Agentic Sites: Hyper-Personalized Web Experiences

The core idea is to understand "What is the user doing? What is the user trying to achieve?" in real time. This approach aims to dynamically personalize web pages to enhance user engagement and drive conversion rates.

The technical foundation for agentic sites relies on AMX delivery, a product that places all content on the edge for rapid distribution. This setup is complemented by a robust backend service that integrates various Large Language Model (LLM) providers. Sanchez specifically highlighted the use of Cerebras for demonstrating exceptionally fast inference capabilities within this stack.

To prevent content hallucination and maintain strict brand guidelines, agentic sites do not generate an entire website from scratch. Instead, they focus on personalizing specific content blocks within an existing site. The current site acts as a Retrieval-Augmented Generation (RAG) corpus, ensuring that all generated content is accurate and consistent with the brand's established voice and guidelines.

The demo showcased several practical personalization features, including instant persona adaptation, where the site adjusts content based on detected user characteristics. Search result pages are dynamically customized for individual users, and recommendations are generated in real-time based on browsing history and query analysis. This allows for a deeply responsive and relevant user experience.

Sanchez emphasized that the performance of agentic sites is highly dependent on the specific site and its requirements. To address this, a continuous evaluation process is in place, utilizing tools like Prompt Foo. This tool helps in continuously testing the accuracy and speed of the AI models, ensuring real-time performance across various model providers.

The development team prioritizes two key metrics: accuracy and speed, with a target generation time of one to two seconds per page. Faster site generation directly correlates with higher conversion rates, making low latency a crucial factor. Benchmarking results demonstrated a latency of 1.1 seconds when using Cerebras and Google's Gemma 4 model.

Performance benchmarks highlighted Cerebras achieving an average page generation speed of 1.1 seconds. In contrast, other models tested exhibited significantly slower times, often exceeding 4.6 seconds. This stark difference underscores the importance for developers to carefully consider model speed in relation to their specific use case requirements for real-time personalization.

Contrary to common assumptions, large-scale LLMs are not always necessary for web UI generation. The tasks involved, such as text generation and the organization of content blocks on a website, can be efficiently handled by smaller, specialized models. This approach reduces computational overhead, leading to faster and more responsive user experiences.

This data allows the system to trigger real-time site personalization, dynamically modifying elements such as hero cards, navigation menus, and product listings. The continuous feedback loop ensures that the site remains highly relevant to the user's evolving needs and interests.

User queries are systematically grouped into specific intent types or personas, differentiating between actions like 'trying to buy something' versus 'trying to get information.' This categorization allows marketers to manage how the AI interacts with distinct customer segments. The AI then selects and combines optimal content blocks based on these defined intent clusters, leading to more targeted and effective engagement.

Maintaining consistency with brand guidelines is crucial for all generated images. The LLM responses are further grounded by leveraging product guides and blogs as source material, ensuring both accuracy and brand alignment in multimedia content.

The user's browser collects signals about their behavior and intent. A backend system, potentially distributed across various cloud providers like Google and Cloudflare, performs reasoning by calling the LLM and utilizing Retrieval-Augmented Generation (RAG) built on the site's existing content for dynamic generation. Edge computing then handles the efficient serving of static content and personalized pages to the user.

These sites deliver unique per-customer experiences, showcasing dynamic product storytelling tailored to each user's specific journey and preferences. This allows for a level of individualized marketing previously unattainable.

The system continuously tracks user behavior, segmenting users into categories such as 'exploring' based on their page visit history and time spent on each page. This real-time data is then fed into the LLM, enabling it to personalize content. By understanding how users navigate and interact with the site, the LLM can dynamically tailor results to individual preferences.

The system demonstrates impressive performance metrics, achieving a total page generation time of approximately 1.1 seconds. This is accomplished using Cerebras hardware running Google's Gemma 4 model, which delivers a throughput of 2,200 to 2,300 tokens per second. These optimized benchmarks highlight the efficiency and speed of the underlying infrastructure in creating personalized web experiences.

The platform offers flexible controls for developers to experiment with various LLM models, adjust temperature settings, and set token limits. This capability allows for direct observation of how different configurations impact content generation and user experience. Alongside automated testing via Prompt Foo, manual intervention is also possible, enabling a comprehensive approach to tuning generative models.

This enables fully personalized, voice-activated interactions, creating a 'living room' experience tailored exclusively to individual user intent. Users can make queries via voice, and the system delivers personalized content directly to their smart TV, for example, on Google TV.

This technology is poised for continuous improvement, becoming cheaper, faster, and more widespread over time. The core value lies in understanding "What is the user trying to do?" and dynamically building page content around that intent, with multi-model integration expected to drive further performance enhancements.