AI Engineer World's Fair 2026

Agentic Sites: Building Hyper Personalized Websites — Carlos Sanchez, Adobe

Loading slides…

Read the talk

Agentic Sites: Building Hyper-Personalized Websites

Carlos Sanchez shows how Adobe’s prototype turns visitor intent into a purpose-built page by changing selected content blocks, grounding them in the existing site, and treating latency as part of model quality.

From a talk by Carlos Sanchez

At a glance

Ideas worth remembering

  • Personalize selected blocks rather than freely generating the whole website; use the existing site as the retrieval corpus so adaptations remain connected to approved content.

  • Treat latency as part of model quality. In the reported 15-prompt example, the chosen configuration averaged 1.1 seconds versus 4.6 seconds for the next option, but that comparison does not establish universal provider performance.

  • Pre-generating a “For You” page moves inference ahead of navigation, reducing visible waiting at the cost of repeated calls and potentially wasted generation.

  • Marketers define the intent groups and strategy; the model chooses and orders suitable blocks for the selected group.

  • The camping-coffee and conference-comparison demos show the central capability: reorganize several grounded blocks around a visitor’s task, rather than returning a conventional list of search results.

  • Behavioral personalization requires more than model engineering. Because the prototype records visited pages and time spent, a production design also needs explicit privacy and data-governance decisions.

Start with what the visitor is trying to accomplish

Carlos Sanchez begins with a different organizing question for a website: what is this visitor trying to do right now? Instead of serving one fixed page to everyone, the proposed system interprets browsing behavior and explicit requests, then adapts the experience in real time. Engagement and conversion are the intended marketing outcomes, although the talk does not present measurements showing that this implementation improves either one.

The important constraint is that the system does not freely generate an entire website. It changes selected content blocks according to a persona or inferred intent while preserving the surrounding site. That smaller action space fits the practical requirements of brands with strict visual and editorial guidelines: the model can choose, arrange, and rewrite approved material without inventing every component from scratch.

To constrain the content itself, the team turns the existing site into a retrieval corpus. Retrieval-augmented generation supplies relevant products, guides, stories, and other site material as context for each response. This connects generated copy to available content, but grounding alone cannot guarantee that every recommendation or claim will be correct; it narrows the source material rather than eliminating model error.

Sanchez describes three forms of adaptation:

  • Persona adaptation: change selected blocks as the system classifies the visitor.
  • Query-driven assembly: build a results page around an explicit search or request.
  • Behavior-based recommendation: prepare a page from what the visitor has browsed and what the system infers they may want.

The third mode is the least direct because it relies on inferred intent rather than a stated request.

Marketers define the strategy in natural language and use analytics to revise it. The intended loop runs from a business goal to a personalized experience, then from observed behavior back into the next strategy change. Sanchez describes the loop at a conceptual level; he does not specify an optimization algorithm, attribution method, or success threshold.

0:130:43
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:13 · section reference included

A page generator has an interaction budget

The frontend composes adaptable blocks and updates them through the AI service. Behind it, the team continuously evaluates prompts across multiple models and providers. Model selection is deliberately site-specific: corpus size, subject matter, commerce type, and the required outputs can all change which configuration is accurate enough and fast enough.

The team uses Promptfoo for comparisons across local and hosted providers, evaluating accuracy alongside speed. Latency is not an infrastructure footnote here. If generation happens after a visitor asks for a page, inference time becomes page-load time, so Sanchez sets a target of roughly one or two seconds.

For one example site and a set of 15 prompts, the selected Cerebras configuration averaged 1.1 seconds to generate a page; the next configuration averaged 4.6 seconds. That large gap drove the provider choice for the demonstration. These results describe one prompt set and site, not a universal ranking of providers or models.

This produces a practical selection rule: choose enough capability to satisfy the task within the interaction budget. The model is mostly generating bounded copy, selecting blocks, and arranging them, so Sanchez argues that the workload does not require the largest available LLM. A configuration can be imperfect yet useful if its quality is acceptable for the site and its response arrives before the experience feels broken. The talk leaves “good enough” to application-specific evaluation rather than defining a universal threshold.

