The Web iv clock: how close are we to the intelligent and agentic web?
My 2030 forecast for the intelligent and agentic web: what is changing, what the evidence shows, and who receives the value. Updated as the research develops.
I am more convinced than ever that we are in a painful transition from Web2 to what I call Web iv: the intelligent and agentic web.
In The agentic layer, I set out a hypothesis: by 2030, autonomous agents would dominate the transactional layer of the web. Websites would continue to exist, but increasingly as infrastructure supplying information and services to systems acting on our behalf.
The infrastructure is developing faster than I anticipated. The evidence is in access controls, payment arrangements, reporting tools and commerce protocols. At the same time, publishers face a harder question: what do they receive in return for making their content available?
I still use 2030 as my forecast. This page is where I will test it.
The Web iv clock
Forecast yearMy forecast is still 2030.
- Original forecast
- 2030
- Calculated estimate
- Model not yet built
- Evidence checked
What is breaking?
Value exchange, measurement and access.What is becoming possible?
Discovery, permission, interoperability and execution.What are people delegating?
Completed journeys, repeat use and human intervention.What I mean by Web iv
My shorthand for the major changes is publishing, participation and delegation. Web1 made information available through linked pages. Web2 made participation central, with platforms organising the connections. Web iv is the stage where intelligent systems increasingly interpret our intent and act across services to achieve it.
I am skipping Web3 in that sequence because, in my view, it never delivered the decentralised everyday web it promised. I explained that position in Web iv is the intelligent and agentic web. These are my descriptions of changing behaviour, not official version numbers.
The distinction between the human and functional web still matters. People will continue to gather, express themselves and enjoy things online. Delegating a hotel booking does not remove the desire to look at holiday photographs. It removes some of the work between deciding to travel and having a room.
The transition I am investigating is the point where delegation becomes the normal way to complete functional web journeys across a broad range of services. Finding, comparing, booking, buying and managing a service increasingly happen through an agent. The website supplies what the agent needs; it may no longer host the customer’s whole journey.
An AI-generated answer is evidence of a changing interface. An agent completing a task is a further step. The clock needs to distinguish them.
The bargain is coming apart
When I say Web2 is crumbling, I mean the value exchange, measurement and access to content are becoming less reliable. Websites can still operate while the businesses publishing them receive less value for their content.
Content creates value elsewhere
The early web had an unusually open foundation. CERN released the Web software into the public domain in April 1993, and later made a release available under an open licence. Anyone could build on it. That openness did not make every commercial exchange fair, but it gave people room to participate without asking a platform for permission.
My concern is that each successive arrangement puts more of the relationship, information and bargaining power inside the largest technology companies. The people supplying useful material become increasingly dependent on someone else’s terms.
Search offered a recognisable exchange: make useful information accessible and receive visitors who might subscribe, enquire or buy. It was imperfect, but the return had a form a business could measure.
That connection weakens when a platform can use the information to satisfy the customer itself. In Pew’s study of 900 US adults, traditional results received clicks on 8% of Google visits associated with an AI summary, compared with 15% without one. Links in the summaries received clicks on 1% of visits with a summary.
Those are observations from March 2025 browsing, with results reconstructed in April. They are not a universal estimate of traffic lost because of AI. They are a reason to examine the exchange. Content can be useful to a platform while sending very little business back to its source.
Measurement follows the visit, while value moves elsewhere
A website analytics system can report activity on the website. It cannot, by itself, tell the owner every time their information influenced an answer, a shortlist or a decision somewhere else.
That leaves a difficult distinction: has the business become less useful, less visible, or simply less able to observe what happened?
There are attempts to close the gap. Google now documents dedicated generative AI impression reporting for AI Overviews and AI Mode. That provides useful visibility. It does not establish which content caused a sale or how much value an intermediary retained.
The measurement problem is therefore more specific than everything being broken. Our familiar units of sessions, clicks and on-site conversions cover only part of the changing exchange. New reports can improve one part without reconstructing the entire journey.
Access becomes a negotiation
Content owners have to decide which automated systems can enter, for which purposes, and on what terms. Providers need the material that makes their products useful. I see a contest for control underneath the product announcements: who can obtain the information, who sets the conditions and who owns the customer relationship afterwards?
Cloudflare’s AI Crawl Control shows crawler activity and lets site owners set access rules. Paid customers can configure a payment-required response to a blocked request. That response does not mean a crawler has agreed to pay.
There are technical constraints too. Cloudflare’s crawling API can discover and render pages, but it respects site controls by default and cannot bypass its bot detection or captchas. New tools make crawling easier, while access controls limit which systems can obtain the content.
