The Architecture of Agentic Sites

Agentic sites move beyond static pages by using LLMs to assemble pre-defined content blocks in real-time. Rather than generating an entire page from scratch—which risks hallucination and violates brand guidelines—the system treats the entire website as a corpus. It uses Retrieval-Augmented Generation (RAG) to ground the AI's output, ensuring that only specific sections like hero banners, product lists, and calls-to-action are dynamically adapted to the user's current intent.

Prioritizing Speed for Conversion

For personalization to be effective, latency must be minimal. Carlos Sanchez emphasizes that for web experiences, a generation time exceeding two seconds leads to lost engagement. To achieve sub-second performance, the team:

• Evaluates models per site: Performance varies based on site size and content type, so they run continuous evaluations using tools like Promptfoo to test latency and accuracy across different providers.
• Selects for speed: They prioritize fast inference providers (e.g., Cerebras) over frontier models, noting that the task is primarily about choosing and arranging existing blocks rather than complex reasoning.
• Uses pre-generation: As users browse, the system gathers signals and buckets them into personas, allowing the site to pre-generate "For You" pages or recommendations before the user even requests them.

Intent-Driven Personalization

By tracking user signals—such as time spent on pages and navigation paths—marketers can define strategies in natural language to group users into intent-based personas. The AI then dynamically selects the sequence of blocks and media that best serves that specific user's goal. This "audience of one" approach allows brands to deliver highly relevant experiences without the manual labor of creating thousands of static page variations. The system is designed to be modular, allowing for the rapid deployment of agentic capabilities to any existing URL by indexing its content into a vector database.