Every learning platform today has a recommendation engine. Some suggest the next article. Others suggest the next quiz, video, or practice problem. But a recommendation that keeps a learner busy is not the same as a recommendation that helps them learn. The difference comes down to what the engine is optimizing for — and most of them are optimizing for the wrong thing.
What most recommendation engines get wrong
The typical approach in consumer technology is engagement optimization: recommend what keeps users clicking, watching, or scrolling. This works for media and e-commerce. It fails in learning because engagement and learning are not the same metric.
A learner who breezes through easy content will show high engagement. A learner who struggles through appropriately difficult content will show lower engagement. An engagement-optimized engine will recommend more of what the first learner finds easy and less of what the second learner needs. The result is a system that makes learners feel good without making them learn more.
A learning recommendation engine that optimizes for engagement will quietly widen the gap between struggling and advanced learners.
What makes a recommendation useful for learning
A useful learning recommendation engine optimizes for one of four things, depending on the learner's current state:
- Mastery building: recommending content just above the learner's current level — the zone of proximal development
- Knowledge gap closure: recommending content that addresses specific missing prerequisites
- Spaced repetition: recommending content at intervals scientifically shown to improve retention
- Transfer practice: recommending problems that require applying learned concepts in new contexts
None of these maximize engagement in the traditional sense. Some actively reduce it. But they are what actually move the needle on learning outcomes.
The data requirements that matter
To make useful recommendations, the engine needs more than interaction data. It needs learning data:
- Attempt-level data: not just whether a question was answered, but how many attempts, what errors were made, and how long was spent before answering
- Mastery signals: clear definitions of what it means to have mastered a concept, not just visited it
- Prerequisite graphs: a map of which skills depend on which others, so the engine knows what to recommend next
- Time-series learning data: how the learner's performance on specific skills changes over time
Most EdTech platforms collect interaction data — what learners clicked, how long they stayed, what they searched for. Fewer collect the attempt-level, mastery-oriented data that enables useful recommendations. This is not a technology gap. It is a data design gap that most platforms do not think about until they try to build the recommendation engine.
Signals that indicate a recommendation engine is working
If you are evaluating a learning platform's recommendation engine, watch for these signals rather than the flashiness of the suggestions:
- Recommendations change based on performance, not just history — if the engine suggests easier content after a struggle, it is adapting to learning state
- Recommendations sometimes surface content the learner would not have chosen — if recommendations only reflect what the learner already likes, the engine is optimizing for engagement, not growth
- The platform can explain why a recommendation was made — vague or generic explanations usually mean the engine is guessing
- Progress metrics improve over time — the ultimate test is whether the recommendations correlate with actual skill development
What Nivorius builds
When Nivorius designs recommendation engines for learning products, the first question is never 'what will keep the learner engaged?' It is 'what does this specific learner need next to make progress?' The answer comes from analyzing attempt patterns, mapping prerequisite relationships, and tracking mastery over time — not from predicting what will keep someone on the platform longer.
The best recommendation a learning platform can make is not the one that feels smart. It is the one that the learner does not yet know they need — the skill just outside their comfort zone, the concept that bridges what they know to what they need to learn next. That is what separates a useful recommendation from a sophisticated guess.
Part of the Nivorius research and consulting team, focused on practical applications of AI in education and enterprise contexts.
