Company contribution
Company contribution
Judgement Over Hype: Rethinking GenAI in Content Discovery
According to Chris Ambrozic, vice president and general manager of discovery at TiVo, knowing when to deploy AI in content discovery is key.
Watch how people search now. Where someone once typed‘ best hiking boots’ into a search bar, they now type‘ best waterproof hiking boots for wide feet under a hundred and fifty dollars’. Something shifted over the last two years. People learned that offering more context buys them a better answer, and generative AI is what taught them. They make that trade willingly, because it pays off.
We see the same behaviour arriving in video discovery. After nearly two decades of running natural language understanding solutions at scale, the pattern in our data is clear: conversational queries are on the rise. Viewers are increasingly willing to tell a system more about what they are in the mood for, because they have been trained to expect that the exchange is worth it. That is a real change, and it is tempting to read it as a signal to make discovery conversational everywhere.
It is not. The more useful lesson from the data is about restraint. Discovery is not one problem, it is several, and generative AI is a precision instrument that earns its place in some of them and gets in the way in others. The operators who win the next few years will not be the ones who deploy GenAI most aggressively. They will be the ones who deploy it most deliberately.
generative response is not a feature. It is friction masquerading as innovation.
The skill, then, is recognising which moment you are in. High-intent searches want to be answered and gotten out of the way. But there is a second kind of moment, rarer and far more valuable, when a viewer is genuinely open to suggestion. Think about your own behavior. You finish a series you were invested in, or a film you had been meaning to watch, and you sit back and think,“ I want something different now, and I am not sure what?” Those moments do not come often. Our research puts them at roughly fifteen per cent of searches. The other eighty-five per cent, the viewer knows what they want.
But that fifteen per cent carries weight because it is precisely when a viewer is most suggestible that they are also most at risk. When people reach that open, collaborative moment and the system fails to meet it with something that lands, they do not simply shrug and scroll. Our data shows that around seventy percent of the time, an unsatisfying answer in that window leads the viewer to disengage, and often to churn. The suggestible moment is a small share of activity and an enormous share of the risk. Generative AI is uniquely suited and able to rise to the occasion, but only if the system knows it is in one. Fire it at the wrong moment and you have
The‘ why’ behind a recommendation
For over 20 years, TiVo has been a leader in personalised content discovery, licensing our recommendation technology to media and entertainment companies around the world. Through dozens of deployments, across numerous languages, we’ ve helped tens of millions of viewers find content they’ ll love.
That experience taught us an important truth: the greatest challenge in recommendations is explanation.
We could tell a viewer,“ We think you’ ll like this show.” We could rarely tell them why, in language they actually trusted.
That‘ why’ matters because it reduces friction at the exact point where viewers hesitate. Committing to a film is two hours of someone’ s evening. Committing to a series is twenty or thirty hours. That is a real investment, and before a viewer makes it, they need to feel that the system understands them well enough to be worth the gamble. This is how recommendations have always worked between people. A friend says,“ I know you, and you’ ll like this because of the specific thing you loved in that other show” is the reasoning, delivered in familiar language, that earns the trust.
Generative AI can finally hold up its end of that conversation. An LLM( Large Language Model) can explain a recommendation in the language a human actually uses, rather than the language of a data model.
Consider the industry’ s older attempts at the‘ why’: the ninety-eight per cent match score. What is a viewer supposed to do with
Most searches are not an invitation to talk
Here is the counterweight to the rising tide of conversational queries: the vast majority of searches are still what we would call high intent. Someone looking for Game of Thrones is looking for Game of Thrones. Someone looking for tonight’ s match is looking for tonight’ s match. When a viewer says“ Top Gun”, they do not want a paragraph about the cast, the director, and the plot. They want to press play. Wrapping that moment in a chatty,
14 EUROMEDIA added cost and friction to a viewer who just wanted Top Gun. ninety-eight per cent? Most people do not reason about their preferences in decimals.