Model selection is a site-specific tradeoff

Site-specific promptsevaluation-input — Prompts reflect the site’s size, domain, commerce type, and desired adaptations.
Accuracy evaluationselection-criteria — Outputs must be suitable for the site’s task and content.
Latency evaluationselection-criteria — On-demand page generation targets roughly one or two seconds.
Per-site configurationdecision — Select a model and provider that are accurate enough and fast enough for this experience.
The chosen configuration must meet both the site’s quality requirements and the page’s interaction budget.
Read the diagram as text
  • Site-specific prompts. evaluation-input — Prompts reflect the site’s size, domain, commerce type, and desired adaptations.
  • Accuracy evaluation. selection-criteria — Outputs must be suitable for the site’s task and content.
  • Latency evaluation. selection-criteria — On-demand page generation targets roughly one or two seconds.
  • Per-site configuration. decision — Select a model and provider that are accurate enough and fast enough for this experience.
  • Site-specific prompts → Accuracy evaluation: test output quality.
  • Site-specific prompts → Latency evaluation: measure generation time.
  • Accuracy evaluation → Per-site configuration: must be good enough.
  • Latency evaluation → Per-site configuration: must meet interaction budget.
4:305:00
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

4:30 · section reference included

Move some generation ahead of the click

The runtime inputs combine browsing activity with explicit queries. In the example, those signals can change the hero card, product selection, blog feed, navigation, and calls to action. Personalization therefore affects emphasis, ordering, recommendations, and the next action—not just a line of greeting text.

A “For You” page creates a useful timing option. As the visitor browses, the backend can repeatedly generate or refresh recommendations and prefetch the resulting page. When the visitor opens it, the expensive work may already be complete. This relaxes the immediate latency requirement, but it spends additional inference before the system knows whether the visitor will use the result.

Freshness creates the corresponding cost. Every new signal can make the prepared recommendation stale, so keeping it aligned with current behavior may require more generations and more LLM calls. The design trades waiting at navigation time for background inference cost.

Queries can also place visitors into marketer-defined personas or intent groups, such as buying a product versus gathering information. Marketers decide which groups exist and how they should be treated; AI selects the blocks, suggestions, sequence, and potentially media for the selected group. This keeps the segmentation strategy under business control while delegating page assembly to the model.

Sanchez treats generated images as a possible extension, with a sharper brand risk than text. Image generation is slower, and marketing acceptance depends on quality and brand fit. The demonstrated architecture does not require it: product pages and comparisons can be assembled from existing products, guides, experiences, and site media.

The stack separates observation, generation, retrieval, and delivery. The browser collects signals. A backend calls the LLM and reasons over retrieval results from the site’s vector database. Adobe Experience Manager serves pages and static content at the edge. This lets the static site remain fast and controlled while selected blocks take a dynamic path through retrieval and inference.

That signal path also creates an important production concern that the presentation does not explore: the demo records visited pages and time spent, then makes those observations available to the LLM. A deployed system would need explicit decisions about consent, retention, access, and what behavioral data may be sent to an inference service.

From visitor behavior to an adapted page

Visitor activitybrowser — Browsing and queries provide intent signals.
Persona or intent grouppersonalization — The system classifies the visitor using groups defined by marketers.
Site retrieval corpusbackend — A vector database retrieves relevant existing site content.
LLM assemblybackend — The backend selects, orders, and rewrites adaptable blocks.
Edge-delivered pagedelivery — Static content is served with personalized heroes, products, feeds, navigation, and calls to action.
Static site content remains at the edge while selected blocks travel through a retrieval-and-generation path.
Read the diagram as text
  • Visitor activity. browser — Browsing and queries provide intent signals.
  • Persona or intent group. personalization — The system classifies the visitor using groups defined by marketers.
  • Site retrieval corpus. backend — A vector database retrieves relevant existing site content.
  • LLM assembly. backend — The backend selects, orders, and rewrites adaptable blocks.
  • Edge-delivered page. delivery — Static content is served with personalized heroes, products, feeds, navigation, and calls to action.
  • Visitor activity → Persona or intent group: signals inform classification.
  • Persona or intent group → Site retrieval corpus: guides relevant content selection.
  • Site retrieval corpus → LLM assembly: grounds the response.
  • LLM assembly → Edge-delivered page: updates selected blocks.
8:168:46
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