This is why I want to distinguish a broken page from a blocked request, an expensive crawl from a poor return, and a declared permission from an enforced boundary. They can all interrupt the old arrangement. They do not mean the same thing.
The new web is being built
Companies are building tools that let agents discover what a business offers, establish terms and complete a transaction.
Information can have terms and a price
Pay Per Crawl lets participating site owners price crawler access. Cloudflare still describes it as a closed beta. Its significance is the mechanism: automated access can carry a price rather than relying entirely on a hoped-for future visit.
Cloudflare’s September 2026 Pay Per Use beta addresses what happens after access. A buyer proposes a use and price; a publisher decides whether to participate. The buyer reports uses, and Cloudflare handles billing and payment.
There is a measurement dependency inside that arrangement: the buyer reports the use. The ability to invoice is progress towards an exchange. Whether the reporting is complete and the return worthwhile are further questions.
AI participation becomes a website setting
Google’s Search generative AI control allows site owners to manage inclusion in named generative Search features. Its documentation records worldwide rollout on 31 August 2026. This is separate from a universal permission for every AI use or model-training activity.
The existence of that setting matters. Participation in an AI-mediated experience has become something a website owner must explicitly consider, alongside the visibility and business consequences.
Commerce becomes a capability an agent can call
Google’s Universal Commerce Protocol, UCP, describes common commerce operations and a way for businesses to expose their capabilities. A system can discover what a business supports and use a checkout interface without navigating the same sequence of pages as a person.
Agentic Commerce Protocol, ACP, introduced by Stripe and OpenAI, addresses programmatic commerce between buyers, agents and businesses. Its September 2025 announcement accompanied an initial ChatGPT checkout deployment for US customers buying from US Etsy sellers.
There is evidence of implementation beyond the initial specifications. THE ICONIC documents a limited Australian UCP integration: selected products, linked accounts, Google Pay and single-item purchases. Its website and app remain its primary shopping destinations. This is a retailer’s account of its integration; I have not tested the purchase journey.
There are setbacks too. In March 2026, OpenAI said its first version of Instant Checkout lacked the flexibility it wanted. It shifted its focus towards product discovery and allowed merchants to use their own checkout experiences. That supports a more limited conclusion: AI can change how people find a product while the merchant’s website still completes the purchase.
These are concrete parts of the transition. They are not interchangeable, and a specification is not a measure of adoption. The maintained UCP and ACP references examine their developments and scope in more detail.
What connects them to my original hypothesis is the change of interface. A business exposes information and operations that another system can use on a customer’s behalf. The visual website becomes one of several ways into the business.
How the clock will work
I want a forecast that can be challenged and updated. Product announcements alone cannot tell us when people will adopt the new web.
The research separates pressure on the old web, readiness of the new infrastructure and actual adoption. A worsening bargain can encourage change, but it can also remove useful content or make access harder. More infrastructure can make delegation possible without persuading people to use it.
The first version retains 2030 as my hypothesis. It does not yet calculate an arrival date or percentage complete. Before building the model, I need to define which activities it covers, how much of that activity must be delegated, and for how long.
The intended milestone is sustained use across services, rather than a single demonstration. Evidence should tell us how often agents complete real tasks, where people intervene, whether they return to delegation, and how much activity still follows the conventional route. A customer confirming a consequential action is compatible with delegation; it is not the same as manually performing every step.
For each development, I record what happened, when it happened, the source, its scope and what it can actually establish. Announcements, beta services, production deployments and measured outcomes carry different meanings. Multiple reports of the same event remain one event.
As suitable data becomes available, the model should estimate when the transition will happen and show the uncertainty around that estimate. It must allow for a date after 2030, or for the transition not to happen within the period being forecast. Every revision will remain available for comparison with my original 2030 prediction.
I will review the evidence weekly. A review may leave the forecast unchanged. A meaningful revision will explain which evidence or assumption changed it. The structured evidence record exposes the same forecast status and source links for people using the material in their own research and for agents processing it.
What would weaken my conviction? Persistent failure to complete tasks economically, limited repeat use, incompatible systems that remain expensive to connect, or people consistently choosing direct interaction. What would strengthen it? Repeated successful delegation across independent services, clear authority and payment arrangements, and adoption that extends beyond a handful of platforms or early users.
Evidence of the transition
This record draws on my Signals, maintained guides and direct source research. I select developments relevant to the forecast; publishing a Signal does not automatically change it. Each entry links to the original source and, where available, the related Signal. Dates refer to the event or documented rollout, with the newest first.
-
W3C Strategy Team proposes principles for the agentic web
The proposal connects duties to users, consent, interoperability and creator terms, and recommends chartering an Agentic Web Working Group.
