re-searching
Users rephrasing the same query two or three times are telling you the first results missed.
Searchers leave from the results page
People who search usually have high intent. Losing them at that point is a relevance problem, not a traffic problem.
Recommendations dominated by best-sellers
Popularity bias shows everyone the same items and hides the long tail, whatever the visitor was looking at.
Staff maintain synonym lists by hand
A growing spreadsheet of "sofa equals couch" rules means keyword matching has reached its limits.
Your own teams cannot find internal content
Search is not only for customers. Case files, product data and knowledge bases suffer from the same problems.
Keyword, semantic or hybrid search?
Semantic search is not a straight upgrade on keyword search. Each wins on different queries, which is why most production systems combine them.
| Keyword search | Semantic (vector) search | Hybrid search | |
|---|---|---|---|
| Exact SKUs, codes, part numbers | Excellent | Unreliable | Excellent |
| Descriptive or vague queries | Weak | Strong | Strong |
| Synonyms and other languages | Manual synonym lists | Handled by embeddings | Handled, with keyword precision |
| Explaining why a result ranked | Easy | Harder | Moderate |
| Infrastructure cost | Low | Embedding and index costs | Highest, usually justified |
| Ongoing tuning | Rules and boosts | Model and field choices | Blending weights and reranking |
Recommendation engines matched to the data you have
The right approach depends on how much behavioural data exists. We start simple and add sophistication only where testing shows it pays.
Similar items
Content-based suggestions from product attributes, descriptions and images. Useful from day one, even with no user history.
Behaviour-based recommendations
Collaborative filtering on views, baskets and purchases, once there are enough interactions to beat simple rules.
Session-aware suggestions
Recommendations that react to what a visitor is doing right now, which matters when most traffic is anonymous.
Business rules on top
Stock, margin, compliance exclusions and promotions shape the final ranking, so the model never pushes what you cannot sell.
Relevance for media owners
With Nodals.ai, an AI advertising platform for media owners, relevance runs on per-publisher models with first-party data under publisher control.
Relevance is only half of the job
Search has a latency budget measured in milliseconds, and every AI step spends some of it. Embeddings are computed ahead of time when products or content change, not at query time, and any language model reranking is limited to a short list of top results. When the budget is tight, we cache frequent queries and keep a fast keyword path as a fallback.
Recommendations create feedback loops. Items that are shown get clicked, clicks get them shown more, and newer products never get a chance. We reserve room for exploration, monitor catalogue coverage alongside conversion, and review categories where a poor suggestion would be embarrassing or non-compliant.
Behavioural data is personal data. Personalisation should respect consent choices and retention rules in the markets you serve, and session-level signals can often deliver useful results without building long-term profiles of individual visitors.
How we build semantic search and recommendations
AI speeds up our own delivery here too. Language models produce first-pass relevance grades for thousands of query and result pairs, which people spot-check before those grades steer any ranking decision. Senior engineers own the ranking logic and experiment design.
Baseline today
Week 1Search logs, zero-result queries, clicks and conversions, so improvement is measured against reality rather than impressions in a demo.
Build a judged query set
Weeks 1 to 2Hundreds of real queries with graded results, used to score every ranking change offline before a customer sees it.
Prototype hybrid ranking
Weeks 2 to 5Keyword and semantic retrieval combined, a reranking step and business rules, tuned against the judged set within a strict latency budget.
A/B test in production
Weeks 5 to 8Live experiments on conversion, revenue per search and engagement, with the definition of a win agreed before the test starts.
Questions we often hear
What is semantic search?
Semantic search finds results by meaning rather than exact words. Text is converted into numerical representations called embeddings, so a query for "quiet dishwasher" can match products described as "low noise" without a synonym list. It works best combined with traditional keyword search.
How do AI recommendation systems work?
They predict what someone is likely to want from signals such as the item being viewed, past behaviour and what similar customers chose. Techniques range from attribute similarity to collaborative filtering and sequence models, usually finished with business rules that reflect stock and margin.
Do we need a lot of data to build a recommendation engine?
Not to start. Content-based recommendations need only good product or content data. Behaviour-based methods need enough interactions to find patterns, so smaller catalogues and newer businesses often begin with similarity and rules, then add personalisation as traffic grows.
Can we use an LLM for search instead of a search engine?
LLMs are valuable inside search for interpreting queries, generating embeddings and reranking the top results, but asking one to scan a whole catalogue on every query is slow and expensive. The effective pattern is fast retrieval first, then AI applied to the shortlist.
How do you measure whether better search is working?
Offline, we score rankings against real queries with judged results. Online, we run A/B tests on outcomes such as conversion, revenue per search and zero-result rate. Offline tests catch regressions early, and live tests confirm the business impact.
Related reading and tools
Tell us what people struggle to find.
The quickest route to a useful conversation is your search analytics: top queries, zero-result searches and exit rates. Send what you have, and within 24 hours we will point out the likely quick wins and whether a bigger rebuild is justified.
Working with companies globally · Response within 24 hours