The Web Agent is the layer of the Internet where AI agents, acting on behalf of humans, discover, read, and transact with websites. It exists alongside the human network and is measured separately.
For most of the history of the Internet, three categories of visitors appeared on a website: humans, search engine bots, and bots running scripts. Agents are a fourth class. An agent is sent by a human with a task, runs autonomously on behalf of the user and performs actions in several steps. Checking availability. Fill out a form. Compare prices. Complete a purchase. Agents read websites like a crawler does and act on them as a user does. This combination is new.
The Web agent is the part of 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 runs this traffic, including protocols, execution times, and measurement tools, has been made public through 2025 and accelerated to April 2026 with Cloudflare Agent Week.
I have been thinking, talking and writing about this for 18 months. On my own website, AI assistants outnumber human visitors 5 to 10 times every day, depending on what’s going on. This ratio was close to zero two years ago. Web agent is the term I explain most often. So here it is, end to end.
This article defines the term, situates it in relation to AI research and AEO/GEO, explains the machine-centric architecture framework to build it and describes the changes for publishers, developers and businesses.
Agents are a new primary visitor class
Today, three categories of visitors read websites: humans, crawlers, and agents. Humans load pages in browsers. Bots scrape pages to create 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 can query a product catalog for a user’s specifications, compare listing options, authenticate via an OAuth flow, add items to a cart, and make a payment. An agent visiting a post can pull the current article, summarize it with other sources, and return a summarized response to the user without the user ever loading the page. Both behaviors are agent web traffic. Retail behavior drives revenue. Posting behavior rarely returns referral traffic. This asymmetry is one reason why agentic network effects are unevenly distributed across sectors.
Agent traffic is the fastest growing category of web traffic in 2026. Automated traffic as a whole is growing about eight times faster than human traffic year over year (CNBC). The growth rate is the obvious part. The interesting part is the conversion behavior. On retail websites, AI-generated traffic now outperforms human traffic in revenue per visit, a year after underperforming it. Such reversals generally do not reverse.
How Agent Web Differs from AI Search and AEO/GEO
AI and AEO research are categories adjacent to the Web agent. They are often confused with this, and each answers a different question on the Internet.
AI research refers to search products powered by large language models, including ChatGPT Search Mode, Perplexity, Google AI Mode, and SearchGPT. AI research is a consumer product that retrieves and synthesizes. The Web agent is broader. It includes AI search agents visiting websites, as well as transactional agents, booking agents, search agents, and custom agents built on top of browser APIs and runtimes. AI search is a subset of agentic web activity. Other categories of agents operate outside of research.
OAS and GEO (Answer Engine Optimization and Generative Engine Optimization) are the content optimization disciplines adjacent to SEO so that AI search systems cite it accurately. AEO is a specific practice in the broader context of the web agent. THE No Hacks Guide to Response Engine Optimization And the SEO-to-AAIO primer cover the practical side.
AXO (Agent Experience Optimization) is an actively used, although contested, term. A product launching in 2026 uses the acronym for a different concept (Agentic Experience Orchestration), so the industry vocabulary is still settling in. Functionally, AXO-as-discipline describes the work of making websites readable and transferable for agents. Machine-centric architecture is the specific framework that structures what works.
Machine-centric architecture defines how to build for it
Machine-first architecture (MFA) is based on 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 general (SEO) or too narrow (schema.org). The pillars are what I test every website against. Episode 221 of the No Hacks podcast presents them in detail, and the Glossary No Hacks defines each term individually.
Identify. A website in the Web agent needs an unambiguous machine-readable identity. Who the website is, what it sells or publishes, and what authoritative source it represents. Concretely, this means canonical URLs, consistent entity naming on pages and off-website, verified presence on queries from platform agents (LinkedIn, GitHub, Wikipedia, industrial directories) and cryptographic signals where applicable. An agent that cannot confidently determine the identity of a website turns to pattern matching, and pattern matching loses to competitors with clearer identity signals.
Structure. Critical content should not rely on client-side JavaScript execution to become viewable. Today, agents primarily read the rendered DOM, but the reliability bar is different than a human browser. Structured data (Schema.org, JSON-LD), server-side rendering, and semantic HTML all fall under this pillar. The lesson of mobile-first indexing applies here: infrastructure that depends on fragile rendering is the first thing to fail when a new class of visitors arrives.
Content. Web agent content is consumed in the form of response units and not in the form of articles. An agent extracts the sentence or paragraph that answers the user’s question, often without context. The content pillar covers response-driven architecture, quotable specificity, provenance signals, and time markers (publish dates, update dates, version numbers). The working rule: every sentence in the content must survive extraction independently. An officer quoting it should not need the surrounding paragraphs to make the quoted sentence accurate. My guide to how AI agents see your website goes through this in detail.
Interaction. The agents act. They don’t just read. The interaction pillar defines how an agent accomplishes a task on a website: what actions the website exposes, how workflows recover from errors, and how an agent’s identity and permissions are verified. This pillar is growing the fastest in 2026. WebMCP allows websites to register structured tools that an agent can call directly. Universal Trade Protocol standardizes agent payment. MCP, A2A, NLWeb and AGENTS.md cover the other protocols of this layer.
What changes for publishers, developers and businesses
Publishers, developers and businesses face three different economic realities in the context of the Web agent. Here is each one.
Publishers. Research-based referral traffic to publishers has dropped about a third globally by November 2025, with local publishers seeing declines of 25-50% (Press journal). The agent layer of the web reads content from the editor and synthesizes it directly, often without returning the user to the source page. The monetization of display ads, affiliates and page views is compressed in parallel. The step forward for publishers is the diversification of income: subscriptions, licensing agreements with AI labsdirect relationships with the public and a recognition that the pageview economy is shrinking structurally, not temporarily.
Developers. A new API surface is active. navigator.modelContext shipping in Chromium 146 in February 2026, allowing websites to register tools that an agent can call directly. Cloudflare Browser Run added production support in April 2026. (For a broader inventory of browser agents, automation frameworks, and enterprise APIs, see The agent browser landscape in 2026.) Model Context Protocol servers, OAuth flows for agents, and agent identity verification layers are active infrastructures, not propositions. The way forward for developers is to learn new primitives early, before the reliability bar rises and upgrading becomes expensive. Cost surfaces to track: inference cost per agent task (screenshot-scan-click loops burn tokens), authentication flows, and error recovery for multi-step actions.
Businesses with transactional websites. Retailers saw AI traffic increase 393% year-over-year in Q1 2026, while converting 42% better than non-AI traffic. Lead generation and SaaS signup flows come next. The next step is to audit the readability of the agents with a tool such as isitagentready.com (see the No article on hacks), correct the signals sent today against real agent execution environments and treat the agent funnel as a second funnel alongside the human funnel. The broader protocol surface for agent buying flows is covered in my agent trade guide.
The short version
Agent web is the part of the Internet in which AI agents act on websites on behalf of humans. It is real enough to show up in conversion data, and its infrastructure is shipped faster that most websites accommodate it. Machine-centric architecture is the framework for building it, with four pillars: identity, structure, content, interaction. The long change is already underway. The question is which side of the fork a given website is on.
I shifted the focus of No Hacks last year because the gap between what ships and what most builders know is wider than it has ever been since mobile. The Web agent is the biggest part of this gap. If this article arrives, send it to someone who will discuss it with you.
More resources:
This article was originally published on No hacks.
Featured image: Collagery/Shutterstock