Why it matters. Shared rules could make delegation easier to support across independent services. This is a step in defining the infrastructure Web iv needs.
Related Signal: Who does the agentic web belong to?
Scope and sources
A discussion proposal does not establish adoption, enforceable content licensing or successful delegated journeys. The 2030 forecast is unchanged.
-
Cloudflare adds payment for content use
Pay Per Use connects buyer offers, publisher participation and payments for reported uses. It complements charging for access.
Why it matters. A commercial exchange can happen without a website visit. Whether it delivers worthwhile returns needs outcome evidence.
Scope and sources
Usage is reported by buyers. Beta availability is not evidence of broad participation or sustainable publisher income.
-
Google reports worldwide availability of AI controls and insights
Search Console documentation records worldwide rollout of the generative AI inclusion control and impression reporting.
Why it matters. AI participation and visibility are becoming explicit website-management concerns.
Scope and sources
Impressions are not completed agent journeys. The inclusion control covers named Search features, not every Google AI use.
-
THE ICONIC documents a limited Australian UCP integration
The retailer describes single-item purchases of selected products through Google checkout, requiring linked accounts and Google Pay.
Why it matters. A retailer describes using the protocol for a defined purchase journey. Its website and app remain its primary shopping destinations.
Related Signal: THE ICONIC documents a limited Australian UCP integration
Scope and sources
No purchase test or transaction volume is included here. The date is the documentation date: its stated launch weekday and numbered date conflict.
-
OpenAI changes its approach to checkout
OpenAI says its initial Instant Checkout lacked the flexibility it wanted. It emphasises product discovery and allows merchants to use their own checkout experiences.
Why it matters. AI product discovery does not necessarily replace a merchant website or complete a delegated purchase. Early approaches can be revised.
Related Signal: ChatGPT shopping shifts its emphasis towards discovery
Scope and sources
This explains a product decision. It does not provide a transaction-success rate or establish that all agent checkout approaches will fail.
-
Cloudflare exposes website crawling through an API
The Browser Rendering crawl endpoint discovers and renders pages and returns HTML, Markdown or structured JSON.
Why it matters. Crawling is being adapted to agent workflows as well as restricted by publishers.
Scope and sources
It respects site controls by default and cannot bypass Cloudflare bot detection or captchas. No success-rate test is included here.
-
Google introduces Universal Commerce Protocol
UCP supplies a shared commerce interface, including discovery of business capabilities and checkout operations.
Why it matters. Businesses can expose functions that agents can use without reproducing the human website journey.
Related Signal: Google announces the Universal Commerce Protocol
Scope and sources
A specification and worked example do not establish deployment breadth, reliability or transaction share.
-
Stripe and OpenAI introduce Agentic Commerce Protocol
Stripe announced ACP alongside an initial ChatGPT checkout experience for US customers buying from US Etsy sellers.
Why it matters. The transaction can move into the interface where a customer is already making a decision.
Related Signal: ACP launches with Instant Checkout in ChatGPT
Scope and sources
This is the historical launch scope. It is not a current market-share estimate or proof of fully autonomous purchasing.
-
Cloudflare moves from AI auditing to access control
AI Crawl Control reached general availability, with custom payment-required responses for paid customers alongside the separate Pay Per Crawl beta.
Why it matters. Automated access increasingly has an explicit policy and commercial boundary.
Scope and sources
A payment-required response is not a completed payment or evidence that a buyer accepted the terms.
-
Pew finds fewer outward clicks alongside AI summaries
The study compared Google search activity from 900 US adults during March 2025, with search results reconstructed in April.
Why it matters. The relationship between supplying information and receiving visits needs scrutiny.
Scope and sources
The association does not establish a universal causal loss of traffic, and it does not measure agent-executed tasks.
Who receives the value?
The timing question matters. So does the distribution of value when we get there.
A more capable agent can make life easier for a customer while strengthening the platform sitting between that customer and everyone supplying the service. Technical progress does not automatically improve the bargain.
The question in The agentic layer remains central: if a system uses someone’s content, data or service to produce an outcome, how does value return to the source? Attribution, referrals, reciprocal access and payment can each contribute. Meaningful permission and a practical ability to negotiate matter too.
That argument helped develop my work on Odexa, where I am Foundation Chair. It also informs the practical AI Discovery guide: businesses need to decide what they are making available, for which uses, and what return they expect.
My conviction about the direction has strengthened. My forecast remains 2030. The purpose of this living article is to make the evidence, the disagreements and any change of mind visible.
Agents should make services easier to use. The people supplying the content, data and services should receive a fair return. I want to track whether both are happening.
Updated 23 October 2026
Reviewed 21 October 2026