Real-time ranking

Most recommendation systems precompute candidates in a nightly batch. Real-time ranking adds an online step: when a page is requested, a low-latency service scores those candidates against the current session, including the last few items viewed, the active search, the cart, the device and the time of day. A visitor who has just looked at three running shoes sees running gear ranked up immediately, not tomorrow.

Architecturally it needs a fast event pipeline that makes actions available within seconds, a feature store or session cache, and a scoring service that responds in tens of milliseconds so the page does not slow down. Models are usually light, such as gradient-boosted rankers or small neural models, and heavy lifting stays offline.

It is not the same as retraining the model in real time, which is rarely necessary. The model is fixed; the inputs are fresh. Nor is it just "recently viewed": it changes the order of everything the user sees based on inferred intent. It is the main answer to cold start for anonymous visitors, because the session is the only history they have.

• Product manager: session intent is the strongest signal you have for the majority of visitors who are not logged in; ignoring it wastes it.
• CTO / Head of Engineering: this is a latency and streaming problem; the serving path must be engineered so personalisation never becomes the reason a page is slow.
• Founder / CEO: for retail and marketplaces the uplift from within-session personalisation is often larger than from profile-based personalisation, because it reaches everyone.
• Data lead: you will need event freshness measured in seconds, which is a different pipeline from the one feeding your dashboards.

Real-time ranking exists because intent changes faster than profiles. What someone wants right now is signalled by what they did in the last two minutes, and a system that only knows last month's purchases will show the wrong things. It also solves the practical problem that most sessions are anonymous, so there is no profile to use. The trade-off is engineering cost: streaming events, session state and a sub-hundred-millisecond scoring path are real infrastructure, and every added feature must be justified by measured uplift in a controlled test rather than by intuition.

• A fashion retailer re-ranking category pages after each product view so the page reflects the shopper's emerging taste.
• A quick-commerce app reordering the home screen once a customer adds the first item to the cart.
• A news or media product adjusting the feed as the reader skips or finishes articles.
• A B2B SaaS app promoting relevant help content and features based on the screens visited in the current session.
• A travel or ticketing site adjusting offers as the user changes dates, class or destination.

Technique / practiceAn architecture and modelling technique within recommendation systems. Eazyware implements it as part of personalisation engines, usually with a streaming event pipeline and a session feature cache, measured against the batch baseline in a controlled test.

Eazyware service that covers it: Personalization Engines. Starting prices are on the pricing page.