The web after websites: our vision for the agentic web

Five predictions

A vague vision costs nothing, so we are putting dates on ours. The predictions are technical, and the rest of the paper explains each one.

1. By mid-2027, most first contact with a B2B service firm happens without a human loading its homepage. By the time a buyer gets in touch, an assistant has already read the site and answered their question.
2. By end of 2027, an agent can find a site's action endpoint and connect to it without anyone wiring it up by hand. This is the hinge: every other prediction on this list depends on it.
3. By mid-2028, organic search sends B2B service firms less than a third of the traffic it sent in 2024. On research questions such as comparisons and how-tos, the click is all but gone: the assistant answers, names a few sources and keeps the buyer in the chat.
4. By 2029, a real share of quotes, bookings and orders in professional services begins agent-to-site, with no page view anywhere in the transaction. The first thing a firm learns about the deal is that it has already started.
5. By end of 2029, new commercial sites take more agent traffic through structured endpoints than through browsers imitating a human. The bridge starts giving way to the road it was built for. As far as we know, we are the first to put a date on that.

They come from our own research and from what we see across the B2B sites we build and run. Some will be wrong, and the dates are there so you can check.

Three things that already broke

The first break: machines became the majority of the web. In June 2026, Cloudflare reported that automated systems generate 57.5% of the HTTP requests on its network, against 42.5% from humans, the first such crossover on record. Its CEO, Matthew Prince, had predicted the moment and was still 18 months late, overtaken by agentic traffic that grew roughly 7,851% year over year. The asymmetry explains the speed: a human shopping visits perhaps five sites, while an agent running the same errand visits five thousand, by Prince's own illustration. Multiply that across every errand on earth, and the surprise is that the crossover took until 2026.

The second break is closer to home: rankings still work, but they deliver fewer clicks. Google answers a growing share of queries on the results page itself, AI assistants answer them before a results page loads, and a first-place ranking increasingly buys a citation instead of a visit. Publishers see it in their server logs at industrial scale: crawlers pull thousands of pages for every referral they send back. The old bargain that paid for the open web, content traded for visits, is being rewritten.

The third break is our own. For a handful of the sites we run, we track whether assistants name the business when a buyer asks the question it sells into. One now surfaces in roughly one ChatGPT answer in three. That visibility is real and winnable, but it sends few visitors. Across the portfolio, four months of being named brought 322 visitors, under half a percent of human traffic, and the monthly share did not grow: 0.63% in May, 0.28% in June, 0.35% in July, 0.29% in August. Industry reports call this traffic explosive. On our servers it held still.

The few who do arrive are the most engaged visitors of any channel: roughly half engage with the page, against 38% across the portfolio, and the gap holds in every window we measure. The assistant has done the qualifying. It answered the question and forwarded only the people who still had a reason to visit.

Together they show what a transition looks like: the reader of the web changed before the web itself did.

Search changed masters

Ask an AI assistant a question and it searches. Behind most answers sits an ordinary query to Google or Bing; the assistant reads what ranks and answers from it. So ranking still matters, but it now buys something new: being the source the machine reads, cites and trusts first. The work is the same; the reader is now a machine.

That reader also has different taste. Your hero animation and art-directed imagery mean nothing to it. It wants the answer to a specific question, stated plainly, in a structure it can parse without guessing. Most B2B sites are autobiographies and few answer questions, which makes answering them the cheapest competitive advantage on the internet right now.

The arrangement is also temporary. Search engines exist to help humans find pages, and agents rely on them only until they can reach a site's answers directly. When that becomes possible, the detour through search collapses. Every euro you spend on ranking today is a loan against the next web, and it pays out only if what you publish is worth reading to a machine.

The catalogue replaces the index

This can go two ways, and most of the industry is building for the first.

In the first, agents keep using the human web: they open a browser, render pages and click buttons like a person. Several major labs ship this today, and it works. Enterprises run it in production at scale, mostly against decades-old portals and legacy software that will never get a modern interface.

