Back to insights
3 October 2026 · Analysis
By Kenneth Bæver Bjerke, Diggs.no

Who gets to define web 4?

The term web 4 is used for different things. The EU described it in a strategy in 2023. Ukraine, the United Kingdom and the United States have since put language models in front of the state's services. My assessment is that such entry points will shape what web 4 becomes, more than definitions on paper.

Revised 4 October 2026: My job title has been corrected from “Chief Product Officer (CPO)” to “product manager”.

I have long used “web 4” for the agentic web: a web where language models act on our behalf, and where we visit websites ourselves to a lesser extent. That is my use of the word, and it is not generally accepted. The European Commission used “Web 4.0” in a strategy from July 2023 for virtual worlds and the merging of the physical and the digital.

What is unsettled is the starting point for this text. The question is who gets to settle it, and by what means.

Names are given in hindsight

I have a hypothesis, and it rests on only two cases: the eras of the web get their content from what comes into use, regardless of what was promised.

“Web 2.0” became a widespread term around 2004, when the services for user-generated content already existed. The name described something that was in operation.

“Web3” went the opposite way. The term was launched in 2014 as a promise of a decentralised web, built on blockchains. My assessment is that most people today associate it with cryptocurrency. The promise gave the era a name, and the use gave it its content.

Two cases are thin evidence. I use the hypothesis to ask a question about web 4: what is it that is now coming into use?

The protocols are open

Language models have become everyday tools. Statistics Norway (SSB) reports that almost five in ten Norwegian enterprises with at least ten employees use AI in 2026, compared with three in ten the year before. The figure applies to AI in general, and enterprises with fewer than ten employees are not included.

The models have also become agents. They are connected to tools and data through the Model Context Protocol (MCP), which Anthropic launched in November 2024. They talk to each other through Agent2Agent (A2A), which Google launched in April 2025. Both have been transferred to the Linux Foundation, A2A in June 2025 and MCP in December 2025.

Measured against the doctrine, this is a good development. The protocols are open and are not owned by a single provider. But open protocols only determine how agents talk to tools and to each other. They do not determine who operates the agent the user meets.

The door

By “door” I mean a layer with three functions: it interprets the question, selects the sources and formulates the answer.

Such a layer in front of the web already exists. Search engines have selected sources for us for decades, and since 2023 they have also given formulated answers. I have looked at three states that have put such a layer in front of their own services. It is not a complete overview.

Ukraine launched Diia.AI in September 2025. The information below comes from the Ukrainian authorities and from coverage that builds on them.

  • The assistant answers questions and carries out services directly in the conversation, including ordering an income certificate.
  • It is built on Google's Gemini 2.0 Flash (OECD.AI) and is connected to the state's systems through MCP.
  • The authorities have said that they are building a national model (Kyiv Post). It is to be built on Google's open model Gemma (Kyivstar).

The United Kingdom launched GOV.UK Chat in the GOV.UK app on 14 May 2026, after a soft launch in March.

The United States launched america.gov on 29 September 2026.

  • The portal answers through a single chatbot and, according to the authorities, draws information from around 29,000 public websites (FedScoop).
  • The answers are produced by Google's Gemini and xAI's Grok, according to design chief Joe Gebbia (CNBC).
  • The executive order makes the portal the common entry point (“single point of entry”) for federal online services with more than 100,000 users a year. Tax filing, defence and intelligence are exempt.
  • The order promises that existing channels will remain, including the agencies' own digital services, and that the agencies keep control of their own records, systems and decisions.
  • The first version answers questions. Transactions, including Medicare enrolment, have been announced (The White House).

The three differ in ambition. The British service is an optional aid that answers from guidance. The Ukrainian one carries out services. The American one is to be the common entry point, with login and transactions. China has long had a combined entry point in WeChat, a private app built with menus and mini programs. I mention it only to show that gathering services behind one entry point is an old idea.

The promise of correct answers

The British publish an accuracy score from their own tests, 90 per cent, and say that the service provides information and not advice. Around the launch, the tax expert Dan Neidle found answers he called misleading (LBC), and the organisation LITRG found the same in its own tests.

The Americans promised more. Gebbia said in an interview with CNBC on launch day that the answers are drawn only from official sources, and that there is therefore no chance of outdated or incorrect information. The first tests show the following:

  • AP reported that the chatbot changed its answer while the president was still speaking. It first answered the question of who won the 2020 election, and later rejected the question as political.
  • Axios found that different users got different answers to the same question, and pointed out that models, instructions and guardrails are not publicly known.
  • CBS News asked 37 health questions. Many answers were correct, but several became unclear because the official sources contradict each other.

I have not examined independent tests of Diia.AI.

The British tests found misleading answers. The American ones found varying and unclear answers. Official sources are therefore not enough to guarantee quality. The difference between the two lies in what is promised. A promise of correctness must be verifiable answer by answer. The American order itself requires the technology to be accurate, reliable and transparent, and that is a requirement the portal can be measured against.

