The Discovery Brain: Why LLMs Will Become the Foundation of Streaming Recommendations

Every streaming product is built around a deceptively simple question: what should this person watch next?

While it is a simple question on the surface, a good answer depends on:

• what the customer has watched
• what they abandoned
• what they finished
• what they searched for
• what they saved
• what they watched recently
• what they tend to watch in different contexts
• what is available to them
• what is appropriate for their profile
• what device they are using
• what language they prefer
• and what kind of experience they may be trying to have at that moment.

Public descriptions of streaming recommendation systems already show how layered this problem is, with signals ranging from viewing history and title attributes to time of day, language, device, row selection, title selection, title order, and continuous feedback loops.

For years, the industry has treated this primarily as a prediction problem. Given a user, a catalog, a surface, and a moment, predict what title should appear next. That framing has taken us a long way. It gave us personalized homepages, smarter search, “because you watched” rows, continuation experiences, title similarity, editorial blends, and increasingly sophisticated ranking systems.

But I think the next generation of streaming recommendations will not simply be a better prediction system. It will be a reasoning system.

The next generation of streaming recommendations will not simply be a better prediction system. It will be a reasoning system.

A prediction system asks, “What is this user likely to watch?” A reasoning system asks a more complete product question: “What is this user trying to accomplish right now, what can they actually watch, what does the catalog really contain, what tradeoffs matter, and which recommendation can we defend?”

This is where frontier models become strategically important. LLMs don’t magically know what everyone should watch. They don’t replace every part of the recommendation stack. A chatbot is not the future of streaming discovery. However, an LLM can become the shared reasoning layer that sits across the entire discovery experience.

If I were building a streaming service from scratch today, I would not start by creating separate recommendation services for every surface. I would not build one intelligence layer for search, another for post-play, another for homepage rows, another for franchise pages, another for collections, another for “because you watched,” and another for conversational discovery. That is how systems accrete over time, but it is not the architecture you would choose if you were starting fresh.

I would build one discovery brain.

That discovery brain would not be one unconstrained model making things up from memory. It would be a grounded, governed, LLM-native decision platform. It would have access to the necessary context:

• click and engagement data
• profile memory
• session context
• content understanding
• catalog relationships
• rights
• maturity constraints
• editorial intent
• business signals
• and feedback history.

But it would not use those inputs casually. It would receive compact, governed representations of the facts and signals the product is willing to stand behind.

The LLM should reason. The platform should know.

The LLM should reason. The platform should know.

The catalog remains the source of truth. Rights systems remain the source of truth. Maturity ratings, profile boundaries, availability, territory, subscription tier, suppressions, and policy constraints remain deterministic. The LLM should not decide whether a title exists, whether it can be shown, whether it is available in a market, or whether it belongs on a kids profile. Those are not judgment calls. They are product truths.

The LLM’s job is different. It should receive an eligible slate of real catalog IDs and rank that slate with more nuance than a collection of disconnected services can provide. It should reason across tone, mood, pacing, novelty, familiarity, fatigue, intensity, format, franchise adjacency, context, and satisfaction. It should understand that “something like this, but lighter” is not the same as “more of the same.” It should understand that a prestige drama on Sunday night, a comfort comedy after work, a true-crime documentary with a parent, and a short mobile session while traveling are different recommendation problems, even for the same customer.

Traditional recommenders can model pieces of this. They are very good at many parts of the problem, especially when there is rich behavioral history. But entertainment taste is not only behavioral. It is semantic, emotional, social, and situational. People do not merely consume genres. They choose experiences.

That is why LLMs are interesting for streaming. They are good at translating human intent into structured meaning. They can connect the way people describe entertainment with the way a catalog is represented. They can reason across soft attributes that matter deeply to customers but are difficult to encode cleanly in a single ranking feature. They can also explain recommendations in language that helps a customer decide faster.

The risk is that fluency can be mistaken for reliability. A model can recommend a title that does not exist, describe a title incorrectly, ignore availability, overstate confidence, or produce a recommendation that sounds plausible but should never render. In a production streaming experience, that is not a small problem. A weak recommendation is forgettable. A confidently wrong recommendation breaks trust.

The winning architecture is not “ask the LLM what to show.” It is “ask the LLM to rank a grounded, eligible slate under a clear contract, then validate the output before the customer ever sees it.”

The practical stack looks something like this:

• Classical retrieval systems generate candidates from collaborative signals, content similarity, search intent, continuation logic, editorial programming, trending systems, franchise graphs, and business context.
• Deterministic filters remove anything that cannot be shown because of rights, territory, maturity, tier, device, profile, policy, or suppression rules.

The LLM then receives a compact decision packet: who the profile is in broad terms, what the session context is, what surface is asking, what objective matters, what eligible titles are in the slate, and what facts are known about each title.

The model ranks the slate. It may also produce grounded explanations, but only from supplied facts. Then a validator checks the response.

• Did the model return only valid catalog IDs?
• Did it duplicate anything?
• Did it violate eligibility?
• Did it produce a grounded explanation?
• Did it leak spoilers?
• Did it make an unsupported claim about the user or the title?
• Did it stay within the schema?

If the answer is no, the system repairs, retries, or falls back to a deterministic ranking path.

That is not a compromise. That is how probabilistic intelligence belongs inside a production product. The model output is a proposal. The product output is what survives validation.