We are building for the second: the web grows a machine layer. Sites expose structured endpoints, agents call them directly, and nothing renders, because nothing needs to.

We bet against the browser agent even though it works. It works the way a horseless carriage worked. The first cars were shaped like carriages because that was the only form anyone knew, and the first mobile sites were desktop pages shrunk to fit. A browser agent imitates a human for the same reason: the human interface is the only door most sites leave open. So it renders pixels, reads layouts and clicks through forms to reach information a machine could request in a single call. Inside enterprises it thrives where no better path exists or is coming; a government portal from 2003 will never ship an endpoint. The public web runs the other way. There, every browser agent is proof of demand and proof of waste. The day a site offers an endpoint, agents stop rendering its pages, just as developers stop scraping a site once it offers an API. The browser agent is the bridge that makes its own destination inevitable.

That destination is close, because the hard part has shipped. The Model Context Protocol (MCP) gives AI systems a standard way to use external tools; it runs in production across the major AI products, and thousands of companies operate MCP servers. The interaction layer is built. The missing piece is discovery: connecting an endpoint to an AI tool is still done by hand, and an agent cannot stumble onto your MCP server the way a browser stumbles onto your homepage.

The web has solved this problem before. DNS, links and crawlers made pages findable; the same pattern, pointed at capabilities, makes endpoints findable by any agent. The first attempts are live. In June 2026, Google, Microsoft and nine other companies published Agentic Resource Discovery (ARD): a site lists its machine interfaces in one catalogue file at a fixed path on its own domain, and registries crawl those catalogues the way search engines crawl pages. A W3C community group led by Google and Microsoft engineers is also drafting WebMCP, which lets a site register callable tools directly in the page. Both are weeks old and adoption is near zero: a June census found that even the companies behind ARD were not yet serving a catalogue file. That makes this the ground floor of a standard, and the window to be early is still open. Discovery is the contested layer, and whoever wins it owns the next web.

Once discovery works at scale, the website turns inside out. Today a site is a stack of pages a human reads and a machine scrapes. Tomorrow it is a set of capabilities a machine calls: answer this question, check this availability, produce this quote, book this meeting, place this order, put these two people in touch. The human pages stay, the way phone lines stayed, but they become the side entrance. Search engines rank pages; the next layer catalogues capabilities.

The early failures make the case better than the successes. OpenAI launched Instant Checkout in September 2025 and pulled it roughly five months later with almost nothing to show: tax handling unsolved, inventory out of sync, fewer than fifteen merchants live out of millions eligible, and shoppers who researched in the assistant but paid in the checkout they already trusted. The interface failed; the layer underneath held. In March 2026, after AI-attributed orders rose elevenfold through late 2025, Shopify switched agent-facing storefronts on by default for eligible US stores. By the first quarter of 2026, its AI-driven traffic was up eightfold year over year and orders from AI-powered search nearly thirteenfold. The payment protocols survived by consolidating under industry bodies: this spring, Google handed its agent payments protocol to the FIDO Alliance, alongside some sixty other organisations. The early web went the same way: push channels died and HTTP won.

The demand is here; only the rails are late. Ninety-four percent of B2B buyers now use generative AI as a research tool, and most say they would place large orders through self-service with no sales call. What happens next is contested. Gartner projects fifteen trillion dollars in B2B purchases flowing through agents by 2028; soberer forecasts land two orders of magnitude lower by the end of the decade. The direction is unanimous; the speed is anyone's guess. Meanwhile, the browser agent is about to get more common. No mainstream agent calls a site's structured tools today, and Google shipped an agentic browser mode in Chrome this year that will add to that traffic as it rolls out beyond paid tiers. Our bet is that browser agents are the peak of an existing approach rather than the start of a new one, and that the structured path passes them on new commercial sites once discovery is solved.

One caveat that usually gets skipped: the destination is clear to us, the road much less so. Discovery still runs through search, and search belongs to two companies. How fast the catalogue layer arrives depends on how Google, and to a lesser extent Microsoft and the AI labs, play their hands. Anyone promising a certain timeline is selling something. We give you the direction and take the risk on the dates.

