Generative Recommender
A Generative Recommender is a recommendation system that predicts a user's next click or purchase by continuing their past usage history like a piece of text. Instead of scoring ready-made suggestions by similarity, it generates the recommendation itself, step by step.
When you watch a video on a video platform, the site immediately suggests the next one. Such suggestions come from a recommendation system, that is, a program that infers from your past behavior what might interest you. A Generative Recommender is a newer design of such systems. It treats your history list like a sentence: first video A, then B, then C. And just as writing programs predict the next word, it predicts the next entry in this list. So it doesn’t select from a list, but generates the answer itself.
Why platforms are rebuilding their recommendation logic
Classic recommendation systems work in two stages. First, they pick out a few hundred candidates that might suit you. Then a second model evaluates each candidate individually and ranks them. This works, but it’s costly, because a separate calculation is needed for each candidate.
Generative Recommenders partially eliminate this second pass. They deliver the most likely next items directly from a single pass. Above all, though, they benefit from an effect known from language models: more data and more computing power reliably make them better. With the older systems, these gains eventually leveled off.
Economically, this is a major lever. At services like Instagram, TikTok, or Amazon, a considerable share of usage time and revenue is generated through recommendations. In 2024, Meta demonstrated with an architecture called HSTU that a generative approach can significantly boost click-through rates. Since then, the topic has regularly appeared in quarterly reports and tech news.
From click history to prediction
The first step is to translate each product or video into a code. So-called semantic IDs are often used for this: short sequences of numbers that describe the content. Two cooking videos then get similar codes, while a cooking video and a racing game get very different ones. This is important because it allows the system to also handle new items that no one has clicked on yet.
After that, your history becomes a chain of such codes. The model learns from millions of other people’s histories how such a chain typically continues. At runtime, it then generates the next entry code by code. In the end, the generated code is mapped back to an actual product in the catalog.
A common misconception is that there’s a chatbot behind this. That’s not true. The Generative Recommender doesn’t generate language, but item identifiers. It merely borrows the design of language models, namely predicting the next element in a sequence.
Where generative suggestions are already in use
The technology is most visible in feeds that scroll endlessly. Short-video apps, music streaming, and large online shops are the typical fields of application. The difference often becomes noticeable in that suggestions react faster to fresh behavior. If you watch two guitar videos, the topic sometimes already shows up on the very next screen.
In the news, the term usually appears alongside Meta, Google, ByteDance, or Netflix. This is often about investments in data centers, since these models are expensive to operate. It’s just as often about criticism: systems that optimize even more precisely for short-term stimuli may reinforce filter bubbles and addictive behavior.
For you as a user, the technology remains invisible. Only the consequences are visible, namely more fitting suggestions and more time spent in the app. That’s exactly why it’s worth knowing that behind the next video lies no coincidence, but a very targeted prediction.