AnswersPublished29 sources
How Adobe’s Agentic Sites Personalize Web Pages in Real Time
Adobe demonstrated a prototype that assembled an intent based coffee machine page in 1.64 seconds, including about one second for the model call; the reported $0.01–$0.02 page cost and performance are prototype estima... Instead of generating an entire website from scratch, the system retrieves approved AEM content...
Published byEdited with GPT-5.6 LunaImages generated with GPT Image 1.5
Research answer
Key takeaways
- Adobe demonstrated a prototype that assembled an intent based coffee machine page in 1.64 seconds, including about one second for the model call; the reported $0.01–$0.02 page cost and performance are prototype estima...
- Instead of generating an entire website from scratch, the system retrieves approved AEM content and adapts discrete blocks such as hero cards, product feeds, navigation, and calls to action for an individual visitor.
- The concept is promising but is not yet established as a production scale, per visitor Adobe service; governance, integrations, traffic economics, measurement, and commercial pricing remain open questions.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What did Adobe principal scientist Carlos Sanchez demonstrate about “agentic sites”—including how the system uses Cerebras inference hardwar. Article summary: Sanchez demonstrated a prototype “agentic site”: rather than serving one pre-authored page to a segment, it infers an individual visitor’s intent and assembles approved page components around it in real time. It is a cre. Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fa
Adobe principal scientist Carlos Sanchez demonstrated a prototype “agentic site”: a website that interprets a visitor’s intent and assembles a page around that intent in real time. The goal is an “audience of one,” replacing some predefined audience segmentation with page experiences that adapt to what an individual visitor appears to need. 2
7
The important caveat is maturity. Public reporting describes a working demonstration and related Adobe products, not a generally available service that generates a unique production page for every visitor at scale. 4
34
What Sanchez demonstrated
The demonstration centered on a coffee-machine website. After a visitor searched for a machine suitable for camping, the system changed the page’s content, recommendations, and presentation to reflect an outdoor-use intent. Sanchez’s session at the AI Engineer World’s Fair was titled “Agentic Sites: Building Hyper Personalized Websites.” 33
The underlying idea is more controlled than asking a language model to invent a website. The system identifies an intent or persona from signals such as a visitor’s query and browsing behavior, retrieves relevant material from an existing content corpus, and composes the page from that material. 2
34
How the architecture works
A simplified version of the workflow looks like this:
- Interpret intent: The system evaluates the visitor’s search or interaction context to determine whether they are exploring, researching, preparing to purchase, or pursuing a more specific use case such as camping.
- Retrieve approved material: Retrieval-augmented generation, or RAG, supplies the model with relevant content and product information from the organization’s existing systems.
- Choose page components: Rather than treating the page as one indivisible document, the system can select or adapt individual blocks.
- Render the experience: The resulting hero content, product recommendations, navigation, and calls to action are assembled into a page suited to the inferred intent.
This approach matters for brand control. Grounding the response in enterprise content and approved assets can reduce the risk of unsupported product claims and keep the experience closer to the organization’s established visual and editorial identity. Reports on the demonstration describe the system as retrieving and reassembling existing content rather than generating pages from nothing. 4
34
Why Cerebras and Gemma 4 matter
The reported implementation used Google’s Gemma 4 model on Cerebras inference infrastructure. Cerebras documents Gemma 4 31B at approximately 1,850 output tokens per second in its model catalog, while Cerebras has reported roughly 2,300 tokens per second for a Gemma 4 use case. 3
6
High token throughput is relevant because an agentic site may need to perform several steps during one web interaction: interpret intent, retrieve context, decide which components to use, generate or revise copy, and return the result. Faster inference makes that retrieve–reason–compose loop more plausible in an interactive experience rather than limiting it to offline content production.
Token throughput is not the same as full-page latency. Network conditions, retrieval, catalog calls, consent checks, browser rendering, caching, and backend delivery all contribute to what the visitor actually experiences.
Reported speed and cost
According to reporting on Sanchez’s discussion, a camping coffee-machine request produced a personalized result in 1.64 seconds, with approximately one second attributed to the model call. The same reporting compared that result with 4.6 seconds on another setup and cited an estimated generation cost of roughly $0.01–$0.02 per page. 1
Those figures should be read as reported prototype benchmarks and estimates, not audited service-level commitments or published Adobe pricing. They also do not establish that every page, visitor, or traffic pattern would perform identically.
What an under-two-second target would enable
For personalization to feel immediate, the system must finish before its decision becomes a noticeable source of friction. A roughly one-to-two-second target could allow a site to react to a search, arrival context, or meaningful change in visitor intent while the interaction is still fresh.
