Mallika Rao, a longtime practitioner in search, discovery, and personalization, makes a case in a recent presentation that the most common mistake organizations make with recommendation systems is treating them as isolated modeling problems. In her view, the real challenge lies in building the entire end-to-end system, the infrastructure that lets a recommendation engine continuously learn, adapt, and deliver value under real-world constraints like latency, cost, observability, compliance, and customer trust. That insight, she argues, is one that the broader AI industry is only now arriving at, and that recommendation systems have been grappling with for years.
From Static Heuristics to Adaptive Ecosystems
Several years ago, most recommendation systems looked quite different from what exists today. They were built on handcrafted heuristics, carefully tuned business rules, and deterministic ranking functions. These systems worked, and in many cases worked well, but they had structural limitations. They struggled when signals were sparse, when user intent shifted quickly, or when scale demanded faster adaptation. Refresh cycles were periodic, meaning a model trained yesterday operated on a world that may have already changed.
Modern systems have moved toward what Rao describes as AI-native architectures. Embeddings capture semantic similarity in ways that handcrafted features could not. Contextual signals let the system adjust to changing intent. Multistage retrieval pipelines and ranking systems handle scale. But this evolution introduced a new class of operational challenges. The system is no longer static, and the feedback it receives from users changes the system in ways that must be monitored, measured, and controlled.
Rao frames the distinction sharply: a static system makes decisions based on yesterday's understanding of the world, while an adaptive system is making decisions in the present, based on what is happening right now as the user interacts with the product. That shift is subtle, but it fundamentally changes how the system must be architected. Latency, freshness, and the speed of feedback loops take on a different kind of importance.
Retrieval Is a Breadth Problem
When most engineering teams think about recommendation systems, they focus on ranking, which is understandable. Ranking models are the most visible component, the layer that directly shapes what a user sees. But ranking operates only on candidates that have already been retrieved. If an item never enters the candidate set, no amount of ranking sophistication can surface it.
Rao describes retrieval as a breadth problem. The goal is not precision at that stage; it is to find a sufficiently large and diverse set of candidates so that the ranking layer has a rich universe to work with. Modern systems typically run multiple retrieval strategies simultaneously: embedding-based retrieval for semantic similarity, behavioral retrieval for user history, popularity retrieval for trending content, and context-aware retrieval for temporal or situational signals. Each strategy captures a different signal type, and together they define the search space.
Rao points out that many teams spend disproportionate effort fine-tuning downstream ranking models while neglecting upstream retrieval quality. Improving retrieval freshness or expanding the candidate set can have a larger impact than any model optimization, because retrieval determines what the system is even allowed to consider.
Ranking, Intent, and the Shift to Orchestration
Once candidates are retrieved, ranking decides what deserves the user's attention. Modern ranking systems derive their value not from increasing model complexity but from signal complexity. Behavioral data, engagement signals, contextual information, session history, temporal patterns, content features, and evidence-based signals all feed into the model. Increasingly, intent signals are the most differentiating factor among them.
A user who hovered over a product at 9 a.m. may have had a different intent than the same user at 6 p.m. Goals shift based on time of day, location, and changing circumstances. Some of those shifts happen within seconds or minutes. This means ranking systems must answer a more nuanced question than simply what this user generally likes; they must answer what this user likes right now. Personalization becomes not just about predicting who the user is, but about understanding what matters to them in this specific moment.
Beyond ranking, a more fundamental shift is underway. Systems are evolving from ranking pipelines into orchestration systems. Modern experiences must balance multiple surfaces, multiple objectives, multiple intent types, and multiple user goals, often simultaneously. A recommendation that is relevant to one user might create a worse experience for another viewer on the same page. A user might need exploration rather than refinement, or diversity rather than precision. Those are orchestration decisions that go beyond what any single ranking model can determine.
The Feedback Trap
Rao considers the feedback system the most important concept in the presentation, because every recommendation system is ultimately a feedback loop. A recommendation is made, the user reacts, that reaction becomes data, and the cycle repeats. From the top, this looks straightforward. In practice, it introduces surprising complexity.
Positive outcomes can compound in ways that narrow the system's scope. Negative outcomes can do the same. Systems accidentally optimize themselves into narrow paths, reducing exploration and becoming too conservative based on skewed feedback. They become repetitive, trapping users in local optima where the same content keeps surfacing regardless of the user's stated or implicit desire for variety. Rao emphasizes that these are not model problems. They are system behavior problems, and they require system-level solutions, not just better algorithms.
Evaluation as a First-Class Citizen
Rao spends significant time on evaluation because she believes it is the most underserved aspect of recommendation system design. Too often, organizations treat modeling as the core discipline and evaluation as an afterthought. In her experience, that gap is critical. Evaluation is harder than modeling because the feedback signal is continuous, noisy, and often delayed. Metrics that look good offline may not translate to online behavior, and experimentation frameworks must account for the fact that every change to the system affects the feedback loop itself.
She argues that evaluation must be treated as a first-class concern from the start, not bolted on after a model is deployed. Organizations that get this wrong deploy systems that appear to perform well in testing but degrade the user experience in production, often without anyone noticing until the feedback has already compounded into something harder to reverse.
Where Adaptive Systems Are Heading
Rao's presentation covers migration strategies and operational lessons, though those portions were cut short. The throughline across the talk is clear: the teams that succeed with recommendation systems are not the ones with the best models, but the ones that treat the system as a whole, with deliberate attention to how retrieval, ranking, orchestration, feedback, and evaluation interact. The lessons from recommendation systems, she notes, are becoming broadly applicable as the broader AI industry builds more adaptive and agentic systems. The question of how to operate under uncertainty, manage feedback loops, and maintain control while allowing adaptation is not unique to recommenders, and the approaches developed in that domain offer a template for what comes next.