What this means for a business

Preparing for the agentic web takes more than a checklist or a file you upload, and anyone selling a one-step cure is selling snake oil. Four things hold whichever protocol wins, and all of them are familiar work.

Know exactly who you serve and the question they arrive with. An agent carries a real buyer's constraints, and a site pitched at a vague audience cannot answer a specific question. Context is the asset; the rest is rendering.

Build the site around your buyers' questions. The next reader is running an errand for someone. It needs your pricing logic, service boundaries, availability and proof, laid out plainly enough to find without interpretation.

One of our clients runs Meta ads for a single kind of business: high-ticket home improvement, the contractors and custom-build firms where the average job is large and trust decides everything. Because they know their buyer that precisely, we built the site around the questions that buyer asks, backed with the proof those buyers look for, instead of another page describing the agency. Now, when someone asks an assistant for the best Meta ads agency for contractors in their area, that firm comes up. It took a clear foundation and the right context.

Keep the content legible to a machine: clear structure, real answers in full sentences, and every important point in the text itself, outside images and animations. When a capable reader cannot lift your offer off the page in one pass, fix the structure.

Run your web presence as an ongoing system. The four-year rebuild assumes a web that sits still. A site now has to stay current against a moving target: new formats, new readers and new interfaces as the catalogue layer settles. The firms that win a transition are the ones standing in the right place when it arrives, and they are rarely the ones who called it.

All four describe what a good website should always have been. The agentic web keeps the same standard of quality and removes the option of faking it.

Our position

We build websites for a living, and here we are arguing that the website as everyone knows it is a passing format. A firm that lives on four-year rebuilds has every reason to keep quiet about that. We are saying it out loud because we read the traffic across our clients every week, and the numbers stopped telling the old story.

So we build for the reader we think is coming: structured context first, rendered pages second. Our Growth service, which runs on our own software, Orbit, keeps that structure current as the standards change, because a site built for this shift has to be maintained towards a moving target.

We published five dated predictions at the top of this paper. Twelve months from now we will come back and score them in the open. If we called it wrong, you will see exactly where.

Transitions never feel like transitions from the inside. The homepage still loads, the rankings still update and the analytics still fill with sessions. Meanwhile, most of the traffic is already machines, reading a web that was never built for them and waiting for one missing piece. The last shift of this size put the storefront on a screen. This one takes the screen out of the transaction. We would rather be early and wrong on the dates than late and right on the direction.

Notes

Sources are grouped by the claim they support. Each figure was checked against the cited announcement as of August 2026.

The crossover

Bots passed human traffic for the first time, 57.5% against 42.5% of requests to HTML content, roughly 18 months ahead of Matthew Prince's own SXSW prediction. Cloudflare, June 2026. https://searchengineland.com/cloudflare-bots-webpage-requests-479608 (also Tom's Hardware: https://www.tomshardware.com/tech-industry/artificial-intelligence/bots-have-now-passed-human-traffic-online-cloudflare-boss-laments-says-agentic-traffic-wasnt-expected-to-eclipse-real-people-until-next-year)

Agentic traffic grew roughly 7,851% year over year, and Prince's five-versus-5,000 illustration of the human-agent asymmetry. https://ppc.land/us-sends-53-5-of-global-bot-traffic-decodo-analysis-finds/

Search and referrals

The share of Google searches producing any click fell 22.9% between 2024 and 2026. SparkToro. https://searchengineland.com/google-zero-click-searches-2026-study-479717

Where an AI answer appears, organic click-through falls 38 to 61% and zero-click searches rose from 54% to 72%. SSRN / Seer Interactive. https://nobori.ai/blog/google-ai-mode-zero-click-rate-100m-users-2026

Google's AI Mode ends roughly 93% of its queries without a click, at more than 100 million users. https://www.omnibound.ai/blog/ai-seo-statistics

