Leveraging AI in OTT and CTV Content Discovery
Content abounds on today’s streaming services, yet it’s getting harder than ever for many consumers to find something to watch. Many viewers become overwhelmed by the sheer number of possibilities available to them and find themselves trapped in an unending scroll through a provider’s electronic program guide due to the so-called “paradox of choice.” However, emerging AI technologies may now incorporate user data to provide a much more personalized experience when it comes to finding content.
Although content discovery is becoming less difficult thanks to generative AI-based apps, issues remain with this developing technology, and there are also ethical concerns about using audience data to improve content discovery.
In this article, experts from key streaming vendors weigh in on both the benefits and drawbacks of leveraging AI in OTT and CTV content discovery while speculating on how it will impact the future of personalization. Various methods of filtering make AI recommendations fully transparent to users, ensuring that personalization does not become too rigid for them, and advancements in IP distribution are among the many issues under consideration in this evolving field.
How AI Is Already Changing the Ways That People Find Content
Based in Brno, Czech Republic, but with a global analytical reach, 24i delivers a strong, adaptable, data-driven video platform that uses cutting-edge technologies to generate creative revenue streams and provides a smooth content discovery experience.
“Artificial intelligence is redefining how audiences discover content,” says 24i CEO Sebastian Braun. “At 24i, we’ve seen firsthand, through projects like our collaboration with Crossmedia and RTVE Play, how AI-driven personalization transforms streaming from a passive browsing exercise into an intuitive, deeply engaging experience. But while the technology is powerful, its value lies in how thoughtfully it’s applied.”
“At 24i, we’ve seen first-hand… how AI-driven personalization transforms streaming from a passive browsing exercise into an intuitive, deeply engaging experience. But while the technology is powerful, its value lies in how thoughtfully it’s applied.”—Sebastian Braun, CEO, 24i
Headquartered in Columbia, Md., LTN enables clients to monetize live sports and news more successfully, create new channels in a matter of hours rather than weeks, and automate content development for increased workflow efficiency when compared to VOD-only options. “AI is making big strides in content discovery and personalization,” says LTN chief strategy officer Roger Franklin, “streamlining user experiences and helping major content providers aggregate services and surface relevant recommendations across platforms. Personalization has become a fundamental part of the modern media experience. Almost every interaction with content today is personalized to some degree, from targeted advertising to local language commentary, customized audio tracks, tailored graphics, and data overlays in live sports. For rightsholders and content providers, personalization is essential to maximizing the ROI on costly content and sports rights, while engaging audiences with experiences that feel relevant and valuable. The traditional ‘one-to-many’ model of distribution has been replaced by a new reality in which platforms must deliver tailored, platform-specific live experiences to diverse audiences at scale.”
“The traditional ‘one-to-many’ model of distribution has been replaced by a new reality in which platforms must deliver tailored, platform-specific live experiences to diverse audiences at scale.”—Roger Franklin, Chief Strategy Officer, LTN
With its global headquarters in India and 12 additional offices worldwide, Amagi is a SaaS provider that uses cloud-native technologies to link media firms with their audiences. According to Srinivasan KA, Amagi co-founder and president of global business, “AI algorithms now drive most content discovery on streaming platforms, with most of the viewing coming from personalized recommendations rather than manual search. These systems analyze each viewer’s habits and preferences alongside content attributes to surface what someone is likely to enjoy next. For example, generative AI can automatically enrich a title’s metadata with multilingual summaries, keywords, and other descriptors, making shows more searchable and discoverable across different regions.”
The Differences Between Content-Based Filtering and Collaborative Filtering
The two primary categories of today’s top recommendation systems are content-based filtering, which matches content qualities to auser’s past, and collaborative filtering, which looks for similarities in the behavior of like-minded users. Taking a hybrid approach may be the most effective way to reap the benefits from each method.
“At 24i, we don’t believe in choosing one over the other,” says Braun. “Instead, we blend multiple models, from similarity-based and metadata-driven suggestions to popularity-based algorithms. This hybrid approach ensures viewers discover both the content they expect and the surprises they didn’t know they’d love.”
Headquartered in South Korea but focusing on the U.S. market, SpoonLabs offers Spoon, a platform that makes it simple for anybody to start a live audio broadcast and that is growing into an audio content platform. In July 2024, SpoonLabs launched Vigloo, a video platform that features short-form programs of roughly 2 minutes, which became live worldwide in seven languages.
