a16z: The Internet’s focus has shifted from humans to agents—how should businesses prepare for the B2A economy?

Authors: Olivia Moore, Justine Moore, a16z

Translated by: TechFlow

TechFlow Summary: For decades, major e-commerce and content platforms spent years building formidable risk-control walls to block crawlers, but the tide has now completely reversed. AI agents already command over half of all web traffic, write more than half of online articles, and generate vast amounts of production-grade code, making them the internet's true primary users.

From mimicked clicks in web browsers to MCPs, plugin connectors, and one-click agent authorization, a distribution revolution comparable to the desktop and mobile internet eras is underway. a16z’s latest research report breaks down the four categories of beneficiaries in the “B2A Economy” and presents the first complete ecosystem map.

Before diving into today’s main article, congratulations to Counsel Health on its successful acquisition by OURA. For more details on this news, you can click here.

The primary users of the internet are no longer humans, but AI agents.

Agents currently account for the vast majority of all web traffic, write more than half of all newly published online articles, and generate most of the code deployed to production environments.

The Structural Shift in Distribution Paradigms

This is a typical characteristic of the early stages of platform-level paradigm shifts: growth in user adoption can already be clearly quantified through data, but most of the new business models it spawns remain hidden beneath the surface.

Every historical technological generational shift has simultaneously reshaped “what people can do” and “how businesses reach users.” The desktop web era gave companies an online storefront that any visitor could walk into, while the mobile internet era shoved that store directly into everyone’s pockets.

Today, AI has fundamentally expanded human capability boundaries, yet its restructuring of “commercial distribution channels” has only just begun.

As agents move beyond simply answering questions to autonomously completing complex tasks, they will increasingly make decisions on our behalf: which software to choose, which service to subscribe to, and where to spend our money.

Companies that treat agents as “prime clients” upfront will have the chance to ride the next wave of explosive growth; those who remain indifferent may face marginalization. For decades, retailers and internet platforms built high walls specifically to keep bots out; going forward, they must instead compete for the favor of “bot customers.”

Once agents truly reach mainstream consumers and achieve scaled daily usage, they will unlock the greatest untapped dividend of the entire AI era: a genuine platform-level paradigm shift.

Past great migrations (the desktop web and mobile internet) were invariably driven by two wheels: “brand-new product capabilities” and “brand-new distribution mechanisms.” Currently, AI performance in product capabilities vastly exceeds expectations, but lags significantly in distribution. This is precisely why, apart from ChatGPT, the adoption of most consumer-facing AI applications remains highly limited.

Agent Traffic Conversion Shows Early Promise

Early signs of dramatic shifts in distribution channels are already visible. Large language models like ChatGPT, Claude, and Grok are rapidly becoming highly valuable traffic sources for numerous businesses.

Data released by Shopify shows that traffic from AI referrals to merchants on its platform grew 8 times year-over-year in Q1 2026, while actual orders facilitated by AI search surged nearly 13 times.

This growth rate is accelerating further: according to monitoring by cybersecurity firm Human Security, agent browser traffic jumped 8 times in just two months from August to October 2026. Meta’s Muse accounted for a massive 60% share, while OpenAI’s products combined made up 30%.

Just as graphical web pages can target audiences far more precisely than roadside billboards, AI agents acting as discovery and referral portals are theoretically at least as powerful—if not more so—because they hold extremely granular and accurate data on user needs and preferences.

Real-world data to date confirms this: compared to non-AI traffic, AI-referral consumer traffic exhibits higher user retention, better conversion to actual revenue, and significantly lower bounce rates. Even before mass agent adoption, it is estimated that 11% of online shopping journeys in the US already involve consulting an AI before clicking through to a merchant site.

At this stage, most agent access behaviors still rely on “browser use”: either the agent takes over the user’s current browser tab, or the system spins up a virtual machine and browser instance, then clicks through webpages sequentially like a human to complete tasks.

However, we are already witnessing the nascent emergence of an “App Store for Agents”: various data connectors, plug-ins, and the MCP protocol are evolving rapidly, allowing agents to skip cumbersome simulated clicks and directly invoke core capabilities from third-party tool websites.

For example, in a practical test: we instructed ChatGPT to coordinate Krea, Descript, and Eleven Labs plugins to create and edit a video ad. It smoothly managed cross-product coordination and seamlessly prompted us to upgrade to Descript’s paid subscription during the final export step to complete the task.

As underlying large model platforms allow users to securely “lend” their account credentials or login tokens to other products for task execution (such as Sign in with ChatGPT authorization logins), the friction for agents to automatically register, place orders, and even pay on behalf of humans will continue to diminish.