Discovery and protocols

Agentic Resource Discovery (ARD), a single catalogue file at a well-known path declaring a site's machine interfaces, published June 17, 2026 by Google, Microsoft, GitHub, Hugging Face, Cisco, Databricks, GoDaddy, NVIDIA, Salesforce, ServiceNow, and Snowflake. https://developers.googleblog.com/announcing-the-agentic-resource-discovery-specification/

A June 2026 adoption census found zero discoverable catalogues among 39 major sites, including the working-group members themselves. https://www.synscribe.com/agentic-discovery/agentic-resource-discovery

WebMCP, which lets a site register callable tools directly in the page. W3C Web Machine Learning Community Group draft, edited by Google and Microsoft engineers. https://webmachinelearning.github.io/webmcp/

Commerce

OpenAI launched Instant Checkout on September 29, 2025 and pulled it March 4, 2026 with fewer than 15 merchants live and tax and inventory handling unsolved. https://www.forbes.com/sites/jasongoldberg/2026/03/10/why-openais-checkout-retreat-spells-trouble-for-its-commerce-strategy/

Shopify switched agent-facing storefronts on by default for eligible US merchants in March 2026. https://aiadvantageagency.com/shopify-agentic-storefronts/ AI-attributed orders rose elevenfold and AI traffic sevenfold from January to November 2025 (Q3 2025 earnings, https://techcrunch.com/2025/11/04/shopify-says-ai-traffic-is-up-7x-since-january-ai-driven-orders-are-up-11x); by Q1 2026, AI traffic was up eightfold year over year and orders from AI-powered search nearly thirteenfold. https://finance.yahoo.com/markets/stocks/articles/shopifys-ai-push-sustain-more-161700877.html

Google donated the Agent Payments Protocol (AP2) to the FIDO Alliance on April 28, 2026, with sixty contributing organisations. https://blog.google/products-and-platforms/platforms/google-pay/agent-payments-protocol-fido-alliance/

Gartner projects $15 trillion in B2B purchases intermediated by AI agents by 2028; more conservative forecasts land near $144B by 2029, and Gartner separately expects more than 40% of agentic projects to be cancelled by 2027. https://www.digitalcommerce360.com/2025/11/28/gartner-ai-agents-15-trillion-in-b2b-purchases-by-2028/ and https://www.ibtimes.co.uk/future-b2b-agentic-commerce-challenges-1804342

94% of B2B buyers use generative AI as a research tool, and 73% would self-serve orders over $50,000. https://www.marketscale.com/industries/retail/zero-click-commerce-arrives-ai-agents-set-to-intermediate-15-trillion-in-b2b-purchases-by-2028

No mainstream agent calls a site's structured tools as of mid-2026; all still scrape or use computer vision, and Google shipped an agentic browser mode (auto browse) in Chrome in early 2026, available to paid tiers. https://studiomeyer.io/en/blog/webmcp-reality-check-may-2026 and https://www.openhermit.com/blog/browser-ai-agents-2026

First-party measurement

Luniq, 15 August 2026. Across a portfolio of small Belgian and Dutch business sites we operate and instrument with our own analytics, 29 April to 15 August 2026:

• AI assistants referred 322 human visitors, 0.38% of 83,526 human sessions, across the sites (ChatGPT 239, Claude 29, Perplexity 25, Gemini 22, Copilot 6, Poe 1).
• The monthly share did not rise over the period: 0.63% in May, 0.28% in June, 0.35% in July, 0.29% in August. Detection was widened on 5 August 2026, so the flat trend is not an artefact of weaker earlier measurement.
• Assistant-referred sessions engaged at 50.3% against a 38.5% portfolio baseline (two-proportion z = 4.35, p < 0.0001).
• Answer-surface tracking on one site recorded citation in 22 of 75 ChatGPT responses to buyer-intent prompts, 28 July to 11 August 2026.

Two limits. The measurement is a script that runs in the browser, so it sees only visits that render the page and run JavaScript. It records nothing about crawlers or server-log traffic, and no figure here describes crawler activity. The instrument is also young, live for roughly four months, so these numbers may still carry gaps or errors we have not yet caught.