The match math underneath may be perfectly sound but exposing it as a number, while historically a limited way to make a connection with a user, is now an obviously inferior method when compared to how LLMs can‘ talk’ to the user in emotionally resonant language. A number asks the viewer to trust arithmetic instead of the superior method a LLM conveys: feeling understood.
There is a hard requirement attached to this, and that is where the substance lives. An LLM explaining why you will like something is only as good as what it is grounded in. Left to improvise, a model will happily hallucinate descriptors, inventing tones and themes a title does not have. The fix is deep, accurate video metadata. The language the model reaches for has to be anchored in a factual understanding of the content itself, so that“ a foreboding British crime drama” is a true statement about the show and not a confident guess. Metadata is what keeps the‘ why’ honest.
The problem this solves is measurable, and it is not unique to our own data. Across the marketplace, viewers routinely spend ten to fifteen minutes deciding what to watch. That delay is rarely a failure of recommendation quality. The candidates on screen are often good. It is a failure of trust. The viewer does not yet feel that the system gets them, so they hesitate. A well-grounded explanation, in plain language, is often what closes that gap and gets someone to say,“ This system understands me”, and then they press play.
Where it belongs, and where it does not
Recognising the suggestible moment is one part of a larger question: mapping where generative AI genuinely improves the experience and where it quietly adds cost for no gain.
It clearly helps with the genre carousels that fill every interface today.‘ Action-packed movies for you’.‘ Comedies that’ ll make you laugh’. At a shallow level, these work. But when a system truly understands a household, GenAI can do better on both the label and the contents.‘ Funny films’ becomes‘ funny family films’ because the system understands there are teenagers in the home. Context shapes not just what the carousel is called but what goes into it.
It also helps with one of the oldest and hardest problems in recommendation: the cold start. A new release, by definition, has no viewership yet. Recommender models built on behavioural signals have nothing to work with on the day a title drops. This is where rich metadata does the heavy lifting again. If the system knows a household gravitates toward foreboding British crime dramas, and a new British crime drama with exactly that tone premieres today with zero views, GenAI grounded in metadata can make the connection that a purely behavioral model cannot. It can recommend what no one has watched yet.
Now the other side. There are use cases where reaching for an LLM is simply the wrong economic choice. Consider Continue Watching, or Recently Watched. These carousels are workhorses. In our data they can account for as much as a quarter of all click engagement, because so much viewing is a return to something already in progress: the series you are binging, your team’ s next game. This is predictive behavior, grounded in what you have already done, rather than suggestive behaviour reaching toward something new. Routing full viewership history through an LLM every time you need to render Continue Watching burns tokens, and cost, to reproduce something a traditional method does faster and cheaper.
The conclusion writes itself once you look at discovery this way. The personalisation platform for the next several years is a hybrid. Lower-cost, well-understood methods handle the high-frequency, predictive surfaces. Generative AI is reserved for the suggestive moments and the cold-start problems where it earns its keep. Treating GenAI as the default for everything is not just a quality mistake. It is a budget line item that will not survive contact with a finance team.
The model is a component, not the entire system
There is a final point that gets lost in the enthusiasm, and it may be the most important one for anyone planning a roadmap. An LLM sits inside a system. It is not a system unto itself.
A recommendation platform was never just a recommender model, and generative AI does not change that. It is one more model inside a platform, and the platform still has to surround it with the scaffolding that makes the whole thing safe and reliable. Business rules, for instance, have to remain hard constraints, not probabilistic outputs. Removing adult content from a recommendation cannot be left to something that might hallucinate. It has to be a rule that always holds. The same logic applies to availability. Few things erode trust faster than convincing a viewer that a title is perfect for them, only for them to discover it is not available on their service, or costs extra to rent. So, metadata matters not only for describing content, but for grounding availability: what is watchable, where, and for how long. The same can be said for editorial input. Curators are not replaced by LLMs, rather they are simply another tool to wield to allow editorial vision to be crafted and delivered with personalisation. The LLM has to be wrapped in tools that already exist in most mature recommendation platforms today.
That leads to an important caution worth stating plainly. For anyone tempted to rip out their recommendation backend and replace it wholesale with an LLM, slow down. A generative solution fits inside all this other infrastructure. It is not a panacea. Used selectively, grounded in solid metadata, and surrounded by the right business rules and availability data, it delivers real value in the moments that matter. But it does not replace the scaffolding that holds the system up, and the operators who mistake the model for the system will learn that the expensive way.
The through-line across all of this is not that generative AI transforms discovery. It is that GenAI is a powerful instrument that must be deployed with precision. Know the moment you are in. Explain the‘ why’ in language people trust. Reserve the tool for where it earns its cost. And remember that it is one component within a much larger system. That is a less thrilling story than‘ everything changes’, but it is the one that will actually be true, and it is the one worth building on.
EUROMEDIA 15