What is the agentic web, and what should a business do about it?

The short version

• The agentic web is the internet used by AI agents that take actions for people, rather than only by people reading pages.
• A crawler collects content, a chatbot replies and an agent acts, which is why agent mistakes can have real consequences.
• The agentic web depends on four things: readable content, callable actions, agent identity and user authorisation.
• Published adoption figures often lack a shared definition of 'AI agent', so they do not compare cleanly.
• Clean data, narrow APIs and human confirmation for irreversible actions are worth doing even if agent traffic stays small.

What is the agentic web?

The agentic web is the internet used by software that acts for people. An AI agent takes a goal, works out the steps, reads what it needs and then does something: it compares prices, fills in a form, books a slot or places an order.

Take a person who says 'book a table for four near the office on Thursday'. Today they open three sites and compare. In the agentic version, an agent checks availability, applies their preferences and confirms the booking. The restaurant's site becomes a data source and a booking endpoint, with no person looking at the page.

The term has no formal definition, and vendors stretch it to fit what they sell. The useful core is simple. The web was built on the assumption that a person is at the screen. The agentic web removes that assumption for some tasks, and asks what pages, forms and checkouts need when the visitor is a program.

Cloudflare, which says more than 20% of the web sits behind its network, put the idea in a headline: 'The Internet has a second audience'. That is a fair summary of the direction. It tells you nothing about the size.

How is an AI agent different from a crawler or a chatbot?

A crawler collects, a chatbot replies and an agent acts. That difference changes the risk. A crawler that misreads your page produces a bad summary. An agent that misreads it can produce a wrong order.

What does each type of software need from your site?

Is the agentic web the same as AI search?

No. AI search answers a question using your content. An agent completes a task using your content and, often, your systems.

The two ask different things of a site. For AI search your job is to be quotable: clear claims, named sources and facts a model can lift. For an agent your job is to be usable: stable fields, predictable errors and actions that behave the same way each time. A business that sells online needs both. A business that only publishes information mostly needs the first.

What does the agentic web run on?

It runs on four things: content a program can read, actions a program can call, a way to identify the agent and a way to prove what the user allowed. Most of the underlying parts are familiar. What is new is needing them to work together without a person checking each step.

These standards are young and they change. Check the current version of any protocol before building on it, and do not assume that a standard under discussion is one agents actually use. Most businesses directly control only the first two layers, so avoid designs that depend on one identity or authorisation scheme winning.

Which standards belong to which layer?

• •Readable content: clean HTML and structured data, plus proposals such as llms.txt.
• •Callable actions: APIs, the Model Context Protocol (MCP), which gives agents a standard way to find and call tools, and Agent2Agent (A2A), for agents that hand tasks to other agents.
• •Agent identity: signed requests such as Web Bot Auth, which aim to let a site verify who is calling.
• •User authorisation: scoped, revocable permissions, commonly built on OAuth.

How much of the web is agentic today?

Nobody has a good number yet, so be suspicious of any figure quoted without its definition. The measurements that exist each describe one part of the picture, and they do not add up to a total.

Cloudflare's post on the agentic web draws on requests across its own network. It gives no figure for agents' share of total traffic, no formal methodology and no definition of 'AI agent'. So even a network of that size has not published the number people most want.

Surveys have their own gaps. The SBE Council's September 2026 online survey covered 516 small business owners with 2 to 99 employees, with a credibility interval of ±4.4 points. Its release gives no question wording and no definition of 'AI agent', so you cannot tell what respondents took the term to mean. Menlo Ventures surveyed 495 US enterprise AI decision-makers in November 2025, but only at enterprises already using AI, and Menlo invests in AI companies.

Government data answers a different question. The US Census Bureau's Business Trends and Outlook Survey runs every two weeks, and its article covering 14 December 2025 to 3 May 2026 names firms with at least 20 employees as the biggest AI users. That is evidence about AI use at work. It does not tell you how many agents are visiting websites.

How can you tell if a website is ready for AI agents?

Give an agent a real task on your own site and watch where it fails. Can it find the price, understand the policy and complete the action without guessing? That test is cheap and it shows you the exact point of failure.

Automated checks also exist. Cloudflare has published an Agent Readiness score, and Cloudflare Radar used its isitagentready.com tool to scan the 200,000 most visited domains, after filtering out categories such as redirects and ad servers. The post dates from April 2026 and gives no scan date, and the Radar chart updates weekly, so any count from it can move.

A scanner is a useful checklist. It sees what is visible from outside. It cannot see whether the action behind a form is safe for software to trigger, or what happens when an agent retries it.

Can AI agents do useful work on the web yet?

Some can, on some tasks, but the fair test is whether a person would accept the finished result. A smooth demo proves little, because demos show the path that works. Real tasks fail on the missing field, the changed page and the request nobody anticipated.

The Remote Labor Index is a useful model of that test. Developed by the Center for AI Safety and Scale Labs, it has agents attempt paid freelance projects, and human evaluators judge each deliverable against a paid professional's. That measures finished work rather than steps completed.

Read the details of any such result. In the Center for AI Safety's post, the top-scoring model had a $150 per-project budget against the $50 default, so that score is not a like-for-like comparison. The index also covers freelance deliverables, so it says little about an agent ordering from a stranger's website.

What goes wrong when agents act on websites?

Most failures come from misplaced trust. An agent reads content from strangers, holds a user's permissions and can take actions that are hard to undo. Each of those is manageable alone. Together they need deliberate limits.

Which failures should you plan for?

• •Prompt injection: text on a page can act as an instruction, because a language model has no reliable way to separate content from commands.
• •Excess permission: an agent given a broad login can do anything that login can, including things the user never intended.
• •Irreversible actions: payments, cancellations and deletions need a confirmation step or a way to undo them.
• •Retries: an agent that times out may submit the same order twice, unless the action is built to run safely more than once.
• •Blurred analytics: software visits can distort traffic and conversion data, and decisions based on that data drift with it.

What should a business do about the agentic web now?

Fix what pays off even if agent traffic stays small, and hold off on anything that only pays off if it grows. Clean data and clear permissions help your staff and your own internal tools today. An interface built for agents nobody has seen helps nobody.

In what order should you do it?

1. Check your server logs to see which identified agents and crawlers already reach you.
2. Publish key facts in structured form: prices, availability, policies and contact details.
3. Write down what agents may do on your site, and reflect it in robots.txt, your terms and your rate limits.
4. Offer a narrow API for the one or two actions worth automating, starting with read-only ones.
5. Require human confirmation for anything that moves money or cannot be undone.
6. Log each agent action with the identity, the permission used and the result.

When is preparing for the agentic web not worth it?

It is not worth much effort if your customers buy through relationships, your site is mainly a brochure or you can see no agent traffic in your logs. In those cases, a tidy site with accurate facts is enough for now.

It is also a poor first AI project. Agents inside your own business work on systems you control, with permissions you set and outcomes you can measure. That is usually an easier place to learn than the open web, where you control neither the agent nor the pages it reads.

We have no measurement of our own to add here, and we do not know how quickly agent traffic will grow. Neither does anyone offering you a firm date. So plan around what you can see in your own logs.