In short order, agents will be able to independently search, select, and install required tools without human intervention: Claude already supports enterprise administrators connecting external plugins for the entire organization with one click.

Just yesterday, OpenAI launched Businesses in ChatGPT, a commercial entry point that allows businesses to actively claim their company pages, connect real-time product data streams, and directly monitor user feedback and brand interaction metrics within ChatGPT.

Four Core Beneficiaries in the B2A Economy

Looking at the upcoming B2A (Business-to-Agent) economic wave, the following four types of companies are best positioned to capture this era-defining dividend:

• Agents and Agent Foundation Models Themselves: When discussing consumer agents, people often think of standalone product frontends or harnesses like Instinct, Muse, and Wabi.

But we believe that as foundation models grow more action-execution driven, virtually every future software product will become natively “agent-enabled”; ChatGPT, with 1.2 billion weekly active users, holds an especially advantageous hand in this direction.

• The Next Generation of Native Industries Spawned by Agents: Amazon was born reliant on the desktop internet, Uber couldn’t exist without smartphone ubiquity, and we too anticipate agents birthing a new breed of native commercial species.

While their exact forms are hard to pin down yet, we can already faintly glimpse several high-potential themes: fully autonomous small/medium enterprises, AI-matched trading markets or network products, and various automated optimization tools.

• Established Giants Optimizing for Agents First: LLMs have long been a core traffic driver for many content and e-commerce sites, directly spawning the so-called Generative Engine Optimization (GEO) industry.

But traditional GEO often stops at the “search discoverability” stage; agents will extend the battlefield all the way to solution evaluation, form registration, and even final payment decisions, corresponding to an incremental market far larger than search alone.

• Agent-Specific Infrastructure: For agents to truly act on behalf of humans in the real world, they must be equipped with their own dedicated infrastructure toolkit. Even at this early stage when agent traffic volume isn’t yet massive, system components like payment settlement, fraud prevention/risk control, identity verification, and communication interfaces (iMessage, voice calls, email, etc.) are already showing significant load pressure.

Precedent: How Coding Agents Are Rewriting the Software Stack

We actually already hold a real-world prototype capable of previewing future trends: coding agents, which have fundamentally rewritten developer behavior patterns over the past year.

As of September 2026, over 40% of code merge requests (PRs) on GitHub have had their first draft written by agents, whereas just six months prior, that figure stood at only 17%.

Like consumer-facing agents, coding agents equally rely on coordination with various external engineering tools. Imagine a user asking an agent to develop and deploy a website; it must invoke a cascade of backend components including cloud databases, web hosting, authentication, email delivery, and monitoring/alerting systems.

Consequently, alongside spotlight performances from agent models and dev kits like Cursor, Codex, Claude Code, and Cognition, as well as vibe coding platforms such as Replit, Lovable, and Base44, supporting infrastructure providers across the entire supply chain are experiencing explosive growth:

• Supabase: disclosed in June 2026 via this link that over 60% of new databases created on its platform were automatically generated by AI tools.

• Neon: revealed back in May 2025 via this report that newly deployed databases handled by automated agents accounted for as high as 80%.

• Stripe: reported in June 2026 via this update that agents now drive nearly 40% of traffic to its developer documentation and account for 70% of API resource CLI requests.

• Vercel: announced in October 2026 via this recap that the number of application deployments automatically triggered by agents surged 28 times over the past 12 months.

Consider that software developers make up less than 1% of the global workforce. Now imagine what happens when the remaining 99% of non-technical users start fully adopting agents—the energy unleashed would be staggering.

Full Ecosystem Map of the Agent Economy

This piece serves as the opening installment of our deep-dive series on the “Agent Economy.” Merely labeling it “AI e-commerce” vastly underestimates the historical opportunity at hand: certainly, agents will leverage massive volumes of physical goods and lifestyle service transactions, but they will also fundamentally reshape traffic entry points for digital products, token consumption, and even the flow paths of commercial revenue.

While most traditional enterprises currently still block agent traffic out of concerns over fraud prevention and chargebacks, we are convinced they will quickly realize they must embrace this increasingly high-value emerging trend.

Over the coming weeks, we will roll out more deep insights and frontline data, exploring the actual usage landscape of current consumer agents, the rules governing how agents filter and distribute high-value traffic, how underlying infrastructure continues to evolve around agents, and finally, the last-mile hurdles standing between agents and fully closed-loop transactions.

This ecosystem changes almost daily. Attached below is the most complete market map our team has been able to chart based on the current industry landscape, encompassing both the agent applications themselves and the core underlying infrastructure that powers them.