I know the problem from my own work. The Scorecard on diggs.no is a prototype that lets a language model assess software against the doctrine. In September, I found that for up to five months the doctrine documents had never reached the model, and that the evaluations were stored without a prompt version. I paused the tool and described the findings in “The Voice in My Own Code”.

From quality to control

What is documented above is a quality problem: the answers vary and are sometimes misleading. That it is also a question of control is my inference. It concerns what can happen. I have not observed it, and GOV.UK Chat has been in operation for several months without my having found signs of it.

The inference is this. The layer that interprets, selects sources and formulates influences what the citizen gets to know, without owning the services behind it. How great the influence becomes depends on three factors. Two are technical and one is organisational:

  • The model: it can be replaced or updated without the answers being marked with a version.
  • The instructions: they can be changed without the public seeing it.
  • The alternatives: the more that can only be done through the common entry point, the less the other channels are worth. That is an organisational decision.

The American order promises that the other channels remain. The doctrine asks about something more: whether the way out is realistic. A channel can exist formally and still become harder to find and less well maintained, or lack functions that the common entry point gets.

Three checks can clarify how open such a door is:

  1. Does the citizen's own agent reach the services behind it on the same terms as the state's?
  2. Are instructions and model versions published, and can every answer be traced to them?
  3. Do the agencies' own websites get the same functions as the common entry point, at the same time?

Four questions, two levels

I put four questions to the three examples, based on what is public:

  1. Who supplies the model? American companies in all three: Google in Ukraine, Anthropic in the United Kingdom, Google and xAI in the United States. Jurisdiction also depends on contract, place of operation and data access, and I have not examined that.
  2. Can the model be replaced? The British have changed provider once, between trial and operation. Ukraine is planning a national model on Gemma. The United States uses two providers, but I do not know whether one can replace the other. Open protocols can reduce the cost of a switch.
  3. Can an answer be verified? The British publish model and provider. Whether each individual answer can be traced to model version and instructions is something I have not found documentation of for any of the three.
  4. Is the way round realistic? In the United Kingdom, yes: the service is optional, and the pages it answers from are openly available. In the United States it is promised. For Ukraine I have not examined it.

The doctrine has two levels. The target is that a European state is not dependent on a model under another jurisdiction. The everyday rule, pragmatic patriotism, says that such a dependency can be temporarily acceptable when it is documented, has a defined criticality and has a concrete and funded exit plan with a time frame and a named owner. Ukraine's plan and the British change of provider show that the question has been asked. They do not show that the conditions are met, and I have no basis for assessing that for any of the three. Question 3 is the one I could not answer myself for the Scorecard, and it is part of what my R&D project is to investigate.

So who gets to define web 4? What follows is reflection and not findings. State entry points are only one part of the web, and browsers, search engines and personal agents may come to matter more. But here the direction becomes visible: in the three states I have looked at, the layer that interprets and formulates has been put into operation, while the EU's definition from 2023 is still a document. My assessment is that web 4 will get its content from such entry points. For Europe, the question is then not what we call the era, but whether we can answer the four questions for the entry points we build ourselves.

Disclosures

I am product manager at Akari and built Akari Engine. Akari sells Google Workspace. The Scorecard on diggs.no uses Gemini, the same model family as america.gov and Diia.AI. The tool is paused. The text was edited with Claude from Anthropic, the company that launched MCP and that supplies the model behind GOV.UK Chat.

How this text was made

The text was written together with Claude, which has been the editor and wrote the drafts from my notes and a longer conversation. Each draft was given to four other models: ChatGPT, Gemini, Grok and Le Chat. They were given the site's editorial rules and asked for a fact check, the weakest link, the strongest counterargument and breaches of the rules. They were not told who had written the text. I set a limit of four rounds.

Three times, Claude put in a claim that something was the only, the first or new, without having checked it:

  • Round 1: that the executive order makes america.gov the only entry point. ChatGPT found the error in the order. The other three marked the claim as confirmed.
  • Round 2: that it was new for a state to operate such a layer. ChatGPT pointed to GOV.UK Chat.
  • Round 3: that the United Kingdom was first in Europe. ChatGPT pointed to Ukraine.

Round 4 produced no new errors of that kind. It produced clarifications: that formulated answers in search came in 2023, that Ukraine's national model is to be built on Gemma, and that the exit plan in pragmatic patriotism must be funded.

Le Chat contributed the objection about the search engines, the distinction between quality and control, and the criticism that the British accuracy figure was taken for granted. ChatGPT pointed out that the ending skipped pragmatic patriotism. Grok found no factual errors in any of the rounds.

Everyone who answered thought that the image of the door breaks the rule of mechanisms over metaphors. I have kept it and defined it with three functions. That assessment is mine, and so is the responsibility for the text.

The result says something about this review. It says nothing in general about which models are reliable.