Same Coffee-Maker Site, Totally Different After You Search "Camping"

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

Same domain, same business — but the tone of the copy, the products up front, and the recommendations below all shifted with the visitor's intent. The demo has a second, more open-ended entry point: a visitor can just type a query, and the page is assembled around it on the spot.

What We Used to Call "Personalization" Was Really Just Picking from a Menu

Sanchez says that after years of running websites, "personalization" has always been an ideal — but in practice it's basically meant one thing: choosing from a batch of pre-built options. A merchant would recommend an item based on your last purchase, or sort you into a segment and show you the matching version. That was it. Agentic sites change what's actually happening under the hood.

How the Page Gets Assembled on the Spot

This is the core of the whole system. It first turns a visitor's browsing behavior and search terms into signals, sorting them into an intent category — exploring, researching, or ready to buy — then uses AI to assemble a page matching that intent. The key step is hidden in the middle: the AI isn't inventing content out of thin air, it's retrieving material from the business's existing content library and assembling from that.

Intent classification is the hinge of this whole pipeline. The same person, in "just browsing" mode versus "ready to check out" mode, gets a completely different page from the system — one spreads out possibilities, the other drives straight toward conversion. Three intent tiers, three different ways of assembling the page.

1–2 Seconds, 1–2 Cents: Is This Slow or Expensive?

For anyone doing AI engineering, the first wall real-time page generation hits is latency. Sanchez says Adobe evaluates models not just on accuracy but on speed: "We don't want page generation to take more than a second or two." He also gave a current cost estimate, stressing that the number will only keep dropping.

"And our belief is it's only going to get cheaper," Sanchez said. "This is where it is today. Where it'll be in six months, who knows." Cost already sitting at a cent or two per page means "reassembling the page for every visitor in real time" is no longer something only big companies can afford.

Not Actually Live on Client Sites Yet

Adobe hasn't rolled this experience out at scale on client production sites. Sanchez says the company is currently demoing the concept to clients and looking for organizations willing to run experiments. The first-choice use case is clear — e-commerce, because personalization there 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 wide range of user types or personas could use this."

Uncertainty is real here too. Sanchez says outright he's not sure agentic sites will become common reality. He sums up the position of building products in the AI era in one line: "AI makes it easy to build things, but hard to know what to build. We build it first, then go find customers." And it's not just Adobe living with this uncertainty — the whole industry of site owners is right now simultaneously weighing chat interfaces, structured content, generative UI, personal agents, and a pile of other new options, all while figuring out how to pull users on third-party AI platforms back to their own site.

Websites Aren't Just for Humans Anymore

Websites past 2026 have to serve more than human visitors. As personal agents get more capable, users may hand off some shopping and research entirely to their agent. And when that agent arrives at a site, it may be carrying preferences far richer than anything the site could infer from a cookie or recent browsing history. Sanchez expects sites to evolve for both kinds of visitors at once.

But not every transaction gets the same level of automation. Someone might happily let an agent auto-reorder toilet paper — a low-stakes repeat purchase. But buy a jacket, and they probably still want to see the product themselves and make the final call in person. That means sites need to support varying degrees of "delegated authority," rather than treating "agentic commerce" as one uniform interaction mode. The spectrum below maps the territory between those two ends.

Sanchez says whether sites end up "splitting into two versions" — one for humans, one for agents — is still unclear: "But clearly, you have to serve both."

The Keys That Let Agents In

If a site has to serve both humans and agents, it needs interfaces that let agents get in the door and get things done. The source article names three emerging enabling technologies, each mapping to a different segment of the spectrum above, each solving a different problem.

Sanchez says the end state might be a site that carries both visual components and agent-callable tools side by side, two experiences coexisting; or it might be a human-facing site paired with an agent-to-agent backend service. "This is still something everyone's figuring out," he said, "but it's safe to say there will be agent-facing targeting." In his view, traditional websites probably won't disappear entirely, but their role will change: from a pile of fixed pages waiting for every visitor to browse, into a governed content-and-interaction system that assembles the right interface on demand. That's the direction Adobe is exploring.

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

The system treats your browsing behavior and search terms as signals, and the moment you open the page, it reassembles the whole thing into a version tailored to you. On the same coffee-maker site, a regular visitor and a visitor who searched "camping" see two different websites.

Same domain, same business — copy tone, top-listed products, and recommendations below all changed with your intent (there's a second, more direct entry point too: type "Europe AI conferences" and the page gets assembled around that query on the spot). So how does the machine actually put together the whole page in a second or two?

The system first sorts your signals into an intent (exploring, researching, or ready to buy), then assembles the page around that intent. The most important step is hidden in the middle.

For product people, the first walls real-time page generation hits are speed and cost. Adobe drew its own red lines, and pushed what used to be an expensive proposition — reassembling a page for every visitor — down to a cent or two.