01 Trigger Event
In “Welcome to the B2A Economy,” a16z offered a clear assessment: AI agents already account for most internet traffic, have written more than half of all new articles, and generate most of the code entering production. Shopify, meanwhile, disclosed that AI-driven storefront traffic grew eightfold year over year in Q1 2026, while orders from AI search increased nearly 13-fold.
This is not simply “AI bringing more visits.”
The real shift is that the internet’s primary users are changing from people to agents. A connective layer between browsers, LLMs, and payment infrastructure is forming a new distribution system.
One caveat is necessary: the article does not provide complete data sources for its claims about “most traffic” or “most code,” so it is impossible to independently verify what websites and code repositories those figures cover. Shopify’s and Human Security’s numbers are more specific, but Human Security’s data likewise represents only the bot traffic it observes, not the entire web.
02 What This Really Means
Every platform of the past was designed around human users.
Advertising auctions optimized for clicks, brands purchased SEO rankings, and retailers built storefronts. SEO and SEM were mature distribution systems in the human-computer interaction era: they predicted what a person would see, click, and buy, then competed for that person’s attention.
AI agents are changing the transaction chain.
They do not merely “find products” for you. They compare products, read reviews, decide which service to subscribe to, and may even make payment on your behalf. Agents control the task, not merely the keyword. This gives them an opportunity to bypass traditional advertising inventory and become the product-selection layer themselves.
The issue is not whether websites will be “replaced by AI.” It is who controls the default decision-making entry point for agents.
This is the core of the B2A, or business-to-agent, economy.
The marketplace of the past matched consumers with merchants. The marketplace of the next layer may match merchants with agents. The distribution merchants compete for will no longer be determined only by whether a page ranks first on Google. It will depend on whether their APIs, product structures, review data, and payment interfaces can enter the recommendation sets of agents such as Claude, ChatGPT, and Grok.
If this is true, agent traffic will not be a low-quality copy of human traffic. It carries richer context: user preferences, budget, shopping history, and current task. It is therefore unsurprising that a16z finds AI referrals to have stronger engagement and conversion.
But this also creates an uncomfortable implication: moats built around “traffic scarcity” will weaken. Merchants used to buy keyword auctions; in the agent era, they may need to buy the ability to be understood and invoked by agents. Page copy will matter less than structured catalogs, real-time inventory, shipping policies, return policies, and machine-readable reputation.
What may ultimately be priced is not the agent itself, but eligibility to enter the agent decision set.
03 Historical Analogies / Structural Comparisons
The iPhone was a capability inflection point in 2007, but the App Store was the distribution inflection point.
The iPhone made mobile internet possible. Yet only when the App Store integrated installation, payment, updates, and discovery did developers gain a distribution mechanism that had not existed before. Apple’s primary value capture did not come from hardware margins; it came from control of that distribution entry point.
AWS presented a similar structure in 2014. EC2 was not the first cloud-computing capability, but it extracted computing from lengthy procurement cycles and converted it into an on-demand API. Developers could experiment with new product forms without first purchasing a server. AWS’s moat came from distribution and switching costs: once a service depended on interfaces such as IAM, S3, and Lambda, migration costs accumulated with every dependency.
Today’s AI agent stack closely resembles the intersection of the iPhone and AWS.
Models provide new capabilities; browser use, agent protocols, payments, and tool calling provide the distribution mechanisms. This time, however, the strongest distributors may not be websites. They may be model vendors.
Google controls browsers and search. OpenAI, Anthropic, and Google simultaneously control foundation models and consumer-agent entry points. If agents decide “which product is worth buying,” model vendors can combine model capability and distribution into a single pricing proposition.
There is an even deeper historical analogy.
In the web era, commercial websites competed for human attention. In the agent era, they will compete for agent permission. The former requires SEO, advertising, and brand; the latter requires API compatibility, structured data, and trust scores.
This may also explain why a16z says AI has overdelivered on “capability” while still trailing on “distribution.” ChatGPT demonstrated that people are willing to interact with models. It did not yet demonstrate that models can replace the full web experience of browsing, checkout, and after-sales service.
