Web3 Never Happened. Maybe Now?

Web3 has been through several incarnations, but now that design and content have become especially cheap to produce, I expect a real revolution in what we call Web3. My bet is that the web will move towards personalisation, blurring the boundaries between services, chats and conventional websites. Whether that turns out to be true or not, I believe that with AI and ideas taking less time than ever to become reality, we are standing at an open door to the next major change in how we experience the internet.

From reading to interaction#

I count versions by what people see on a page, how they relate to it and perceive it, and what the web means to them as users, rather than by its protocols. The first web was all about publishing and reading. Someone published a page and everyone else opened it. O'Reilly described the second web more broadly, as a platform and an architecture of participation. People contribute their own content, the system provides the structure, and what the next visitor sees depends on what previous visitors have done.

Web 3.0 has had at least two lives. In public perception, the idea was reduced to blockchains and tokens, although the original thesis was more specific. Both attempts defined the next web through data, protocols and technology, but I think what matters more in practice is what changes on the page a person opens.

The first grew out of the Semantic Web. In 2001, Berners-Lee, Hendler and Lassila described a web whose content machines could interpret, and by 2006 the label Web 3.0 had attached itself to that direction. The second began in 2014, when Gavin Wood used the same name for a different architecture in which identity, data and transactions no longer require trust in a central platform.¹

I think that in the third web, a page is created for a particular user and situation, and composed around the question I arrived with. It does not exist in a finished form. The user's context determines which argument comes first, how deeply it is explored, which material appears and which stays hidden. The site itself is assembled when I arrive, within the possibilities its author has defined in advance.

A website made for you#

The dynamic web has existed for twenty years. The YouTube homepage, Instagram feed and Amazon homepage are assembled from a person's previous behaviour, and millions of people see different versions of the same product at once. But these systems have an enormous inventory of finished pieces of content, which personalisation selects and ranks.

The corporate website was an exception for an understandable reason. Content was expensive, the user's intent was unknown, and you had to build one system for everyone while keeping it compact enough not to overwhelm the navigation, using the menu as a set of reference points within the information. A site can still be dynamic through a CMS, changing prices, dashboards or A/B tests, but in all these cases the final states were designed in advance or the algorithm was deliberately given narrow boundaries. Writing, approving and maintaining a separate good website for every intent was too expensive. So web design became a craft of selection. Identify a clear set of audiences, prioritise their interests and find one composition that loses as little as possible when averaged across them. Half of a project's success lay in cutting what did not belong before it was written. That is a production constraint.

When the cost of one more variant approaches zero, the constraint disappears, and we can create a separate page for each user, taking their likely intent, profile and explicit request into account. For this to work, we need to define general rules instead of final content for every case. These rules establish what counts as a fact, what position the site takes, which arguments are acceptable, what must survive every render, which visual language belongs to it and which changes to the interface are allowed. The page is assembled from this material later, when a particular person arrives. The author defines the space the system is allowed to occupy.

Web 3.0, in my definition, is a site whose content is composed for a particular reader, while its form, order and position belong to the author, the person responsible for them. The Brand Constitution I described earlier can help define the principles these generated explanations should follow.⁴

Customisation should not be guesswork. We can use signals already used in online advertising, or context we can observe directly, such as the campaign behind a banner click, the referral source, audience data or a request in the visitor's own words. These describe a situation nobody has written a separate page for, because creating one for every case was impossible. The signals have different levels of certainty, and the site should make its interpretation easy to correct. Someone who arrived through a pricing campaign may still need an introduction, but ordinary navigation may be enough to get them there. The opportunity is to make the explanation more specific once the visitor has indicated a direction.

The way the content is presented can follow the task too. A question about switching from one product to another may call for a comparison, while a question about cost may call for a calculator using verified prices. The author can define these possibilities in advance and let the site choose and compose the relevant explanation during the visit. Google Research explores the broader possibility of generating interfaces, tools and simulations in response to a prompt.²

Among existing products and technologies, Optimizely already describes assembling content, layout and next steps around visitor intent within boundaries set by the marketing team. I discovered them while preparing material for this article. They are making a really interesting product, and credit to them for it! Meanwhile, OpenAI's Sites makes creating another site from a prompt cheaper. I think they are doing something logically similar, but approaching it from the other side. They make creating a site so accessible that it takes minutes to publish, rather than making existing sites flexible.³

As the web becomes more personal for a human, its presentation matters less to a machine reading it. An agent does not need my hero section. It needs objective facts, a position, capabilities, prices, limitations and the provenance of information in a form it does not have to extract from a layout. Conveniently, this requires almost the same set of principles and source material as generation within the site itself.

The llms.txt proposal points in this direction with a simple, structured description of a site for language models. It is still a proposed convention rather than a standard.⁸ My own sites provide such a document alongside the version for people.