“Collaborative filtering leverages the preferences and behaviors of similar user groups to create recommendations, while content-based filtering analyzes specific attributes of content to suggest titles similar to what a user has already enjoyed,” says SpoonLabs machine learning engineer Hayden Sang-won Ha. “In streaming environments, audiences are often trend-sensitive, so collaborative filtering usually performs better.”
“Collaborative filtering bases recommendations on user behavior patterns—essentially leveraging the collective tastes of similar users—while content-based filtering relies on item attributes (like genre, cast, or storyline) to recommend titles similar to those a user already likes,” says Amagi’s Srinivasan. “In streaming, neither method alone is a silver bullet; a hybrid approach is often most effective, combining crowd wisdom with AI-driven content understanding (for example, using enriched metadata and knowledge of content themes) to deliver well-rounded recommendations.”
The ‘Cold Start’ Dilemma
The term “cold start” in streaming can relate to two different issues: a lack of data for recommendation engines and serverless services’ performance delay. How can these issues best be tackled?
“For new users, simple onboarding steps such as a short survey or prompting them to choose preferred genres can provide valuable initial signals to mitigate the cold start problem,” says SpoonLabs’ Sang-won Ha. “For new content, extracting key information—like genre, cast, or themes—and giving it greater weight within content-based filtering ensures that fresh titles are surfaced to the right audiences more effectively.”
“For new content, extracting key information—like genre, cast, or themes—and giving it greater weight within content-based filtering ensures that fresh titles are surfaced to the right audiences more effectively.”—Hayden Sang-won Ha, Machine Learning Engineer, SpoonLabs
“Every recommendation platform faces the ‘cold start’ problem: How do you serve great recommendations to a brand-new user or surface a brand-new title?” says 24i’s Braun. “Our strategy is to combine smart defaults (editorial picks, trending content) with rapid feedback loops. In RTVE Play’s case, we implemented a system that adjusts in real time based on user interactions, so relevance improves with every click. It’s about shortening the path from ‘just joined’ to ‘just one more episode.’”
Srinivasan says, “We tackle the cold start challenge by leaning on content understanding and broad trends when specific user data is sparse. For a new user, the system might start with popular or editor’s pick titles—possibly after a quick preference input—and then rapidly adapt as soon as the viewer begins watching and providing feedback. As part of the Smart Scheduler offering, we tend to include social trends and cast details such as birthdays to [provide] better recommendations. For a brand-new piece of content, we rely on content-based signals: AI-generated metadata (like summaries, keywords, and genre tags) allows us to connect a new show or movie with the right audience from Day 1, even before it has an engagement history.”
How Designing AI Algorithms for User Data to Create Personalized Content Raises Ethical Questions
Several ethical concerns persist when using AI algorithms that incorporate user data to provide tailored content.
“With great personalization power comes great responsibility,” declares 24i’s Braun. “We’re committed to designing AI systems that are fair, transparent, and inclusive—avoiding algorithmic bias and ensuring that personalization serves all audiences, not just the majority profile. Ethical AI is not an optional extra—it’s the foundation for sustainable viewer trust.”
24i’s mission, Braun explains, “is to help broadcasters, OTT platforms, and pay TV providers deliver personalized content and ad experiences that are as ethical as they are effective. AI-driven discovery isn’t just about serving ‘more of the same.’ It’s about connecting people with content that inspires, informs, and delights, while safeguarding the trust that keeps them coming back. When we get it right, the future of streaming will be less about finding something to watch and more about feeling like the platform knows you, respects you, and helps you discover the stories that matter most.”
How Can AI-Driven Recommendations Be Made More Transparent or Explainable to Users?
Providing transparency regarding how recommendations are served to users is essential to guarantee continued trust in a service, along with helping users attain a clear understanding of how to use the search features.
“We know that users are more likely to engage with recommendations they understand,” Braun says. “That’s why explainability is key. Imagine a small label under a suggested show: ‘Because you enjoyed Series X and follow Actor Y.’ In our future-facing designs, transparency will be a default—not just a compliance requirement—helping users feel in control rather than manipulated.”