That would make it possible to personalize at a finer level than a conventional segment such as “returning customer” or “outdoor enthusiast.” The system could instead respond to the visitor’s immediate goal: finding a coffee machine for camping, comparing products, or looking for a purchase-ready recommendation.
The target does not remove the rest of the web stack. Consent and identity controls, product and price validation, retrieval latency, CDN behavior, page rendering, and failure handling still need to fit within the overall experience budget.
How this relates to Adobe’s existing products
The prototype should not be confused with every current Adobe AI product. Adobe’s documented AEM agents are available for AEM as a Cloud Service and Edge Delivery Services, with availability and rollout conditions varying by environment. 17
Adobe Experience Platform Agent Orchestrator is described as the reasoning and coordination layer for Experience Platform agents. It is designed to use Experience Platform data and coordinate Adobe and third-party agents with human oversight. Adobe has also announced general availability for Agent Orchestrator, while documentation describes request-based access and usage-bound trials for some agent capabilities. 18
19
21
22
Other related capabilities include:
- Brand Experience Agent: Adobe describes this as helping update pages, create new content, and modernize existing digital experiences for AI-driven discovery.
20
23
44
- Content Advisor Agent: This helps teams find relevant enterprise content and create visual variants from approved source assets.
- Agent Orchestrator: This provides a platform layer for managing and coordinating agents across Adobe and third-party ecosystems.
These products show that Adobe is building an agentic technology stack around content, customer data, and digital experiences. They do not, by themselves, prove that Sanchez’s per-visitor page-assembly prototype is already a standard, generally available production feature.
What must happen before large-scale adoption
The technical demonstration answers whether real-time page assembly is plausible. It does not yet answer whether it is operationally or economically sensible for every site.
Customers need controlled experiments
Organizations would need to test the approach against their own AEM repositories, product catalogs, identity systems, consent policies, and conversion goals. The key question is not only whether a page can be generated quickly, but whether the resulting experience improves useful outcomes without increasing errors or operational risk.
Public reporting says the specific agentic-site concept has not yet shipped to client sites at scale. 4
The system needs broad integrations
A production deployment would need reliable connections among inference, content management, customer data, identity and consent, commerce or catalog systems, analytics, experimentation, CDN delivery, and brand-asset libraries. A fast model call cannot compensate for stale product data or a slow catalog service.
Per-page economics need traffic-scale testing
A one- or two-cent estimate may appear small for an individual interaction, but the total becomes material when multiplied across large traffic volumes. Costs could rise further if the site regenerates an experience after every meaningful interaction.
Caching, reuse of approved blocks, selective regeneration, and clear rules for when not to invoke a model would therefore be central to the business case. The public material does not establish a definitive Adobe pricing model for a world in which every visitor receives a unique generated experience.
Governance must cover every generated block
A credible deployment would need explicit controls for:
- approved content sources and retrieval boundaries;
- product attributes, availability, and price validation;
- allowable layouts, components, and calls to action;
- brand voice and visual-identity rules;
- accessibility and localization checks;
- prohibited or regulated claims;
- approval and audit trails;
- monitoring, rollback, and drift detection; and
- thresholds for human review.
The strongest version of the concept is therefore not an unconstrained AI website. It is a governed assembly system that uses AI to make a fast decision about which approved components to show and how to adapt them to a visitor’s current intent.
The practical takeaway
Sanchez’s demonstration suggests that the website can become a runtime decision layer rather than a fixed collection of pre-authored pages. Cerebras inference and Gemma 4 make the reasoning step fast enough to investigate real-time use cases, while AEM and retrieval-based grounding provide a path toward factual, brand-controlled output. 1
3
6
17
But the headline is a prototype, not a finished product category. The next test is whether agentic sites can maintain speed, accuracy, accessibility, governance, and measurable business value when deployed across real catalogs, real consent requirements, and real traffic volumes.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "How Adobe’s Agentic Sites Personalize Web Pages in Real Time"?
Adobe demonstrated a prototype that assembled an intent based coffee machine page in 1.64 seconds, including about one second for the model call; the reported $0.01–$0.02 page cost and performance are prototype estima...
What are the key points to validate first?
Adobe demonstrated a prototype that assembled an intent based coffee machine page in 1.64 seconds, including about one second for the model call; the reported $0.01–$0.02 page cost and performance are prototype estima... Instead of generating an entire website from scratch, the system retrieves approved AEM content and adapts discrete blocks such as hero cards, product feeds, navigation, and calls to action for an individual visitor.
What should I do next in practice?
The concept is promising but is not yet established as a production scale, per visitor Adobe service; governance, integrations, traffic economics, measurement, and commercial pricing remain open questions.