Videos, Sorted: Rutube Trains AI to Classify Content With Precision
The Rutubetech team at Gazprom-Media Holding’s digital platforms technology hub has developed and deployed its own multimodal machine-learning model that automatically identifies the subject of videos uploaded to Rutube.
Rutube is a popular video platform with a growing audience. In the first six months of 2026, its monthly audience increased 5% from the same period a year earlier, exceeding 82.3 million users. Rutube’s average daily audience rose 18% to 20.4 million users. More than 448 million videos are hosted on the platform. Content classification has now become more precise, while advertisers can reach more specific target audiences.
More Than 93% Accurate
The algorithm analyzes a video’s visuals, audio track, title and description. The ML model was trained on the video platform’s large-scale dataset, helping it achieve high accuracy. According to the company, category classification accuracy exceeds 93%. At the same time, Rutube has switched to classifying content according to the Interactive Advertising Bureau’s international taxonomy, known as IAB.
The new structure includes 30 top-level topics and more granular subcategories, increasing the number of available categories roughly tenfold compared with the previous system. Instead of a broad “Automotive” category, for example, the system can now separately recognize content about vehicle repair, auto parts, motorcycles, electric vehicles and other specialized topics.
The new capabilities are already running in Rutube’s advertising algorithms. As a result, brands can place video ads in more relevant contexts, with the potential for greater effectiveness and audience engagement. A unified classification system also makes campaign performance easier to analyze by allowing results to be compared across consistent categories.
AI Becomes Essential
Rutube has steadily expanded its use of AI across the platform. In 2025, the company said it was using artificial intelligence for search, recommendations and moderation. In June, platform representatives specifically identified video recognition as a technology expected to shape the future quality of search and recommendations.
According to Rutube, its recommendation system generates more than half of all views. The new, more precise classification system should eventually help improve both search and recommendations. Rutube currently considers user actions, viewing history and video metadata. The more accurately its algorithms understand what a video actually contains, the less recommendations have to depend solely on titles, descriptions and an audience’s initial reactions. Eventually, the system could account for a video’s content even when its description contains no relevant keywords.
Other platforms use similar content-recognition technologies. In November 2025, VKontakte deployed a multimodal model for its recommendation systems. Trained on more than 3 million Russian-language pieces of content, the model can jointly analyze video, images, audio and text to determine a piece of content’s meaning and subject. In its multimodal approach, it is one of the closest Russian counterparts to Rutube’s new technology.
YouTube has also long integrated AI into video search and content analysis. The service has tested using AI to generate video collections and descriptions directly within search results.
An In-House ML Model Marks a New Stage
The broader global trend is for video platforms to move beyond relying solely on metadata and user behavior toward machine analysis of the videos themselves. Rutube is not only following that trend but also moving a critical part of the processing of hundreds of millions of pieces of content onto its own ML stack.
Launching an in-house multimodal ML model marks a new stage in Rutube’s technological development. The platform is gradually bringing automated video recognition, recommendations, search, moderation and advertising tools into a unified system.
Advertisers will feel the impact of these changes first: a more granular understanding of video topics allows content and ads to be matched more precisely. Over time, the algorithm is expected to evolve toward deeper recognition of storylines, objects, people and individual semantic segments within videos.
