What is the agentic web?
I’ve built on the web since 2005. Today I build and test agents and AI workflows every day, across WP Creative’s WPO platform, internal tools and client work.
The short answer
The useful question is what the agent can actually do, and who checks the result.
The agentic web is the web reshaped around a second kind of visitor: software that acts for a person. Agents already read pages through screenshots, the page structure and the accessibility tree, and some can fill forms and call tools a site offers. Adoption is real but early and uneven: in September 2026, AI assistants sent 0.43% of website visits, against Google’s 23.96% (Ahrefs). For a marketing team, the practical job today is making the website readable, trustworthy and usable by agents without breaking it for people.
How does the agentic web work?
Cloudflare describes its Agentic Internet through four qualities: readable, discoverable, callable and payable. They cover finding information, using an offered action and, where supported and authorised, paying. Not every agent or website supports every step.
01
Readable
The agent can get the content. Google’s guide to AI features says browser agents work by “analyzing visual renderings (like screenshots), inspecting the DOM structure, and interpreting the accessibility tree”.
02
Discoverable
The agent can find the right page and trust it. Today that mostly runs through search: Google’s AI features use Google’s index, and Copilot runs on Bing.
03
Callable
The agent can do something, not just read. Forms with proper labels are the basic version, and for most sites the only one needed. Proposals such as WebMCP would let a page hand an agent named tools, like “book a call”, but adoption is thin and agents keep getting better at using ordinary pages.
04
Payable
Cloudflare means payments. For a service business, I adapt the practical question to “can the buyer book?” A booking still needs the right authority and confirmation.
Why does the agentic web matter for marketing teams?
The agentic web matters because some buyers research a business inside an AI answer before visiting its website. That creates two separate jobs: make the facts easy to find and verify, then make the next action work. Fewer, better-informed visits arrive. The site that is easy for an agent to read, trust and act on is more likely to be named, and more likely to convert the visits it still gets.
That last number is why I don’t tell teams to abandon search. Google’s AI features use its Search index and ranking systems; a ChatGPT or Claude answer may use different sources, or no live search at all. A Google position is not an AI citation position. The value of the visit that follows is what I call post-click value.
01
Research moved first
45% of B2B buyers used generative AI in a purchase, mostly to research vendors (Gartner, 645 buyers, May 2026).
02
Clicks moved second
68% of US Google searches ended without a click from January to April 2026 (SparkToro, using Similarweb panel data).
03
Traffic from AI is still small
AI assistants sent 0.43% of visits across 115,254 sites in September 2026 (Ahrefs).
What does the agentic web look like in practice?
01
An agent reading
A marketing manager asks Claude or ChatGPT to shortlist three Sydney firms that rebuild B2B websites. A search-enabled assistant may run several searches and use the pages it retrieves. Clear services, locations, evidence and limitations give it something useful to work with. They do not guarantee a citation.
02
An agent acting
In supported ChatGPT desktop configurations, an agent can call JavaScript tools a page registers through WebMCP, which OpenAI calls site tools. Availability depends on the app, model, workspace policy and rollout. A registered booking tool can offer a direct path, while normal permissions and confirmation still apply.
03
Me, building it
I work on agents that use a defined set of tools rather than an all-powerful login. My team uses Claude Code inside a written project structure, with access scoped to the job, changes reviewed and a recovery path. The tools and the account need separate checks: a WordPress Author can still edit or delete their own published posts even if a particular MCP tool refuses those operations. I want the authority boundary written down and tested, not assumed from the word “agent”. That is the same approach I bring to AI website security.
Agentic web vs the web we have now
The web most sites are built for
- Who visits A person in a browser
- How they arrive A search result or an ad
- How they read Visually, top to bottom
- What wins attention Design and headlines
- How they act Clicks and forms
- What you measure Sessions and rankings
The agentic web
- Who visits A person, or an agent acting for them
- How they arrive An AI answer, a link inside it, or an agent’s own fetch
- How they read Screenshots, page structure and the accessibility tree
- What wins attention Clear, self-contained answers and facts that check out
- How they act Labelled forms, and tools a site offers (for example WebMCP)
- What you measure Being named in answers, and the value of each visit
Caption: How the agentic web changes who visits a website and how. Sources: Google AI optimisation guide (updated 2026-07-10); Cloudflare (6 August 2026).
Common misconceptions about the agentic web
“People will stop browsing by 2027.” The evidence here does not establish that. The Ahrefs referral sample measures visits sent by assistants, not all agent activity or future adoption. Plan for people and agents together.
“You need a new kind of website.” Most of what makes a site usable by agents is what made it usable by people: real buttons and links, labelled forms, content in the HTML, fast pages. Start there.
“An llms.txt file makes you agent-ready.” A link index cannot prove that an agent can complete a task. Google Search does not use llms.txt for visibility; developer tools may use it for navigation. Keep those jobs separate. More on whether llms.txt works.
“The agentic web is about payments.” Agent payments get the headlines, but for most service businesses the first job is being found, read correctly and booked. Payments are a watch-and-write topic for now.
What should you do next?
- Ask the assistants. Run the questions your buyers ask through ChatGPT, Claude, Gemini, Perplexity and Google. Note whether you’re named and whether the facts are right. Repeat it; single runs vary.
- Test both reading paths. Compare a raw fetch with the rendered browser page. Check labelled controls, then run a named task with a named agent; JavaScript being present is not itself a failure.
- Fix access. Make sure robots rules and your CDN aren’t silently blocking AI search and user agents.
- Say it plainly. Put who you serve, what you do and how to start in text, on pages that already rank.
- Read the definition of an agent-ready website and use it as your checklist.
If you want a second opinion, book a free call. You’ll talk with me or my team, and my team at WP Creative builds the technical foundation.
Questions about the agentic web
Who coined the term “agentic web”?
The term has several public uses. Microsoft discussed an “open agentic web” at Build 2025. Pieter Abbeel and Dawn Song are among the 18 authors of Agentic Web: Weaving the Next Web with AI Agents, submitted in July 2025. Cloudflare uses “Agentic Internet”. That record is useful context; it does not establish a single originator.
Is the agentic web the same as agentic commerce?
No. Agentic commerce is one part of it: agents comparing and buying products. The agentic web is broader. It covers agents researching, comparing, booking and building, on any kind of website. For most service businesses, being found and booked matters before agent checkout does.
Does the agentic web replace SEO?
No. Google’s AI features are built on its Search systems, so SEO remains relevant there. Other assistants have their own retrieval and answer systems. Keep technical SEO, useful content and real proof in place, then measure mentions and citations separately from rankings.
How do I know if AI agents are visiting my website?
Check server or CDN logs for claimed user agents such as ChatGPT-User, Claude-User and Perplexity-User, then verify identity using each provider’s documented method where available. A user-agent name can be copied. Keep fetches, answer citations, referral sessions and completed enquiries separate: one does not prove the next. Search Console and Bing Webmaster Tools provide additional, platform-specific visibility data.
Is my website ready for AI agents?
Start with three checks: can the intended reader get the content, can a browser agent use the main form, and do permissions stop it at the right point? Keep anti-abuse controls; design a clear handoff when a person must complete a challenge or approve a commitment. The agent-ready website checklist goes further.
