Meta shifts Facebook Reels beyond engagement metrics with new user survey model

Meta today published research revealing how Facebook Reels moved beyond traditional engagement metrics to directly integrate user feedback through surveys, achieving significant improvements in content matching and platform retention. The company's engineering team detailed the technical framework in a blog post and accompanying research paper published January 14, 2026.

The approach centers on what Meta calls the User True Interest Survey (UTIS) model, which collects randomized in-feed survey responses from users asking "How well does this video match your interests?" on a five-point scale. According to the announcement, this methodology emerged from recognition that conventional signals like watch time and likes failed to capture what people genuinely want to see. Traditional interest heuristics achieved just 48.3% precision in identifying true interests, the research found.

"Traditional recommendation systems often rely on engagement signals—such as likes, shares, and watch time—or heuristics to infer user interests," according to the announcement. "However, these signals can be noisy and may not fully capture the nuances of what people actually care about or want to see." The research indicates that effective interest matching encompasses factors beyond topic alignment, including audio quality, production style, mood, and motivation.

The engineering team deployed surveys across Facebook Reels and other video surfaces, collecting thousands of daily responses from randomly selected viewing sessions. Meta weighted these responses to correct for sampling and nonresponse bias, building what the company characterizes as a comprehensive dataset reflecting real user preferences rather than implicit engagement signals.

Technical architecture and implementation

The technical implementation involves training a lightweight alignment model layer using existing predictions from Meta's main candidate ranking model as input features. The UTIS model operates in parallel to existing recommendation systems rather than replacing them, providing probability scores that videos will satisfy user interests.

According to the documentation, Meta binarized survey responses for modeling purposes while engineering new features to capture user behavior, content attributes, and interest signals. The model outputs are designed for interpretability, allowing the engineering team to understand factors contributing to interest-matching experiences.

Integration into the ranking funnel occurred across multiple stages. In late-stage ranking, UTIS provides an additional input feature into the final value formula, allowing fine-tuning while balancing other considerations. Early-stage retrieval systems use UTIS to reconstruct user true interest profiles by aggregating survey data, enabling re-ranking and sourcing of candidates more relevant to genuine interests.

Large sequence-based user-to-item retrieval models received alignment through knowledge distillation objectives trained on UTIS predictions from late-stage ranking. The architecture allows videos predicted as high interest to receive modest ranking boosts while low predicted interest content faces demotion.

Performance metrics and testing scale

Meta reported offline accuracy improvements from 59.5% to 71.5% compared to heuristic rule baselines. Precision increased from 48.3% to 63.2%, while recall jumped from 45.4% to 66.1%. These gains demonstrate improved ability to identify user interest preferences, according to the announcement.

Large-scale A/B testing involved more than 10 million users. The testing confirmed real-world performance improvements, with the UTIS model consistently outperforming baseline approaches. High survey ratings increased 5.4% while low ratings decreased 6.84%. Total user engagement rose 5.2%, and integrity violations decreased 0.34%.

The results aligned with broader performance trends Meta has reported throughout 2025. The company's advertising revenue reached $46.6 billion in the second quarter, representing 22% year-over-year growth driven substantially by AI-powered recommendation system improvements. The Generative Ads Recommendation System improved ad conversions by approximately 5% on Instagram and 3% on Facebook Feed and Reels during that period.

Meta's focus on Reels intensified throughout 2025, with the format reaching a $50 billion annual run rate by the third quarter. Video time spent on Instagram increased more than 30% year-over-year. Time spent on Facebook increased 5%, driven by AI recommendation system improvements that now surface twice as many Reels published that day compared to the start of the year.

Impact on content distribution patterns

The UTIS integration produced observable shifts in content distribution. Meta reports increased delivery of high-quality niche content alongside reduced distribution of low-quality generic popularity-based recommendations. Metrics including likes, shares, and follow rates improved following implementation.

These changes address longstanding challenges in recommendation systems that tend to favor content with high short-term engagement measured by watch time and interactions but fail to capture interests important for long-term product utility. Models trained exclusively on engagement signals often miss dimensions of content relevance that drive sustained platform use.

The survey-based approach provides measurement capabilities beyond what implicit signals can deliver. According to the research, interest matching involves not just topic alignment but also audio characteristics, production techniques, emotional tone, and viewer motivations. These elements prove difficult to infer from binary actions like clicks or shares.

Facebook has emphasized content quality throughout 2025, introducing stronger enforcement against unoriginal creators in July. Platform monetization access depends on policy compliance, with creators violating guidelines losing earning opportunities through Performance Bonus, In-stream ads, and Ads on Reels programs. Content recommendation algorithms received updates to better identify mass-produced material, evaluating video fingerprints, metadata patterns, and engagement metrics to distinguish legitimate content from repetitive posting.

The quality emphasis extends to advertising contexts. Meta previously rolled out updated video players prioritizing short-form content in early 2024, with improved recommendations surfacing more relevant videos based on user interests regardless of length. The platform accommodates all video formats while clearly prioritizing Reels to compete with TikTok and YouTube Shorts.

Industry context and competitive landscape

The UTIS development occurs within intensifying competition in short-form video markets. Research released in November 2025 found YouTube Shorts leading consumption at 56%, ahead of TikTok and Facebook at 50% each. Instagram Reels captured 41% of usage among survey respondents. The study found 73% of consumers watch short-form video multiple times daily, with 81% viewing primarily on smartphones in vertical format.