In other words, limits on consumer AI adoption may not mean the models are insufficiently intelligent. Agents may be missing a reliable, open final mile with a closed payment loop.
04 What This Means for AI Builders
The question to reconsider this week is not “should we build an AI product?” It is whether the product serves humans or agents.
First, evaluate users without looking only at DAU. Separate agent requests, agent API calls, structured action completions, and paid conversions. A human user may submit one prompt while an agent performs a dozen tool calls. Pricing only by session will underestimate the work actually performed.
Second, product surfaces should support both browser UI and API. Browser use remains the dominant path today, but agents that need to complete tasks continuously will eventually prefer stable APIs over clicking through interfaces anew. For AI builders, tool-calling reliability, structured output, and error recovery are becoming core product capabilities rather than demo features.
Third, redesign SEO and distribution.
If ChatGPT, Claude, or Grok begins selecting products on users’ behalf, “will people open our app?” will no longer be the only metric. The practical questions will be: Can an agent discover the product, understand its value, and invoke it in the correct context?
Catalogs, pricing, availability, and policies must all be machine-readable. Information scattered across marketing pages will be effectively invisible to agents.
Fourth, build a dedicated data flywheel for agents.
Do not assume every referral comes from a human. Session definitions, user identities, and conversion paths differ between human and agent traffic. Combining them in one analytics dashboard may cause you to underestimate agents’ contribution or misjudge the quality of their traffic.
Fifth, evaluate token and inference economics using the full task cost rather than the cost of a single model call.
An agent task may include planning, browsing, retrieval, comparison, payment, and recovery. The model API’s price per token remains important, but the true unit economics are the cost per completed task and the tail costs created by failed retries.
I may be underestimating this point. The bottleneck in many agent workflows may not be inference, but UI friction, login, payment authorization, and identity. Betting all growth on cheaper tokens could mean optimizing the wrong variable.
05 Counterarguments / Risks
I may be overestimating the impact of agent traffic on commercial distribution.
“Most web traffic” sounds like a platform migration, but bot traffic does not equal purchasing power. It may include crawlers, SEO scrapers, automated testing, and malicious bots. Human Security’s observed eightfold growth proves that traffic composition changed, not that agents have developed stable purchasing habits.
The second risk is that the conversion narrative may suffer from selection bias.
AI-referred traffic may be more likely to complete purchases because users have already filtered products through entry points such as ChatGPT and arrive with clear intent. This does not show that agents can introduce users to brands they otherwise would not have bought. In other words, agents may reinforce existing intent rather than create demand.
The third risk is that a B2A economy will not emerge naturally as an open marketplace. Instead, it may be captured in a closed system by a small number of model platforms.
If agent recommendations are controlled primarily by proprietary model vendors, merchants will not gain a competitive auction environment. They will simply have to adapt to one proprietary interface after another. The supposed B2A economy may not be an equal exchange between merchants and agents; it may be a new toll imposed by model vendors through control of distribution.
The fourth risk is that payments and trust will slow adoption.
Agents can recommend products, but platforms still handle purchases, refunds, identity verification, and fraud. When responsibility is unclear, agents are unlikely to be permitted to complete high-value transactions. Browser automation is particularly vulnerable to friction from CAPTCHAs, login, and consent flows.
The fifth risk is that open models may make agent distribution fragment faster.
If Claude, ChatGPT, Grok, and numerous open-source agents coexist, merchants may need to maintain multiple tool schemas, prompt policies, and routing systems. The supposed “agent selection layer” may not converge on a stable entry point. It may instead become another app-store war.
I still believe distribution is shifting, but it will not easily produce a unified, flourishing B2A marketplace.
A more likely outcome is that OpenAI, Google, Anthropic, and browser vendors each control a segment of the agent stack, while merchants compete for invocability across multiple entry points.
The real validation will not be continued growth in agent traffic. It will come down to three questions: Can agents complete payments independently? Will merchants price based on agent-attributed sales? Can businesses remain profitable while switching across models? If all three become true, B2A will be more than a traffic shift. It will be a new business.