-->

Fix the Catalog Before You Personalize It

Article Featured Image

Most streaming services have spent years learning more about their viewers. The next gains in recommendations will come from learning more about their titles.

Personalization has two halves. One is what a service knows about the viewer: what they watched, finished, abandoned and searched for. The other is what it knows about the title. Over the past decade, most of the effort has gone into the first half. Services have become very good at describing behavior. Many are still poor at describing content.

That imbalance now limits what personalization can do. A recommendation engine can only match a viewer to a title along the dimensions it knows about. If every title is described by genre, cast and a two-line synopsis, the engine can recommend another thriller. It cannot tell a slow, atmospheric thriller from a loud, fast one, and for most viewers those are not substitutes.

The gap is widest exactly where personalization should help most. A new release, or a film that arrived last week in a licensing deal, has no viewing history yet. Until enough people have watched it, the only thing a recommender can go on is what the service knows about the title itself. If that is a genre and a synopsis, the title is close to invisible in the weeks that decide whether it finds an audience.

Where the catalog lets personalization down

Genre is too coarse. A genre label sorts a catalog the way a video store sorted its shelves. It says nothing about how a title feels to watch: “Comedy” covers a gentle family film and a dark satire. A viewer who has just finished one is rarely in the mood for the other.

Metadata is inconsistent. Licensed catalogs come from many suppliers, and each has its own house style. One calls it “Sci-Fi”, another “Science Fiction & Fantasy”; one writes three paragraphs of synopsis, another a single line, and the oldest titles may arrive with almost nothing. A model trained on that data learns the quirks of each distributor as much as the qualities of each title.

Older titles are the worst described. That matters because audiences watch them. Luminate’s midyear report found that on major U.S. services, library titles drew nearly four times as many streaming hours as new originals in the first half of 2026. The part of the catalog that carries most of the viewing is often the part the recommender understands least.

Describe titles the way viewers experience them

The fix is to describe every title in one vocabulary, at the level where viewers actually respond: mood, pacing, tone and story. How a service gets there matters less than the discipline behind it. Some start with editors tagging their most-watched titles; most libraries are too large for that alone and add automated analysis of the video, audio and script. Either way, three things decide whether it works:

  • A vocabulary people can use. Small enough for editors to apply consistently, rich enough to separate a slow thriller from a fast one.
  • A test set. A few hundred titles that editors have described by hand, so automated descriptions can be checked against human judgment before anyone relies on them.
  • One source of truth. If the recommender, the search index and the editorial tools each read a different version of the metadata, the work is wasted.

When titles are described this way, the catalog becomes consistent: a film from 1985 and a series from last month are described in the same terms, whatever their source. And recommendations can match on feel, not just category. A viewer who binged a warm, low-stakes comedy can be offered something with the same mood from a different genre.

There is evidence that it pays. Plex, which runs an ad-supported service, built a consistent layer of mood data across more than 50,000 titles and tested it. “We have seen significant uplift in minutes based on using mood data properly,” says Scott Olechowski, Plex’s co-founder and chief product officer. The word that matters is “properly”: descriptions pay only when they are consistent and actually used.

Where this does not help

Better descriptions do not replace viewer data. Knowing that a title is tense and fast-paced only helps if you also know who wants tense and fast-paced tonight. The two halves work together.

Description is also harder than it looks. Dark comedy sits awkwardly in any vocabulary, and a slow film can be slow because it is gentle or because the tension is building. Automated systems should say how confident they are and send uncertain titles to an editor rather than guess.

And it costs money. Not every title justifies the effort. Start where the viewing is: the library titles people already watch, and the new arrivals that have no history yet.

Five questions before your next personalization project

  1. Would two similar titles from different suppliers be described in the same terms?
  2. What do our recommendations know about a title beyond genre, cast and synopsis?
  3. How are new and newly licensed titles described in their first weeks, before anyone has watched them?
  4. How well described are the older titles that carry most of our viewing?
  5. Do our recommender, search and editorial tools all read the same metadata?

The industry has spent a decade perfecting its picture of the viewer. The services that now give their titles the same attention will get more out of every model they already run, and more out of the library they already own.

[Editor's note: This is a contributed article from Vionlabs. Streaming Media accepts vendor bylines based solely on their value to our readers.]

Streaming Covers
Free
for qualified subscribers
Subscribe Now Current Issue Past Issues
Related Articles

Personalization Is a Continuous Capability

Many still treat personalization as a problem that can be fixed once and for all by deploying an AI recommendation engine. There's a tendency to think that once a personalization model has been implemented, the personalization requirement is met, and the model can be left alone to run. When actually, what services need to understand is that personalization is an ongoing capability, not a one-off problem to be solved.

Elevating Content Verticalization With Emotion-Aware AI: A Q&A With Vionlabs' Arash Pendari

Vionlabs makes full content libraries "discoverable, engaging, and revenue-ready" using its "emotion-aware" AI. It's a company paying attention to the shift toward vertical and mobile-first viewing, including the micro-drama boom. Content creators that work with Vionlabs get the verticalization of long-form content into chapters and micro-episodes, preview clips for every title that can be used for social media and promotion, scene-level metadata that surfaces content to match viewers' moods, and non-intrusive ad breaks with contextual targeting. This Q&A with Vionlabs Founder Arash Pendari explores his thoughts on the use of AI in the streaming industry and how Vionlabs' platform works. 

Personalization Is More Than What You Think It Can Be

With advances with generative AI, just-intime transcoding, SSAI stitching, and other streaming video tech stack components, companies like Infuse Video are demonstrating that the true vision of video personalization—changing the video content itself—is finally at hand.