# Agentic Sites: Building Hyper Personalized Websites Carlos Sanchez, Adobe | AI Engineer World's Fair 2026 | 20:42 Source: https://www.youtube.com/watch?v=jebp4V0vh30 Channel: AI Engineer (https://www.youtube.com/@aiDotEngineer). Summarised by AIE Talks. Page: https://aietalks.com/talks/agentic-sites-building-hyper-personalized-websites Published: 2026-08-29 Tags: latency, product-strategy, rag ## TL;DR - Agentic sites personalize selected blocks from a real website according to a visitor's intent, instead of generating an entire page from scratch. - The site itself becomes the retrieval corpus, which keeps generated copy, recommendations, and layouts grounded in existing brand content. - Model selection has to be tested per site, with latency treated as an important part of the user experience alongside accuracy. ## Summary Carlos Sanchez presents Adobe's approach to agentic sites, where a website adapts itself to a visitor's intent in real time. The system records browsing signals, groups visitors into intent or persona types, and uses those signals to change selected blocks such as the hero, product list, navigation, blog feed, and calls to action. The whole site is used as a retrieval corpus, so generated content stays tied to the brand's existing products, guides, and experiences. Sanchez describes continuous testing across models and providers because each site has different content and latency needs. In his example, the fastest setup averages 1.1 seconds, compared with 4.6 seconds for the next option. A demo shows a coffee site producing a camping-focused page from a natural-language request. He also shows a tool that turns any URL into an agentic site in less than an hour, including an AI Engineer conference comparison page. ## Key ideas ### An agentic site adapts its pages to the visitor's intent [00:01](https://www.youtube.com/watch?v=jebp4V0vh30&t=1s) Sanchez defines an agentic site as a website that tries to understand what a visitor is doing and what they want to achieve. It then personalizes the experience in real time for that person. The intended outcome might be higher engagement or more conversions, depending on what the marketing team wants. Adobe's stack puts content at the edge and uses a back-end service with different LLM providers to power the personalization. Sanchez says the system can adapt to a user's browsing behavior, a search query, or a recommendation need. ### Personalization changes blocks while the brand controls the page [02:20](https://www.youtube.com/watch?v=jebp4V0vh30&t=140s) The system does not generate the whole website. It changes selected sections because marketing teams have strict brand guidelines and do not want uncontrolled content or hallucinations. Sanchez names the hero card, products, blog feeds, navigation, and calls to action as blocks that can be personalized. This lets customers avoid authoring thousands of separate page variations. The same mechanism can adapt a search-results page to a user's query and create recommendations after the user has browsed for a while. ### The existing site grounds everything the model produces [02:50](https://www.youtube.com/watch?v=jebp4V0vh30&t=170s) Adobe treats the whole site as a corpus and builds retrieval over it. Generated responses draw on the site's existing products, guides, experiences, and other blocks. Sanchez says this grounding is used for tailored product pages and product comparisons as well as ordinary page personalization. The architecture includes a vector database, inference infrastructure, and back-end reasoning that calls an LLM with the site's retrieval system. The result is a generated arrangement and wording based on material already present on the site. ### Model evaluation depends on the individual site [04:30](https://www.youtube.com/watch?v=jebp4V0vh30&t=270s) Sanchez says there is no single model choice that works for every site. Adobe runs prompts across many models and providers, then checks accuracy and speed. The evaluation is continuous because the result depends on factors such as the site's size, target area, and type of commerce. Adobe uses Promptfoo to compare prompts across multiple models and providers, including local models and OpenAI-compatible services. Sanchez also describes a practical threshold: page generation should take about one or two seconds, since slower experiences can reduce conversions. ### The fastest tested setup averages 1.1 seconds [06:35](https://www.youtube.com/watch?v=jebp4V0vh30&t=395s) For the example site, Adobe tested 15 prompts. Sanchez shows Cerebras running the Gemma 4 model with an average page-generation latency of 1.1 seconds. The second option averages 4.6 seconds, a large difference for an interactive website. He also notes that different models may be suitable for different tasks. A model does not need to be perfect for every operation if it is accurate enough and fast enough for that particular use. This makes the evaluation a site-specific engineering decision. ### The task can use a smaller model because it arranges known material [07:55](https://www.youtube.com/watch?v=jebp4V0vh30&t=475s) Sanchez argues that this use case does not need a very large LLM. The system generates some text, chooses blocks, and decides how to organize a website. It is not asked to supply a large body of new information because retrieval provides the site's content. Browsing and query data becomes input for personalization. Depending on the visitor's persona, the system can change the hero, products, navigation, blog feeds, and calls to action. The work is mainly selecting and arranging content that already exists. ### Some personalized pages can be generated before the visitor asks [08:48](https://www.youtube.com/watch?v=jebp4V0vh30&t=528s) The recommendation page, which Sanchez calls a For You page, can be generated while a person browses. The site collects signals and keeps preparing recommendations in the background, so a visitor could immediately see products suited to what they appear to want. Sanchez says the page should be updated as the person moves around the site. This approach reduces the need for the same latency at request time, although it creates cost considerations because the system makes multiple generations and LLM calls. ### Marketers define the intent groups while AI selects the page blocks [10:05](https://www.youtube.com/watch?v=jebp4V0vh30&t=605s) When a visitor runs a query, the system generates a personalized page and groups the query into a persona or intent type. Sanchez gives examples such as someone trying to buy something and someone looking only for information. Marketers can decide which groups to create, how many groups they want, and how to handle those customers. The AI then chooses the blocks, their sequence, the suggestions, and potentially the media for each group. Sanchez also mentions possible on-the-fly image generation, while questioning whether the quality would be sufficiently on-brand. ### A natural-language query produces an audience-of-one page [12:35](https://www.youtube.com/watch?v=jebp4V0vh30&t=755s) In the live demo, a coffee-machine site records visited pages, time spent, and a bucket called 'exploring.' It uses these signals to prepare a For You page. Sanchez then asks for a coffee machine to use while camping. The site immediately changes the copy, adds camping coffee tips, and recommends machines suited to a camping trip. He reports a total generation time of 1.64 seconds in the debug view, with about one second spent in the LLM and roughly 2,300 tokens per second from Gemma 4 running on Cerebras. The demo also allows manual model and parameter changes. ## Notable quotes - "We don't want the whole site to be generated." (02:30) - "We use the whole site as a corpus." (02:51) - "We don't need a huge LLM to do this sort of work because you are generating text, you are deciding where to put blocks and how to organize the website." (07:55) - "So, we call this audience of one because the idea of in marketing they always dream on being able to personalize things for each individual." (12:35) - "In less than an hour, you have an agentic site." (16:40) ## Tools & references mentioned - Adobe Experience Manager - Adobe - Cerebras - Gemma 4 - Amazon Bedrock - Promptfoo - Cloudflare - Google - Google TV - OfOneLabs - AI Engineer ## Who should watch - You work on a content-heavy or commerce website and need personalization without asking marketing teams to author large numbers of page variants. - You are evaluating LLM providers for a production web experience where response time matters as much as answer quality. - You want to see an implementation pattern for grounding generated page changes in an existing site's content and structure.