Relevant products, content or people from your own catalogue, ranked by a model, bounded by your business rules and explained in a sentence the customer can trust.
Large catalogues hide most of their value. Search only returns what the customer already knows to type, and "popular" lists show everyone the same few items. Merchandising rules written by hand help the top categories and never reach the long tail, where margins are often better.
A recommendation engine built on embeddings understands what each item is, matches it to what each customer is doing, and lets the merchandising team shape the result rather than fight it.
How it works
From catalogue and behaviour to a ranked, explained list
Retrieval finds a few hundred plausible candidates in milliseconds; the ranker and the rules decide the final dozen. Merchandisers shape the output and the live test decides whether it ships.
1. Understand every item
Descriptions, images and attributes are embedded so the engine knows that two products are similar even if nobody has bought them together yet. Customer context is embedded the same way.
2. Retrieve, then rank
Vector search narrows the catalogue to candidates. A ranking model trained on your clicks and purchases orders them, and rules for stock, margin and diversity are applied last.
3. Explain and test
Each recommendation carries a reason: "similar to what you viewed", "pairs with your basket". Results run as an A/B test against the current system before full rollout.
What you get
An engine your merchandising team can steer
Embedding and retrieval layer
Item and context embeddings in a vector index sized for your catalogue, using our semantic search stack on pgvector or Milvus.
Ranking model with business rules
Trained on your behaviour data, retrained on a schedule, with stock, margin, diversity and campaign rules applied after ranking and reported on separately.
Merchandiser console
Pin, exclude and promote with an audit log and an effect report, so manual rules are kept only when they earn their place.
Explanations and placements
A recommendation API with a reason string for product pages, basket, email and app placements. Media companies use the same API for articles and video, as covered under media.
Experimentation and reporting
A/B framework, offline evaluation on historical sessions and a dashboard for conversion, basket value, diversity and coverage of the long tail.
Guardrails we build in
Accuracy thresholds: offline metrics must match or beat the current system before a live test, and the live test must show a statistically meaningful lift before rollout. Human-in-the-loop: merchandisers control pins, exclusions and campaign rules; sensitive categories can be fenced off entirely. Audit trail: every recommendation served is logged with its candidates, rank and the rule that shaped it. PII handling: behaviour data is pseudonymised, consent and opt-out are respected per region, and profiles can be deleted on request, following our AI data privacy controls and the wider AI security and governance practice.
Typical rollout
One placement first, then the rest
Audit (2 weeks)
Catalogue quality, behaviour data coverage and current recommendation performance. Choose the first placement and the metric it is judged on.
Build (4 weeks)
Embeddings, vector index, candidate retrieval, first ranker and the merchandiser rules. Offline evaluation on historical sessions.
A/B test (3–4 weeks)
Live traffic split against the current system on the chosen placement. Weekly readouts with merchandising, adjustments to rules and ranker.
Roll out and extend (2 weeks)
Winning variant to all traffic, additional placements such as email and basket, retraining schedule and handover.
Illustrative example
What a first release usually looks like
An illustrative scenario rather than a client case
An online retailer with tens of thousands of products shows "customers also bought" lists that favour the same bestsellers everywhere. The first placement to change is the product page. Embeddings bring in items similar in style and use, the ranker orders them by likelihood to convert, and a diversity rule allows no more than two items from the same brand. Merchandisers pin seasonal lines for a campaign and see in the report that one of their older promotion rules was reducing conversion, so they retire it. The live test runs for a month before the new list replaces the old one and the basket placement follows. This is the same mechanism our retail and e-commerce practice applies to search and personalised email.
How is this different from the recommendations built into our e-commerce platform?
Built-in modules usually rank by popularity or co-purchase and cannot see your content, your margins or your stock. An embedding-based engine understands what items are as well as what sold together, so it works for new products and long-tail catalogues, and it respects the rules your merchandisers set.
What about the cold start problem for new customers and new items?
Embeddings handle new items immediately because the engine understands their description, images and attributes. New customers get recommendations from context, such as the page they are on or the query they typed, and improve as they interact. See our embeddings strategy service.
Can merchandisers override the engine?
Yes. They can pin, exclude or promote a category, or run a campaign rule, and those rules are applied after the model ranks. Every rule is logged and its effect on conversion is reported, so the team learns which manual interventions actually help.
How do you prove it works?
A/B testing against the current recommendations with conversion, basket value and return rate as the metrics. Offline evaluation on historical sessions is done first to catch obvious regressions, but the decision to roll out is based on the live test.
Does the same approach work for matching people as well as products?
Yes. Our attendee matchmaking for Eventify uses the same mechanism: embed profiles and interests, retrieve candidates, rank with rules such as role and availability, and explain the match. The case study describes it.
Talk to us about your AI project
Tell us what you are trying to do, which systems are involved and what is in the way. We reply within one business day with an opinion and a suggested next step.