Amagi’s Srinivasan also emphasizes clarity and purpose to help build a foundation of user trust. “We believe users should understand why something is recommended to them,” he says. “In practice, this means giving simple, human-readable cues—for instance, explicitly
noting, ‘Because you watched X’ or pointing out that a recommended title shares a lead actor or theme with something the user enjoyed. We can even leverage generative models to create more user-friendly explanation labels in the interface to convey why a title is being suggested in an engaging way. By highlighting key factors like genre, cast, social signals, or mood overlap, we can make the recommendation process more transparent and build user trust.”
“We believe users should understand why something is recommended to them. By highlighting key factors like genre, cast, social signals, or mood overlap, we can make the recommendation process more transparent and build user trust.” —KA Srinivasan, Co-Founder and President – Global Business, Amagi
The Emerging Technologies That Will Shape the Next Generation of AI in Streaming Personalization
Multimodal AI, along with rich and multifaceted forms of data, make up some of the emerging technologies that will shape AI’s next generation of streaming personalization.
“Looking ahead, multimodal AI—capable of analyzing video, audio, and text simultaneously—will transform how we understand content,” says SpoonLabs’ Sang-won Ha. “For example, instead of classifying Vigloo vertical dramas only by genre and themes, we could uncover their mood and subtle scene characteristics. Building on this richer content understanding, we can then map user preferences and behaviors with greater precision, enabling truly granular personalization.”
Braun outlines the ways that 24i has worked to harness data successfully. “The most effective recommendations come from rich, well-structured data,” he says. “This includes not only viewing history but also content metadata, contextual signals like time of day or device type, and real-time behavioral insights. Our work with RTVE integrates multiple data sources—including SAP CDC for user management and Conviva’s ECO technology for real-time activity capture—into 24i’s centralized AI personalization engine. This allows content to be tailored with precision, across video, audio, news, and podcasts.”
Srinivasan breaks down the ways that Amagi is implementing and adapting these emerging AI technologies. “We see AI getting better at deeply analyzing video and audio content to autonomously create useful outputs—for example, generating highlight reels or multilingual metadata—which can be used to tailor the viewing experience on-the-fly. Another exciting area is AI-driven content presentation: systems that automatically extract the best keyframes and generate platform-specific thumbnails for each show or even create personalized channels and dynamic playlists for individual users,” he explains. “Even the viewing experience itself is evolving; for instance, we’re innovating toward a ‘zero-slate’ experience where adaptive, session-aware AI ensures that viewers aren’t interrupted by filler slates in ad breaks. All of these advancements point to a future where personalization is more holistic—seamlessly blending what content is recommended, how it’s packaged visually, and how it’s delivered to keep each viewer maximally engaged.”
How Generative AI (e.g., LLMs) Will Play a Central Role in the Future of Content Discovery
Large language models (LLMs) and other forms of generative AI are set to play a crucial role in content discovery.
“Generative AI models like large language models are poised to become central by enabling a much deeper semantic understanding of both content and user intent,” Srinivasan says. “We already use LLM-based techniques to automatically summarize and tag content, essentially allowing the AI to ‘read’ scripts or audio and extract rich context such as plot points, characters, and themes. This means recommendations and search can go beyond simple keywords, connecting viewers with content in a more nuanced way (for example, finding a show that matches the mood or narrative style a user is looking for). Moreover, LLMs open the door to conversational discovery—imagine a viewer asking an AI assistant for ‘a light-hearted crime series with witty dialogue’ and the system understanding that request and delivering spot-on options. In short, generative AI acts as a smart bridge between the unstructured world of content and the natural language of the viewer, making discovery more intuitive.”
“Imagine a viewer asking, ‘Show me a gripping crime drama set in the 1970s, but without excessive violence’ and receiving a curated shortlist instantly,” says 24i’s Braun. “These conversational, intent-based interactions can break free from the rigid structures of current UIs, making discovery more natural and human.”
24i’s long-term plans, Braun adds, align “with the seamless streaming experience we’ve described in our ‘Streaming in 2035’ series. AI will evolve from simply recommending content to orchestrating an entire entertainment journey, adapting not just to a user’s history, but to their mood, schedule, and even contextual signals from their environment. This could mean universal recommendation engines working across all of a user’s subscribed services; value-added interactivity, where a viewer can instantly learn more about an actor, purchase an item seen on screen, or jump into related bonus content without leaving the story; and human-centered design that anticipates needs before the user even reaches for the remote.”
