Agentic Sites: Building Hyper Personalized Websites — Carlos Sanchez, Adobe
Adobe’s prototype assembles intent-specific page blocks from existing site content in roughly a second, making model latency and per-site evaluation part of frontend architecture.
Adobe’s prototype personalizes selected page blocks while grounding generated copy and recommendations in the existing site corpus. Across **15 prompts**, Cerebras with Gemma 4 averaged **1.1 seconds**, versus 4.6 seconds for the next shown result.
For agent-generated interfaces, evaluate model and provider combinations against each site’s own content and latency target. Keep brand-controlled structure fixed, generate only bounded blocks, and precompute recommendations when live generation is unnecessary.
Adobe’s prototype personalizes selected page blocks while grounding generated copy and recommendations in the existing site corpus. Across **15 prompts**, Cerebras with Gemma 4 averaged **1.1 seconds**, versus 4.6 seconds for the next shown result. For agent-generated interfaces, evaluate model and provider combinations against each site’s own content and latency target. Keep brand-controlled structure fixed, generate only bounded blocks, and precompute recommendations when live generation is unnecessary. The demo records browsing and query signals, creating privacy and cost questions that the talk does not resolve. Its speed results come from one example site and do not establish quality or conversion gains across production properties.
This narrows agentic-site personalization to a controlled rendering pattern: retain brand-owned structure, ground outputs in the site corpus, and generate only selected blocks. The reported latency makes provider benchmarking relevant but remains site-specific evidence, not a quality or conversion result. Privacy, signal collection, and recurring inference cost remain unresolved prerequisites for production adoption.