What If AI Becomes the New Discovery Layer?
It's Not Just About Creation, But Coordination
The show Suits, which aired on USA Network from 2011-2019, was moderately successful for a cable drama. In most years of its run, it was one of USA Network’s top three-rated shows, but by 2019, its ninth and final season, it generated a 0.2 rating in the A18-49 demo, with about 1 million viewers per episode. For reference, a show needed to pull a 0.9 rating to make it into the top 100 shows on TV that year. Vanderpump Rules, Love & Hip Hop 6, and Basketball Wives 8 all squeaked into the top 100. Suits wasn’t close.
In 2023, four years after the show ended, Netflix acquired the streaming rights (which it shared with Peacock) and made all the seasons available. That year, Suits was watched for 58 billion minutes in the U.S., making it the most popular show on streaming and probably the most popular series on all of TV. This surprising resurgence is often referred to as the “Suits phenomenon.”
Why was the same show that was a middling success on USA a blockbuster on Netflix? The simplest answer: Netflix put it in front of tens of millions of viewers and made it easy to choose. In today’s fractured media environment, choosing is hard for consumers. The largest media platforms generate a lot of their value by partially solving this problem. Netflix, Spotify, YouTube, and TikTok all capitalize on the overwhelming difficulty of choosing. The Suits phenomenon showed that coordination—connecting viewers to content—can be as valuable as the content itself.
So, what if AI usurps that role?
Tl;dr:
When people talk about the effect of GenAI in media, they usually mean that falling costs of creation will lead to an explosion of supply. But GenAI can do more than just create. It can coordinate too.
The biggest platforms in media—Netflix, Spotify, YouTube—don't just win on content. They win on coordination: making it easy to find and choose in an environment of overwhelming choice.
Discovery is utterly broken across much of the media landscape and will get worse. A GenAI recommendation layer that operates above the platforms is likely because it could be much better. It could understand intent semantically, recommend across formats and services, incorporate context, and be better aligned with users’ goals. Platforms may adopt AI too, but they likely can’t or won’t surface recommendations beyond their own apps or understand broader context.
Some platforms would hate this, but there might not be much they can do to prevent it.
This idea raises a lot of questions, like who is best positioned to do it, whether it will become a new chokepoint, and how it will affect consumption.
Watch this space. As consumers increasingly turn to AI to help manage parts of their lives, it’s hard to see why content consumption wouldn’t be on this list.
Lowering the cost of creating content will undoubtedly redistribute value in media, but lowering the cost of coordinating it could too. In other words, GenAI will not only affect the battle over who makes the content, but who controls how consumers choose.
GenAI Can Do More Than Create
GenAI models “understand” information, in the sense that they mathematically capture and represent the relationships between units of information. (For a more complete explanation, please see the Appendix.) What can these systems do when they understand meaning? Here are four things:
Create
This is what you think of when you think of GenAI. GenAI models make new combinations of bits: new prose, music, code, proteins, images, and video. This is what Claude, Sora, Midjourney, Suno, and Cursor do.
Coordinate
By coordinate, I mean make information more accessible and useful. This one is less well understood, especially in traditional media circles.
For our purposes, you can think of data as falling into two broad categories: authoritative and semantic metadata. The former is what something is, the latter is what it means.
Authoritative metadata is ground truth—facts established by a recognized source whose authority to assert them is what makes them true. These include unique IDs, legal and contractual information, provenance, or chain of custody. It has a single correct answer that must be maintained in a registry or database.
Semantic metadata is interpretive, like themes, mood, tone, genre, audience fit, cultural resonance, or similarity to other content. It emerges from the content itself and the conversation around it. There is no single correct answer, just better and worse answers.
Historically, creating usable semantic metadata required a lot of upfront work: cleaning it, imposing taxonomies, creating metadata schemas, normalizing fields, building relational databases, and establishing classification systems. It is expensive, relatively rigid (i.e., hard to change once designed), and platform- or system-specific.
For instance: Salesforce structures customer data into standardized objects and fields. LexisNexis catalogs legal documents with intricate indexing systems. Factiva organizes news articles with subject codes. Netflix tags content into microgenres, and so on. (According to this article, Netflix has 30 employees who are taggers and 3,000 different tags.)
While digitization unbundled information from the underlying physical infrastructure, GenAI further unbundles information from the data structure.
By contrast, AI can impose structure at query time rather than requiring it upfront and extract meaning from a big ol’ pile of stuff. It can work directly with unstructured data: raw text, emails, documents, web pages, images, conversations—it doesn’t need pre-tagged fields or pre-ordained vocabularies. It can analyze customer interactions from unstructured emails and Slack messages without Salesforce’s data model; read a legal brief and understand its arguments without LexisNexis’s indexing; and identify relevant news by understanding content and context, not just matching keywords against Factiva’s codes. Put differently, while digitization unbundled information from the underlying physical infrastructure, GenAI further unbundles information from the data structure.
AI can infer what something means much better than humans, but it can’t infer what something is.
AI can’t supplant authoritative data. It could, for instance, pull up a book’s ISBN code from the database maintained by the International ISBN Agency, but it couldn’t infer it. It can enrich that foundation by deriving and adding contextual layers at a scale and granularity that human tagging can’t match. Applied to media, this also means that it can surface and recommend content across Netflix, Apple TV+, Bravo, and YouTube—not to mention Spotify, Apple Podcasts, The New York Times, Steam, Substack, X, LinkedIn, and Audible—based on a semantic understanding of your interests, context, goals, prior behavior, social listening, reviews, and a whole lot else. But we’ll get to that.