Federal court proceedings in November 2025 documented how Reels transformed Meta's platforms. Judge James Boasberg determined that most time Americans spend on Facebook now involves watching videos, with Reels becoming the single most-used part of the application. Similar patterns emerged on Instagram, where Reels consumption accounts for the largest share of user time. The court noted that Meta's Reels, TikTok's videos, and YouTube's Shorts are "virtually—and deliberately—indistinguishable in function and user experience."

Meta invested approximately $4 billion annually in Reels development specifically to counter competitive pressure from TikTok, court records indicated. CEO Mark Zuckerberg told investors that "competition has gotten more intense, especially with the rise of TikTok which is one of the most effective competitors we have ever faced." The company added Reels to Instagram in 2020 and Facebook in 2021, directly copying TikTok's format.

Performance data suggests Reels outperforming other video formats. Research from Emplifi found Instagram Reels generating six times the reach of Instagram Stories and more than three times the median views of other Instagram video content. Facebook Reels saw more than three times the number of median views as other video content on the platform. Longer Reels over 90 seconds generated more than double the median views compared to TikTok videos.

Remaining challenges and future development

Meta acknowledged several limitations requiring ongoing work. The announcement identified better serving users with sparse engagement histories as a priority area. Survey sampling and delivery bias reduction represents another focus, alongside further personalization for diverse user cohorts and improved recommendation diversity.

These challenges reflect broader tensions in personalization systems. Users with limited interaction history present difficulties for models trained primarily on behavioral signals. New users or those who rarely engage provide insufficient data for accurate interest profiling through traditional methods.

Survey-based approaches introduce their own complications. Response rates vary across user populations, potentially skewing results toward more engaged or opinionated users. The announcement noted Meta's use of weighting techniques to address these issues, though bias correction in survey sampling remains an active research area.

The company indicated exploration of advanced modeling techniques including large language models and more granular user representations. Meta planned to use AI chat data for ad targeting starting December 16, 2025, affecting over 1 billion monthly users. Conversations with Meta AI became another data source for personalizing content recommendations and advertisements across platforms.

The personalization update built on extensive AI infrastructure investments. The company's advertising revenue growth has been substantially driven by AI-powered recommendation improvements throughout 2025. Ad impressions delivered across Meta's Family of Apps increased 11% year-over-year in the second quarter, while average price per ad rose 9%.

Implications for content creators and advertisers

The UTIS integration carries implications for creators developing content for Facebook Reels. The system's emphasis on interest matching beyond simple engagement metrics suggests that niche content aligned with specific user preferences may achieve improved distribution compared to generic viral content optimized purely for watch time or likes.

Creators producing content for particular interest communities could benefit from more precise audience matching. The survey-based measurement captures dimensions including audio style, production techniques, and emotional tone that traditional metrics overlook. Content resonating deeply with smaller audiences on these factors may receive distribution advantages.

For advertisers, the improvements in recommendation quality correlate with engagement and retention gains that affect advertising inventory value. Higher user retention and engagement typically translate to increased advertising inventory availability and improved campaign performance through better audience attention and receptivity.

Meta has expanded advertising capabilities across Reels throughout 2024 and 2025. The platform introduced AI-powered video tools including Video Expansion for automatically fitting images and videos to vertical format, and Image Animation for adding dynamic movement to static images on Instagram Reels. These tools address challenges faced by advertisers with limited video resources.

The technical documentation emphasizes that UTIS scores represent one input among multiple factors in ranking systems. Videos receive modest boosts or demotions based on predicted interest rather than wholesale reordering of content. This balanced approach aims to improve relevance while maintaining other platform objectives including content diversity, creator monetization, and policy compliance.

The research paper accompanying the announcement provides additional technical detail for engineering teams working on recommendation systems. Meta has consistently published technical research throughout 2025, sharing implementation details and results from production systems serving billions of users daily.

Timeline

• January 14, 2026: Meta publishes engineering blog post detailing User True Interest Survey model implementation
• December 16, 2025: Meta begins using AI chat data for ad targeting affecting 1 billion monthly users
• November 18, 2025: Federal court dismisses FTC antitrust case, documenting Reels as most-used Facebook feature
• October 30, 2025: Meta reports Reels reaching $50 billion annual run rate with video time on Instagram up 30% year-over-year
• July 31, 2025: Meta reports 22% advertising revenue growth driven by AI-powered recommendation system improvements
• July 15, 2025: Facebook introduces stronger measures against unoriginal content creators affecting monetization and distribution
• April 3, 2024: Facebook rolls out updated video player prioritizing short-form Reels content

Summary

Who: Meta's engineering team including researchers Senthil Rajagopalan, Drew Hogg, Mengxi Lv, Grace Yang, Jieli Shen, Thomas Grubb, Shashank Bassi, Jason Song, Hareesh Nagarajan, Aya Avishai, Zellux Wang, and Min Li developed and deployed the User True Interest Survey model for Facebook Reels.

What: Meta implemented a survey-based measurement system collecting direct user feedback on content relevance through randomized five-point scale questions, training a lightweight machine learning model that integrates into existing recommendation systems to improve content matching beyond traditional engagement metrics like watch time and likes.

When: The research and implementation were published on January 14, 2026, following large-scale A/B testing with more than 10 million users conducted throughout the development period.

Where: The system operates across Facebook Reels and other video surfaces within Meta's platforms, affecting content recommendations for billions of users globally with particular impact on users in markets where Facebook maintains strong presence.

Why: Meta developed UTIS to address limitations in traditional recommendation systems that rely on noisy engagement signals, achieving 48.3% precision with interest heuristics before the new approach improved accuracy to 71.5%, precision to 63.2%, and recall to 66.1% while increasing high survey ratings 5.4% and total user engagement 5.2% in production testing.