This approach also reframes what it means to “replace recommendation services.” The goal is not to delete retrieval, ranking, experimentation, feedback loops, or constraints. The goal is to replace the fragmentation of intelligence across surfaces. Search, homepage, post-play, rails, franchise pages, collections, and assistants should not each have their own disconnected understanding of the customer and the catalog. They should become clients of the same discovery foundation.

That foundation would create leverage. A better content representation improves every surface. A better user memory improves every surface. A better validator improves every surface. A better satisfaction signal improves every surface. A better understanding of “comfort watch,” “shared-screen safe,” “not too heavy,” or “similar but fresher” improves every surface.

The LLM is not merely another model in the stack. It becomes the interface between human intent and machine decisioning.

The LLM is not merely another model in the stack. It becomes the interface between human intent and machine decisioning.

It also changes what teams should invest in.

1. The first investment is content understanding. A streaming catalog needs to become legible to both machines and humans at a richer level than genre, cast, and synopsis. We need spoiler-safe representations of tone, themes, pacing, intensity, humor, emotional payoff, audience fit, setting, structure, franchise relationships, and confidence. The LLM can only reason well if the content layer is truthful and expressive.
2. The second investment is profile memory. Not all preference signals are equal. Some are stable, some are temporary, and some are accidental. A user may love political thrillers in general, but want a light comedy tonight. They may abandon a show because they disliked it, or because the session was interrupted. They may finish a movie and feel satisfied, exhausted, bored, comforted, or inspired. A discovery brain needs memory, but that memory must be conservative, profile-specific, privacy-aware, and legible.
3. The third investment is feedback. Streaming services already learn from starts, completion, abandonment, search behavior, ratings, saves, hides, and rewatching. But if the ambition is satisfaction rather than raw engagement, we need better ways to learn what the experience felt like. That does not mean turning entertainment into homework. It means lightweight, optional feedback that helps the product understand whether the customer wants more of the genre, more of the mood, more of the pace, more of the world, or none of the above.
4. The fourth investment is measurement. An LLM-native recommendation layer should not be judged by demo quality. It should be judged by production outcomes. Does it improve satisfaction? Does it improve retention? Does it reduce abandonment? Does it increase successful discovery without collapsing diversity? Does it protect trust? Does it respect kids profiles, maturity rules, rights, and availability? Does it perform under latency and cost constraints? Does it beat strong non-LLM baselines?

AI excitement should not lower the launch bar. It should raise it.

AI excitement should not lower the launch bar. It should raise it.

I would not begin this transformation with the highest-blast-radius surface. I would not start with the homepage hero. I would start where intent is clearer and the slate is more constrained: post-watch recommendations, “because you watched,” or another focused discovery moment where the system has fresh context and the consequences of fallback are manageable. Prove the contract. Prove the validation layer. Prove the measurement. Prove that the LLM can rank better than the current system without violating trust.

Then expand deliberately.

Over time, the architecture becomes more ambitious. The homepage becomes less like a set of independently optimized rows and more like a coherent expression of the customer’s current and long-term relationship with the catalog. Search becomes less transactional and more intent-aware. Post-play becomes less generic and more emotionally intelligent. Editorial programming becomes more targeted without becoming manipulative. Business strategy enters the system explicitly, with scope, reason, time window, and auditability, rather than quietly distorting personalization.

No streaming product is purely personalized. There are premieres, tentpoles, live events, expiring titles, franchise moments, seasonal programming, and strategic priorities. The question is whether the signals are explicit and governed. In an LLM-native discovery platform, strategic signals should have a seat at the table, not the steering wheel. They can break ties, create measured exposure, and support cultural moments. They should not rescue a recommendation that is otherwise indefensible.

The customer has to feel that the product is listening. If the product starts to feel like it is merely promoting, personalization has failed.

This is why I believe the future of streaming discovery is not a chatbot bolted onto the existing recommendation stack. The opportunity is to rebuild the foundation so that every discovery surface shares the same underlying intelligence: the same grounded catalog understanding, the same governed profile memory, the same context model, the same validation discipline, and the same feedback loop.

The phrase “LLM as recommendation engine” is therefore both right and wrong. It is wrong if it means handing the catalog to a model and asking it to improvise. It is wrong if it means replacing retrieval, constraints, experimentation, and learning with a prompt. It is wrong if it treats fluency as product quality.

But it is right if we mean something more precise: the LLM becomes the primary reasoning and ranking layer over grounded, eligible slates. Classical systems retrieve. Deterministic systems enforce truth. The LLM reasons across context. Validators protect the customer. Feedback teaches the platform what satisfaction actually means.

LLMs do not make recommendation science obsolete. The old problems still matter: retrieval, ranking, cold start, exploration, diversity, feedback, latency, cost, governance, and experimentation. But LLMs give us a new way to connect those systems to human intent. They give us a way to make discovery feel less like a collection of optimized widgets and more like a product that understands what the customer is trying to experience.

The streaming service that wins will not be the one that lets an LLM invent recommendations. It will be the one that uses an LLM to unify discovery around a grounded, trusted, and continuously learning decision platform.

The streaming service that wins will not be the one that lets an LLM invent recommendations.

That is the discovery brain. And I think it is where streaming recommendations are headed.

---

This writing reflects my personal perspectives on product management, AI, and content discovery. It does not represent the official position of my employer or any affiliated organization.