The web agent is an Internet layer where AI agents, working for humans, discover, learn, and use websites. It exists alongside the human web and is measured separately.
For most of the Internet’s history, three categories of visitors have appeared on a website: humans, search engine crawlers, and text-processing robots. Agents are fourth class. An agent is sent by a person with a task, works automatically on behalf of the user, and performs multi-step actions. Checks availability. Filling out the form. Price comparison. To complete the purchase. Agents read websites the way a browser does and act on them the way a user does. That combination is new.
The web agent is part of the web traffic, infrastructure, and protocols dedicated to this class. In Q1 2026, AI traffic to US retailers grew 393% year over year and, for the first time, converted 42% better than non-AI traffic, a year after converting 38% worse (Adobe via TechCrunch). The infrastructure that enables this traffic, including protocols, runtimes, and measurement tools, is publicly deployed through 2025 and accelerated in April 2026 during Cloudflare Agents Week.
I have been thinking, talking, and writing about this for 18 months. On my website, AI assistants outperform human visitors 5 to 10 times on any given day, depending on what’s going on. That rate was nearly zero two years ago. Web of agent is one term that I find myself explaining a lot. Here it is, end to end.
This article defines the term, contrasts it with AI and AEO/GEO search, describes the initial framework for its construction, and outlines what changes for publishers, developers, and businesses.
Agents Are a New Class of Principal Visitors
Three classes of visitors read websites today: humans, browsers, and agents. People load pages in browsers. Search engines download pages to build search indexes. Agents do both and more. They load pages to extract information and perform actions on behalf of the user.
An agent visiting a retail website may query the product catalog with user specifications, compare options across lists, authenticate with an OAuth flow, add items to the cart, and complete checkout. An agent visiting a publication may extract the current article, summarize it alongside other sources, and return a combined response to the user without the user having to load the page. Both of these behaviors are agent web traffic. The sales method generates income. The publication handler rarely sends forward forwarding traffic. This asymmetry is one of the reasons that the results of the web of agents are distributed unevenly across sectors.
Agent traffic is the fastest growing segment of web traffic in 2026. Overall automated traffic is growing nearly eight times faster than human traffic each year (CNBC). The growth rate is the obvious part. The interesting part is the conversion behavior. For retail websites, AI-driven traffic now surpasses human traffic in revenue per visit, a year after underperforming. Such inversions are rarely reversed.
How Agentic Web Differs From AI and AEO/GEO Search
AI search and AEO are the dependent categories of the agent web. They are often confused about it, and each one addresses a different question about the Internet.
AI search refers to search products powered by major language models, including ChatGPT search mode, Confusion, Google AI Mode, and SearchGPT. AI Search is a consumer product that discovers and integrates. The agent web is wide. It includes AI search agents that visit websites, and includes event agents, booking agents, research agents, and custom agents built on top of APIs and browser runtimes. AI search is a subset of agent web work. Some categories of agents work without a search.
AEO and GEO (Answer Engine Optimization and Generative Engine Optimization) are fields close to SEO to optimize content so that AI search systems can interpret it accurately. AEO is a specific practice within the broader context of the web of agents. The No Hacks Guide to Responsive Engine Optimization and SEO-to-AAIO primer covers the practical side.
AXO (Agent Experience Optimization) is a term that is used continuously, although it is debated. The product launched in 2026 uses a different concept acronym (Agentic Experience Orchestration), so the terminology of the field is still being developed. Functionally, AXO-as-discipline defines the task of making websites readable and responsive to agents. The first architecture of a machine is a specific functional framework.
The Architecture of the First Machine Explains How to Build It
Machine-first Architecture (MFA) has four pillars: Identity, Structure, Content, and Interaction. I introduced MFA in 2026 because the existing frameworks for making websites work for AI agents were either too primitive (SEO) or too small (schema.org). Pillars are what I test every website against. Episode 221 of the No Hacks podcast introduces them in detail, and the No Hacks glossary explains each word.