The Risks of Personalization Becoming an Echo Chamber
Ensuring that personalization does not become too rigid for users—in the sense of hyper-optimizing preferences to the point of creating an echo chamber—is another major consideration.
“At Vigloo, we deliver optimized personalization while also keeping about 10%–20% of content diverse,” Sang-won Ha says. “Our recommendations go beyond simply reflecting a user’s viewing history; they are further optimized by taking into account factors such as a title’s popularity, the attributes of content the user has previously enjoyed, and detailed elements like genre, theme, and synopsis. This ensures that while users are primarily shown content aligned with their preferences, they also have the opportunity to discover and engage with new genres or unexpected finds. By striking this balance between personalization and diversity, Vigloo enhances the overall content experience, offering both seamless entertainment and fresh discoveries.”
“Personalization goes too far when it starts to constrain or repeat a viewer’s content universe instead of expanding it,” Srinivasan says. “If someone feels like they’re seeing the same types of shows over and over—or if the recommendations never challenge them with something new—then you’ve entered echo chamber territory. The goal is a balanced diet of content: highly relevant suggestions blended with the occasional discovery that broadens horizons. In the end, effective personalization should help viewers find content they love without making their world of content feel small. Each user’s experience should remain rich and exploratory, not just an endless loop of ‘more of the same.’ ”
“One of AI’s biggest risks is overfitting, showing viewers only what they already like,” Braun says. “The art of recommendation is to balance familiarity with exploration. We design algorithms that deliberately inject diversity, surfacing content from different genres, creators, and even cultural contexts. This keeps viewing fresh, helps reduce churn, and supports content discovery for longtail titles.”
Advancements in IP Distribution
Advancements in IP distribution especially interest LTN’s Franklin. He considers the many dynamic ways that this drives hyper-personalized engagement for users and unique targeting opportunities for advertisers. “IP distribution has unlocked new levels of customization, freeing content providers from the restrictions of satellite ‘world feeds,’ ” he says. “With IP-native workflows, a single source feed can be easily versioned into multiple tailored outputs, integrating local-language commentary, market-specific graphics, regional ad insertion, or even player-focused feeds. For example, sports leagues are increasingly creating alternate versions of live events: One feed might highlight a star player like Lionel Messi for dedicated fans, while another delivers localized commentary and targeted ads to a specific market.”
Franklin argues that the benefits of this approach for streaming audiences, content providers, and advertisers alike are self-evident: “Viewers enjoy richer, more relevant experiences that deepen their engagement, while advertisers and platforms gain new opportunities to drive revenue through hyper-targeted campaigns and tailored ad placements. Sports leagues like the World Surf League and Major League Volleyball and major broadcast groups and national networks like the Tennis Channel are already harnessing IP-native production and distribution tools to bring culturally resonant, localized coverage to global audiences at scale and in real time. Intelligent, automated IP distribution workflows make it possible to deliver infinitely customizable, cost-effective experiences that grow audiences and revenues.”
Related Articles
In today's saturated content market, even on the downward slope of the peak TV era, offering great content isn't enough to deliver satisfying experiences to streaming users. Join media industry thought leaders from Roku, Tubi, Cineverse, and the SVTA at Streaming Media 2025 on October 8 in Santa Monica for a must-see panel on AI-Powered Curation and Discoverability that will explore how machine learning and personalization facilitate more effective curation and discovery and increasingly shape the viewer experience.
06 Aug 2025
On Thursday, November 14, media and entertainment technology strategist Chris Pfaff will moderate the panel "Removing the Friction of Content Discovery on CTV." Incorporating viewership data into the content discoverability process can remove the pain points that have bubbled up in the new streaming landscape, lead to more engagement with the content, and open up more opportunities for advertisers to reach their target audiences. This panel of experts from NBCUniversal, Pluto TV, Revry TV, and Gatsby TV looks at the success of genre channels and provides attendees with insight into how they help their viewers find their content faster.
30 Oct 2024
Gatsby TV CEO Gatsby Frimpong has created a streaming programming discovery app that applies AI and user recommendations to help viewers discover what to watch next on services like Netflix, Hulu, Prime, Disney, and more. In this interview with Streaming Media's Tyler Nesler, he discusses his professional background, what led him to create the app, how it helps customers save money, the AI foundation model it uses, and more.
30 Aug 2024