One source gets two channels. For the agent, it is a machine-readable source of truth. For the person, it is an authored expression of that same source. When the second channel becomes adaptive, the architecture falls into place almost literally. The source of truth, constitution, reader's context and current intent go in, and a rendered page comes out. In the old production model, the result of the work was a page. Now it is a system that produces pages which remain faithful to one source.

AI agents will not kill the conventional web#

You could argue that an AI chat already does essentially the same thing if you ask it about a product or service, have it study the website and prepare an answer specifically for you, and you would be partly right. Chats are getting richer, with widgets, tables and artifacts. Perhaps a year from now a chat will be able to render almost everything a website can. But that does not change the conclusion, because the problem was never the limitations of the markup. A chat does two things at once. It interprets the material in terms I understand and presents the result in a chat window. The interpretation matters, but the presentation simply came along with it because chat was the closest available format, not because it was the right one.

While the semantic layer of understanding the reader and rephrasing for them moves to models, the interface layer that determines how the result is presented can remain with the site's author. Even if chat interfaces become richer and more personal, they will have to remain uniform, leaving no room for the full differentiation I believe is still needed. An assistant's job is to be as predictable and clear as possible, regardless of the source. If I open ten companies through one chat window, each adapts to that environment and loses its distinctive form. A website gives a particular company, publication or person a space with its own character, mood and feeling, allowing me to form a richer personal memory of it.

Ten products studied in one week through one window blur into one after two weeks. You remember the thought but not whose it was, because you do not associate it with a distinctive visual identity and experience. Memory research calls this source monitoring, the ability to attribute a memory to where it came from, which depends on the contextual features surrounding the content (Johnson, Hashtroudi and Lindsay, 1993).⁵ Remove those features and the content remains while the authorship disappears. Differences between sources do semantic work, attaching the content to an address so you can return to it and reconsider it.

Standardised formats exist where a reader needs to understand something complex quickly and the cost of error is high, as with a scientific paper, a blueprint or a safety instruction. But a blueprint sells nothing. A website sells a position, and a position requires an identity, which is why newspapers, with one of the most easily standardised formats imaginable, have looked different from one another for decades. Telegram lets you set a separate background for a particular conversation. When dozens of conversations are open, different backgrounds reduce the chance of writing to the wrong person and also convey a tone, the way a blanket on a sofa conveys a sense of home. Form carries meaning even where the design is intended to be just text.

Branding has a theory for this. Romaniuk describes distinctive assets as memory cues. A colour, shape, character or unusual detail begins to bring a brand to mind without anyone reading its name. There is a gradient here. The easier a mark is to understand, the less memorable it is, and the more unusual it is, the harder it is to read.⁶ A website does the same at a larger scale through typography, rhythm, motion and the strangeness of a single detail. The conclusion runs against intuition. When content becomes variable, recognisable form becomes proof that we are still looking at the same source. The adaptive web does not remove authored form but makes it necessary.

Spatial memory and expected behaviour matter more than adaptability#

A site can change its vocabulary, examples, depth and order while keeping its claims and commitments consistent. A visitor asking for a fully managed service should not turn a self-service product into one. The team can deliberately reconsider its position, but the model should not reconsider it to please the next visitor.

Beyond generation itself, elements in this kind of interface should appear where you expect them, where you remember them from last time. In a CHI study, Findlater and colleagues improved menu selection speed through adaptation that kept elements in place and changed how attention was drawn to them.⁷ The study concerns menus, and the example can be generalised. Adaptation can work through emphasis while preserving learned locations. A website that rearranges the apartment every morning makes its visitor look for the light switch again. This brings us to the question of where the boundary lies between an adaptive website and a randomly generated page that looks different every time you open it.

My bet is on a flexible structure rather than specific content, and the rules you set for that content may address this risk. The system acquires semantic, visual and interaction invariants, with a hierarchy between them that makes it clearer which things are more likely to change.

• What the product claims and holds to be true
• Which features preserve the source's identity when the content changes
• What remains stable enough for a person to learn the product and trust their previous experience

Within these boundaries, the system can adapt the material by choosing what to emphasise, what to move to a secondary level, which example to show, which workflow to suggest and which temporary tool to assemble for the task at hand. This is the same question I asked about brands in the second article. What needs to remain unchanged so that everything else can change? This time, though, it applies to one screen for a few seconds rather than a system for a few years.

In products, the same logic becomes more radical because large professional systems currently show people the product's ontology rather than their current work. Photoshop, Salesforce, Blender and even, to a lesser extent, Ableton feel complex, largely because of the legacy of previous versions and the habits of their established users. They manage this complexity through navigation, permissions, presets and workspaces. All of these are ways of manually narrowing the range of possibilities, and most people never touch the settings.