Ownership. An agent’s web site needs an unmistakable machine-readable identity. Who the website is, what it sells or publishes, and what authoritative source it represents. Specifically, this means canonical URLs, naming the business on all pages and outside the website, a guaranteed presence in the forum agents query (LinkedIn, GitHub, Wikipedia, industry directory), and cryptographic signals where appropriate. An agent that cannot confidently resolve the identity of a website falls back on pattern matching, and pattern matching loses to competitors with clear identity signals.
The structure. Important content should not rely on client-side JavaScript for rendering. Agents today mostly read the rendered DOM, but the reliability bar is different for the human browser. Structured data (Schema.org, JSON-LD), server-side rendering, and semantic HTML all fall under this pillar. The lesson from the mobile-first index applies here: an infrastructure dependent on a fragile supply is the first thing to fail when a new class of visitor arrives.
Content. Agent web content is used as response units, not as articles. The agent outputs a sentence or paragraph that answers the user’s question, usually without any surrounding context. The content pillar includes the original architecture, citable specifications, original signals, and temporal markers (dates of publication, dates of revision, version numbers). Rule of thumb: any sentence in the content must live independently of the output. A quoting agent should not need surrounding paragraphs to make the quoted sentence accurate. My guide to AI agents recognizing your website goes through this in detail.
Working together. Agents are active. They don’t just learn. The interaction pillar describes how an agent completes work on a website: what actions the website exposes, how the workflow recovers from errors, and how the agent’s identity and permissions are verified. This pillar is developing very rapidly in 2026. WebMCP allows websites to register structured tools that the agent can call directly. The Universal Commerce Protocol measures agent output. MCP, A2A, NLWeb, and AGENTS.md include other protocols in this layer.
What Changes for Publishers, Developers, and Businesses
Publishers, developers, and businesses face three different economic realities under the agent web. Here is each one.
Publishers. Search-driven referral traffic to publishers fell by almost a third worldwide in the year to November 2025, with local publishers seeing a 25-50% drop (Press Gazette). The web agent layer reads the publisher’s content and compiles it directly, usually without returning the user to the source page. Display ad, interest, and monetization compression per page view. Moving forward for publishers is diversification of revenue: subscriptions, licensing deals with AI labs, direct relationships with audiences, and an acknowledgment that the pageview economy is declining structurally, not temporarily.
Engineers. The new API area is active. navigator.modelContext shipped with Chromium 146 in February 2026, allowing websites to register tools that the agent can call directly. Cloudflare Browser Run added production support in April 2026. (For a comprehensive list of agent browsers, default frameworks, and enterprise APIs, see Agentic Browser Landscape in 2026.) Model Context Protocol servers, OAuth flows for agents, and agent identity infrastructures, not live authentication layers. The best way for developers to learn new things early, before the reliability bar rises and retrofits become expensive. Track cost areas: cost considerations for each agent’s activity (screenshot-analyze-click token-burning loops), validation flow, and error detection for multi-step actions.
Businesses have commercial websites. Marketers saw AI traffic grow 393% year over year in Q1 2026 while converting 42% better than non-AI traffic. Lead generation and SaaS subscription flows follow. To go further check the readability of the agent with a tool like isitagentready.com (see the text No Hacks), adjust the signals sent against the real agent’s runtime today, and treat the agent conversion funnel as a second funnel next to the human one. A comprehensive protocol for the buying agent flow is covered in my trading guide.
Short Version
A web agent is a part of the internet where AI agents create websites on behalf of people. It is accurate enough to show conversion data, and its infrastructure is deployed faster than most websites adapt to. Machine-first architecture is a framework to build on, with four pillars: Identity, Structure, Content, Interaction. A long shift is already underway. The question is which side of the bifurcation a given website is on.
I took the focus off No Hacks last year because the gap between what’s being shipped and what most developers know is wider than it’s been at any time since mobile. The web agent is a big part of that gap. When this topic comes up, send it to one person you can argue with.
Additional resources:
This post was originally published on No Hacks.
Featured Image: Collage/Shutterstock
