TLDR: Meta has reported substantial improvements in ad performance across its platforms, Facebook and Instagram, attributed to its new Generative Ads Recommendation Model (GEM). This AI-driven foundation model, inspired by large language models, has led to a 5% increase in ad conversions on Instagram and a 3% increase on Facebook Feed in Q2 2025, demonstrating enhanced personalization and efficiency in ad delivery.
Meta’s Generative Ads Recommendation Model (GEM), an advanced AI-powered foundation model, is significantly boosting ad performance across the company’s social media platforms. Launched earlier this year, GEM has already contributed to a 5% increase in ad conversions on Instagram and a 3% increase on Facebook Feed during the second quarter of 2025. This success highlights Meta’s ongoing efforts to leverage artificial intelligence for more effective and personalized advertising.
Described by Meta as the ‘central brain’ behind its ad recommendation system, GEM is built on an LLM-inspired paradigm and trained across thousands of GPUs. Its primary objective is to enhance how Meta personalizes ad delivery by learning from vast amounts of user interaction data, encompassing both organic and paid content. The model’s architecture is notably efficient, reportedly four times more efficient than previous versions in driving ad performance gains for a given amount of data and compute. Furthermore, it achieves double the effectiveness in knowledge transfer compared to standard methods, allowing for broader optimization across Meta’s ad stack.
Meta’s engineering blog post on November 10 detailed the progress, noting that subsequent architectural changes in Q3 doubled the performance benefit per unit of data and compute, supporting further scalability. The company has implemented advanced training strategies, including distributed training, hybrid sharding methods, and custom GPU kernels, to manage variable-length user data and accelerate processing speed. These innovations have resulted in a 23-fold increase in training performance and a 1.43-times improvement in hardware utilization. Additionally, job startup times have been reduced fivefold, and compilation speed improved through caching strategies in PyTorch 2.0.
GEM learns from user behaviors across Meta’s applications, utilizing customized attention mechanisms to understand long-term engagement patterns and refine ad targeting accuracy. It also applies domain-specific optimization, leveraging cross-platform activity, such as Instagram video engagement, to improve predictions on other surfaces like Facebook Feed. The model is designed to learn across various modalities—text, image, audio, and video—to better capture the nuances behind user interactions and long-term value, with a long-term vision to unify ranking for both organic content and ads.
Also Read:
- Generative AI Drives Substantial Growth in October E-commerce Sales, Adobe Analytics Reveals
- VaynerMedia APAC Unveils ‘CREAITORS’ AI Platform for Scalable, Culturally Relevant Campaigns
This initiative is part of Meta’s broader multi-year strategy to automate more of the campaign setup, retrieval, and ranking processes, while integrating generative AI tools for creative development. The company envisions a future where GEM will enable a self-learning marketing ecosystem capable of adapting to user intent in real-time, potentially automating the entire ad creation and optimization process with minimal human input.