An adaptive system could avoid showing you two hundred functions and teaching you to find the ones you need. It could tell you that two deals have changed status, one client has replied to yesterday's proposal and another conversation has stalled, then offer three actions that make sense now. The interface stops being a map of the product's capabilities and becomes a projection of the system's state onto the person's current intent. I am not ready to predict what this will look like. I suspect good adaptive software will be visually much more stable than we imagine today. It will have a stable framework, adaptive emphasis, generated material and, occasionally, a new tool generated specifically for the task.

For the same reason, I cannot yet describe how all of this works in a mobile app. Spatial memory carries more of the load there, the screen is small, and every decision about where to put your thumb has a cost. I expect the adaptive layer to be thinner there.

What I have already tried to build#

The idea for this article came to me while I was working on a website for my experimental side project. The site presents one position across seven screens arranged in a repeating orbital scroll. I could have left it that way, but I tried adding more perspectives to the narrative and wondered whether the copy could vary even more. So I built a simple example to show how this might work in practice. It works on two levels.

• The same page presents the copy from different angles. Because the scroll is infinite, you can scroll down and see the same screen again, but with different text.
• At some point, a discreet button appears in the footer, giving you a little wink and asking whether you have a particular idea in mind. Your answer provides the context for another render. The site rewrites its copy, trying to speak to you more personally and considering your idea through the lens of its own themes.

offpossible.com

Under the hood, the model receives the original claims and constraints. Each screen keeps its function and position and explains it through the material the visitor has just brought. The model is asked to find a connection they may not have noticed, use their vocabulary, avoid flattery and avoid implying support for an idea outside the studio's interests. The visitor returns to a similar design and the same structure, while I draw attention to the changed copy by inverting the colour palette and changing the background graphics.

The useful move here is the connection between a general approach and a specific problem. “Product thinking, design and strategy” become something a founder can get a taste of right there. The model has offered an interpretation that gives the visitor more to think about than a repetition of their own nouns. In my case, this connects to the brand message itself, which makes it work nicely as brand communication. And I am interested in whether people will keep answering the question, whether they will see a pattern and whether noticing it will change a decision.

So where is the web heading, and how far can this approach go?#

I do not know where the economic boundary lies. A site with a few hundred visitors can afford a model call for each of them. At millions of visits, cost, latency and caching start determining how the system works again. Generating a paragraph also has to justify the wait. A fixed pricing table or a direct link to documentation may already be the right answer.

I do not know how to show what has been personalised. Should visitors know their version is different, be able to open the canonical version or ask why they were shown this particular argument?

I ask myself how we, as designers, website creators and product founders, should adapt to these changes. What principles could we adopt today to make websites more useful, more interesting to interact with, more adaptable to change and faster to deliver? And should we do any of this at all?

The web is going through another evolution, and in its new incarnation, perhaps we should stop treating the page as the basic unit altogether. When a particular page does not exist until it is opened, the author has a different object to work on. It consists of the source, position and constraints from which the right page can emerge without becoming the wrong brand.

I see two directions here. On one side, the building blocks of design are becoming more universal and better documented so that AI agents can work with them. On the other, the web is moving towards greater personalisation and a more precise understanding of the individual user, rather than their cluster or group. This raises another question. Is the same thing happening to branding? Is it also moving towards something unique to each person while preserving a common framework? If so, where does the boundary between branding as a whole and the interface really lie? Perhaps we are entering an era in which customer experience becomes a more important discipline, with practical applications for more companies, than ever before.

References#

1. Tim Berners-Lee, James Hendler and Ora Lassila, The Semantic Web, Scientific American, May 2001 (W3C announcement). Gavin Wood, ĐApps: What Web 3.0 Looks Like, 2014. The numbering of versions in this article is my own shorthand. ↩
2. Yaniv Leviathan, Dani Valevski, Vishnu Natchu and Yossi Matias, Generative UI: A rich, custom, visual interactive user experience for any prompt, Google Research, November 18, 2025. ↩
3. Lauren Brennan, Dynamic experiences: Personalization has a scale problem, Optimizely, August 26, 2026. OpenAI, Sites: Build and share hosted sites in ChatGPT. ↩
4. Elisey Soloviev, The Playbook Ran Out and Brand Constitution. ↩
5. Marcia K. Johnson, Shahin Hashtroudi and D. Stephen Lindsay, Source Monitoring, Psychological Bulletin 114 (1993). ↩
6. Jenni Romaniuk, Building Distinctive Brand Assets, Oxford University Press, 2018. ↩
7. Leah Findlater, Karyn Moffatt, Joanna McGrenere and Jessica Dawson, Ephemeral Adaptation: The Use of Gradual Onset to Improve Menu Selection Performance, CHI 2009. ↩
8. Jeremy Howard, llms.txt, proposal, 2024. ↩