8:16 · section reference included

A camping request changes the whole product story

Sanchez calls the ambition “Audience of One”: a marketing experience assembled for an individual rather than a broad segment. His demonstration uses a fully generated example site for coffee equipment. He browses stories and products, spending time on individual pages before opening a debugging panel.

The panel shows the mechanism behind the apparent personalization. The visitor is bucketed as “exploring,” while the system records visited pages and time spent on each. Those signals are available to the LLM, and a “For You” page has already been generated from the browsing history.

The explicit-query demo makes the transformation easier to inspect. Sanchez asks for a coffee machine suitable for preparing coffee while camping. The resulting page changes its introductory copy, adds camping-specific tips, and recommends two machines described as appropriate for that use. The page is more than a ranked search result: several blocks tell one coherent product story around the same intent.

A similar run in debug mode takes 1.64 seconds in total, including the round trip to the LLM. Sanchez reports approximately 2,300 tokens per second for that run; another run shows about one second of LLM time and roughly 2,200 tokens per second. These numbers describe different layers of performance: total generation time includes more than model decoding, while token throughput measures only generation rate.

The demo interface also lets the team switch models and adjust temperature and token settings. Automated Promptfoo tests provide repeatable comparisons, while manual interaction reveals whether a configuration behaves well inside the actual page experience. The two methods answer different questions: batch evaluation measures defined cases; interactive inspection exposes coherence, responsiveness, and presentation problems that a score may miss.

12:3613:06
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

12:36 · section reference included

Carry the pattern to another site in under an hour

Sanchez then introduces Of One Labs, a tool that accepts a URL and generates an agentic demonstration site in less than an hour. The claim concerns creating a demo, not automatically converting any production site with its full brand, security, analytics, accessibility, and editorial requirements intact.

He applied the tool to the AI Engineer site. The generated experience centers on a search box with AI-generated suggestions; a query for European AI conferences is intended to assemble a page focused on relevant events. As in the coffee example, the query supplies the job the page should perform rather than merely a keyword to match.

The strongest result is a side-by-side conference comparison. Once the system infers that a visitor is choosing between two events, it changes the information structure to support that decision. This is the more consequential form of personalization in the talk: the system does not simply rank one item above another; it selects a comparison layout suited to the visitor’s task.

16:4017:10
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

16:40 · section reference included

Intent-driven assembly can outlive the conventional webpage

The closing example asks what this pattern might become if a personal assistant mediates the interaction. Sanchez imagines a spoken shopping request producing a personalized visual response on a television. The user supplies intent by voice, and the system chooses a screen-sized presentation without requiring a phone or computer.

He is careful not to claim certainty about whether the web itself is dying or what will replace it. The supported result is narrower: the demonstration shows that a system can already assemble selected, site-grounded blocks around a visitor’s inferred or explicit intent. Extending that result into a general assistant-driven interface remains a proposed direction.

Sanchez expects generation to become faster, cheaper, and more widely applicable, but those are forecasts. The engineering lesson available today is concrete: constrain what may change, retrieve from the site’s approved content, evaluate models against each site’s quality and latency needs, and assemble the page around the job the visitor is trying to complete.

18:1718:47
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

18:17 · section reference included

Resources

From the talk

  • Sanchez’s technical weblog covers his broader work in AI, infrastructure, Kubernetes, and automation.

  • The speaker’s public profile for updates on his technical projects and presentations.

Transcript download
Full transcript

Loading transcript…