The website of the future assembles itself on the spot for every visitor: Adobe demos pages generated live from intent

The same coffee-machine site, completely changed because you searched "camping"

In the demo, the system treats a visitor's browsing behavior and search terms as signals. On the same coffee-machine site, a regular visitor sees the standard homepage; a visitor who searched "camping" sees copy, product ordering, and recommendations all reorganized around the theme of "brewing coffee outdoors." The toggle below contrasts the two versions.

Same URL, same company — the tone of the copy, the products up front, and the recommendations below all shift to match the visitor's intent. The demo also has a second, more open-ended entry point: the visitor just types a query, and the page assembles itself around that sentence, live.

What we called "personalization" was really just picking from a menu

Sanchez says that after all his years running websites, "personalization" has always been an ideal — but in practice it basically meant one thing: choosing from a set of pre-defined options. A merchant recommends a product based on what you bought last time, or slots you into an audience segment and applies the matching version. That's it. Agentic sites swap out the very nature of this.

How a page gets assembled on the spot

This is the core of the whole thing. The system first treats a visitor's browsing behavior and search terms as signals, sorts them into an intent category — exploring, researching, or ready to buy — then uses an AI to assemble a matching page from that intent. The key step sits in the middle: the AI doesn't invent content from nothing, it retrieves material from the company's existing content library and pieces it together.

Intent classification is the pivot of this pipeline. The same person, in "just browsing" versus "ready to check out," gets a page whose center of gravity is completely different — the first spreads out the possibilities, the second heads straight for conversion. Three intents, three ways to assemble.

1-2 seconds, a cent or two: is this slow, is it expensive

For anyone doing AI engineering, the first wall live page generation hits is latency. Sanchez says Adobe evaluates models on more than accuracy — speed too: "We don't want a page to take more than a second or two to generate." He also gave a current cost estimate, and stressed the number only goes down.

"And our read is, it only gets cheaper," Sanchez says. "That's where it is today. Where it'll be in six months, who knows." With cost already down to a cent or two per page, "reshaping the page live for every visitor" is no longer something only the giants can afford.

Not actually live on customer sites yet

Adobe hasn't rolled this experience out at scale on customers' production sites. Sanchez says the company is currently pitching the concept to customers and looking for organizations willing to experiment. The first-choice setting is clear — e-commerce, because personalization ties directly to conversion rate. But he stresses the opportunity isn't limited to retail: "Any site that needs to lift conversion and has a whole range of user types or personas could use it."

The uncertainty is real too. Sanchez says plainly that he isn't sure agentic sites will become a widespread reality. He sums up building products in the AI era in one line: "With AI it's easy to build things, but hard to know what to build. We build it first, then go find the customers." Adobe isn't alone in this fog — site owners across the industry are right now weighing chat interfaces, structured content, generative UI, personal agents, and a pile of other new things, while also figuring out how to pull users back from third-party AI platforms to their own sites.

Web pages aren't just for people to look at anymore

Websites after 2026 have to serve more than human visitors. As personal agents get more capable, users may hand off a chunk of their buying and research entirely to an agent. And when that agent arrives at a site, the preferences it carries may be far richer than what the site could infer from cookies or recent browsing. Sanchez expects sites to evolve for both kinds of visitor at once.

But not every transaction gets the same degree of automation. A person might happily let an agent auto-reorder toilet paper — a low-stakes repeat purchase — but for a jacket, they'll probably still want to see the product themselves and make the final call. That means a site has to support varying degrees of "delegated authority," and can't treat "agentic commerce" as one uniform interaction mode. The spectrum below is the ground between those two ends.

Sanchez says it's still fuzzy whether sites will split into "two versions" — one for people, one for agents: "But clearly, you have to serve both sides."

The keys that open the door to agents

If a site has to serve both people and agents, it needs a set of interfaces that let agents get in and get things done. The original piece names three enabling technologies taking shape, each mapping to a different stretch of that spectrum and solving one problem.

Sanchez says the end state might be a single site carrying both visual components and agent-callable tools, the two experiences coexisting — or a human-facing site paired with a set of agent-to-agent backend services. "This is still something everyone's working through," he says, "but one thing's for sure: there'll be an agent-facing direction." By his account, traditional websites probably won't disappear entirely, but their role will change: from a pile of fixed pages waiting for each visitor to browse, into a governed content-and-interaction system that assembles the right interface on demand. This is the direction Adobe is exploring.

Rather than a website being a set of fixed pages that every visitor navigates through, the "website" becomes a governed content-and-interaction system that assembles a suitable interface on demand. At least, that's the future Adobe is actively exploring. Carlos Sanchez / Richard MacManus, Latent.Space