In brief: The agentic web is the phase of the internet in which AI agents no longer merely read websites but operate them on behalf of people: searching, comparing, adding to the basket and increasingly paying. Cloudflare now measures more than half of all internet traffic as non-human (as of July 2026). For website and shop owners the question shifts: no longer whether machines turn up, but whether they can get anything done once they do.

What is the agentic web?

The agentic web is the third stage after the PC web and the mobile web. A survey by Yingxuan Yang and colleagues (arXiv 2507.21206, July 2025) describes the transition along three dimensions: intelligence, interaction and economics. The core idea is delegation. Users no longer supply clicks, they supply intent. Instead of comparing five shops, someone tells their assistant to find the cheapest offer for a given product and put it in the basket.

That changes the metric that matters. The interesting figure is no longer only how many people visit your site, but how many of your visitors are still people at all.

How many visitors are already machines?

In its bot report of 1 July 2026, Cloudflare states that more than 50 per cent of internet traffic is non-human. Two further figures from the same report tell businesses more than the headline does: 52 per cent of all crawler requests served AI training in June 2026, up from 22 per cent in spring 2025. And for every hour people spend searching for information, only around 15 minutes are spent on the open web. In heavily crawled categories, Cloudflare saw human traffic fall by up to 40 per cent within a single year.

Those figures say nothing yet about what these machines can actually do. That is the commercially more interesting question.

Can AI agents really shop?

They already handle the mechanical part. The University of Mannheim built WebMall, a benchmark that has agents shop across several online stores: find products, compare prices, add to basket, check out. Ralph Peeters and colleagues report a completion rate of 100 per cent for the basket and checkout task group with a current GPT model. A full purchase journey from search to completion succeeded in 75 per cent of cases.

Comparing is harder. Finding the cheapest variant across several shops succeeded in only about 63 per cent of tasks. The authors attribute this to price comparison posing a challenge of its own on top of product search. Effort and cost are unspectacular: roughly two to three minutes and somewhere between 35 and 55 US cents per task.

For shop owners that means the ordering process is not the bottleneck. The bottleneck is whether an agent finds and correctly understands your offer.

Why your site structure decides the outcome

The most instructive finding of the same study concerns perception. The researchers gave agents three ways of seeing a page: the accessibility tree, the structured representation that screen readers also use, a screenshot, or both together. Agents reading the accessibility tree are markedly more successful. Agents working from the screenshot alone drop off sharply and, with weaker models, fail almost entirely.

The failure modes are telling. According to the report, vision-only agents repeatedly clicked invisible or stale elements until the step budget ran out. The authors conclude that visual input without the structural guidance of the accessibility tree is not sufficient to complete the task.

From this follows something very practical that surprises many people. What makes a website usable for AI agents is essentially what makes it accessible: semantic HTML, real buttons rather than clickable divs, labelled form fields, meaningful alternative text, a coherent heading hierarchy. Anyone retrofitting for the Austrian Accessibility Act is working on agent readiness at the same time. Two obligations, one foundation.

What online shops actually need now

Standards for the purchase itself are emerging right now. The Agentic Commerce Protocol by Stripe and OpenAI (Apache 2.0 licence) describes how agents initiate a checkout on a merchant site. The merchant provides agent-facing endpoints and publishes its checkout configuration, either through a conventional API or through MCP. Important for the risk question: the merchant remains the merchant of record and can accept or decline transactions per agent or per transaction.

Google is taking a similar route with the Universal Commerce Protocol. Merchants publish a profile at the path /.well-known/ucp declaring which capabilities they offer, such as catalogue search or checkout. Google's documentation names three transports explicitly: REST APIs, MCP and A2A.

The risks that come with it

An honest account needs two caveats. First, agents can be manipulated. Zijing Shi, Meng Fang and Ling Chen (University of Technology Sydney and University of Liverpool, April 2026) injected seven typical e-commerce deception patterns into shop interfaces and set agents loose on them. Their result: current web agents are highly susceptible to several classes of deceptive interface, and prompt-based constraints are often insufficient to prevent it. Anyone deploying dark patterns will fool agents in the short term and be filtered out by the platforms in the medium term.

Second, not every machine visit is worth anything. If roughly half of all crawler requests serve training, that traffic brings no visitor with it. Agent readiness is therefore not an end in itself. It pays off where agents act on someone's behalf: research, comparison and purchase.

What I actually recommend

The order matters more than the speed. First, clean semantic HTML and accessibility. That is the foundation without which everything else stays ineffective, and in Austria it is increasingly a legal requirement anyway. Second, structured data and clear direct answers so agents can place your content correctly. Third, a machine-readable overview of the site. Only fourth, interfaces of your own such as MCP or A2A, plus the commerce protocols in retail.

Across my roughly 30 subsidised SME consultancies of the past two years, step four was almost never the first sensible step. For most businesses the leverage sits in points one and two, and there it is considerable. Anyone building a new website today should still design for the agent perspective from the outset, because retrofitting costs more than building it in.

You can see what that looks like in practice on this website: it offers the same six tools over MCP, A2A and REST. Where your own site stands is shown in seconds by the free GEO scan. The content side of this is covered in the article on GEO and AI visibility.

Frequently asked questions about the agentic web

What does agentic web mean?

The agentic web describes the phase of the internet in which AI agents operate websites on behalf of people instead of people clicking themselves. Users delegate an intent, such as a price comparison followed by a purchase, and the agent carries out the steps. In the literature it is described as the third stage after the PC web and the mobile web.

Can AI agents already shop on their own?

They handle the technical sequence. In the University of Mannheim's WebMall benchmark, agents reached a completion rate of 100 per cent for building the basket and checking out, and 75 per cent for the full purchase journey. Price comparison across several shops remains harder at around 63 per cent. For payment itself, the Agentic Commerce Protocol and the Universal Commerce Protocol are establishing dedicated standards.

What makes a website agent-ready?

Structure, above all. Agents that read a page's accessibility tree solve tasks far more reliably than agents working from a screenshot. Semantic HTML, real buttons, labelled form fields and a clean heading hierarchy therefore do double duty: they make the site accessible and machine-operable at once.

Does my online shop need to support an agent protocol now?

For most SMEs, not as the first step. The sensible order is: clean semantic structure, then structured data and direct answers, then a machine-readable overview, and only after that your own interfaces or commerce protocols. If you are rebuilding anyway, design for the agent perspective from the start.

Is traffic from AI agents worth anything at all?

Not all of it. According to Cloudflare, around half of all crawler requests serve AI training, and that traffic brings no visitor with it. Machine traffic becomes valuable where agents act on someone's behalf: research, product comparison and purchase.

Sources

  • Cloudflare, "Content Independence Day, one year on", 1 July 2026: blog.cloudflare.com
  • Yang et al., "Agentic Web: Weaving the Next Web with AI Agents", arXiv 2507.21206: arxiv.org
  • Peeters, Steiner, Schwarz, Caspary, Bizer (University of Mannheim), "WebMall: A Multi-Shop Benchmark for Evaluating Web Agents", arXiv 2508.13024: arxiv.org
  • Shi, Fang, Chen, "Benchmarking Web Agent Safety under E-commerce Deceptive Interfaces", arXiv 2606.13686, April 2026: arxiv.org
  • Agentic Commerce Protocol (Stripe and OpenAI): agenticcommerce.dev
  • Google, Universal Commerce Protocol, UCP profile at /.well-known/ucp: developers.google.com

Note: this article is a practitioner's assessment, as of August 2026. The figures cited come from the linked sources.