For years, product recommendations on most B2B platforms were a polite afterthought. A “customers also bought” strip at the bottom of the page, driven by rules someone wrote once and rarely revisited. It worked, in the sense that it occupied space. It rarely worked in the sense that mattered, which is helping a professional buyer find the right product faster.
That era is ending, and we are building for what comes next. Aleph Front is developing a new recommendation engine, and this post is about why, and where the field is heading.
What changed
Three shifts have reshaped how recommendation and discovery work, and they have all landed at once.
The first is the move from keywords to intent. Search and recommendation no longer depend on a buyer typing the exact product name. Powered by semantic understanding and language models, modern systems interpret what someone actually means, including vague, natural, or conversational queries, and surface the right products even when the words do not match. A request like “a quiet compressor for indoor maintenance work” is now something a system can genuinely understand rather than fail on.
The second is real-time personalisation. Recommendations are no longer computed overnight from last month’s sales. They adapt continuously to live behaviour, including what a buyer is clicking, browsing, and putting in the cart during the current session. The experience adjusts as the visit unfolds, which is what makes it feel useful rather than generic.
The third is the rise of vector-based understanding. Under the hood, the most capable systems now represent products and buyers as embeddings, mathematical signatures that capture meaning and similarity rather than exact matches. Stored in a vector index and queried with approximate nearest-neighbour search, these embeddings let a platform find products that are genuinely related, by attribute, by use case, and by behaviour, at the scale and speed a live storefront demands.
Together these shifts turn recommendation from a static decoration into a discovery layer that understands the catalogue and the customer.
The trap we are designing around
There is a strong pull in the market to solve all of this by pushing everything to a large external AI service: send the behavioural data out, send the catalogue out, get recommendations back. It is fast to demo and convenient to start. It also means your customers’ behaviour and your commercial data continuously leave your perimeter, and it ties your discovery experience, one of the most strategic parts of your storefront, to a third party’s pricing and policies.
We have a clear view here, and it is consistent with everything else we build. The most valuable asset in a recommendation engine is first-party data, your own catalogue and your own customers’ behaviour. That asset should stay yours.
What Aleph Front is building
Our recommendation engine is being designed around the modern techniques that work, intent understanding, semantic and vector-based matching, and real-time adaptation, while keeping the data and the intelligence where they belong: inside your platform, under your control.
The shape of it:
- Semantic, vector-based matching so the engine understands products and queries by meaning and use case, not just keywords, and can surface genuinely relevant complements and alternatives across a large catalogue.
- Real-time signals that let recommendations respond to the current session, including browsing, search, and cart behaviour, rather than relying on stale overnight batches.
- First-party by design. The engine is built to learn from your own catalogue and your own customers’ behaviour, keeping that data inside your environment instead of exporting it to an external service.
- Local, independent models wherever inference is required, consistent with our broader approach: tuned to your catalogue, predictable in cost, and free of lock-in to a single AI vendor.
- Delivered as a managed service and integrated cleanly into the commerce platform, so the capability is operated for you without exposing the moving parts.
Why it matters for B2B
This is not about chasing consumer-retail novelty. In professional purchasing, the right recommendation is worth real money, because it shortens the path to the correct product, surfaces the compatible accessory or the better-suited alternative, and reduces the friction that makes a buyer abandon a basket. A recommendation engine that genuinely understands a technical catalogue, and that improves continuously from real behaviour, is a commercial instrument, not a cosmetic one.
We are building it the way we build everything: modern where it counts, sovereign where it matters, and engineered to keep our clients in control of their own data.
More to come as the engine takes shape. If recommendation and discovery are on your roadmap, we would welcome the conversation.
