Adobe demonstrated a new system called Agentic Sites that personalizes website content in about one second, using fast inference hardware and retrieval grounded in the site's own content. The demo, built by principal scientist Carlos Sanchez inside Adobe Experience Manager, targets the moment when a visitor arrives with a specific intent - and adapts the page to match.
Rather than generating an entire site from scratch, the engine personalizes individual blocks: hero cards, product listings, blog feeds, navigation, and calls to action. All generated text is grounded on a retrieval-augmented generation corpus built from the full site, so brand guidelines remain intact.
How the system works
Content runs on Adobe Experience Manager Edge Delivery Services, a framework that pushes rendering to the edge for speed. A backend service handles reasoning over a vector database and calls multiple LLM providers, with Cerebras used for fast inference and alternatives like Amazon Bedrock also being tested.
Browsing signals are collected continuously - pages visited and time spent on each - and fed to the LLM to infer personas like "exploring" or "buying." Marketers define the strategy in natural language and decide how many persona groups to use, such as buyers versus information seekers.
Why 1.1 seconds matters
Sanchez said Adobe targets 1-2 seconds for generation because faster sites drive more conversions. Evaluation runs continuously with Promptfoo, an open-source testing tool that compares prompts and models side by side.
On a 15-prompt test set for the demo coffee gear site, Cerebras with Gemma 4 averaged 1.1 seconds per generation, while the next-best provider averaged 4.6 seconds. The live demo clocked 2,300 tokens per second and an LLM time of 1 second for a query about a coffee machine for camping, returning tailored copy and product recommendations for Agile and Nano brewers.
Cerebras has previously reported wafer-scale inference above 1,800 tokens per second on Llama 3.1 8B and 3,000 tokens per second on OpenAI models.
Where marketers keep control
Marketers write the personalization strategy in plain language and choose how many personas to define. Query pages and "For You" recommendation pages are grouped by intent and can be pre-generated and pre-fetched while the user browses.
Sanchez also showed OfOneLabs, an internal Adobe tool that crawls any URL and spins up an agentic site in under an hour. In a live example, it turned an AI engineering events site into a search box that built lists of European AI conferences and side-by-side comparisons on the fly.
He then extended the same query to a voice assistant on Google TV, arguing agentic pages will work beyond the browser.
For marketers, this is a shift in how personalization is built. The technology reframes the bottleneck from model quality to inference economics: every page now requires an LLM call and a fresh RAG retrieval. That favors fast inference vendors and puts pressure on CMS rivals to match edge-plus-AI latency.
Following the approach of larger vendors, Adobe views agentic personalization as a product layer, not a research project. Its Agent Orchestrator is already in beta for Experience Platform, and tools like Promptfoo let teams verify whether a provider's speed is sufficient when accuracy is comparable.
Why this matters for marketers
The demo suggests the next phase of personalization isn't about writing more campaigns but about defining intent signals that let a system adapt pages in real time. That's a workflow change for marketing teams, who will also need to monitor cost per generation as audiences are drilled down to one person at a time. The open questions remain: pricing at scale, governance of automatically generated blocks, and how brands will verify that personalized content stays on-message.
Your membership also unlocks:
