The Web Beneath the Web

Last summer I took my family to the Gouffre de Padirac, a cave system in southwestern France. You walk through an ordinary field, and then the ground opens into a chasm 35 meters deep. At the bottom, there is an entire world: underground rivers, vast caverns, formations that have been growing for hundreds of thousands of years. Millions of people drive past it every year without knowing it is there.

I kept thinking about Padirac months later, but not because of the geology. I kept thinking about it because of what is happening to the internet

.

The Web Below the Web

The internet most people see, the one with websites, search results, shopping carts, and social media feeds, was built for humans. Every design decision assumes a person is on the other end: someone who scrolls, reads, clicks, and decides. That assumption held for thirty years.

It is now breaking. Underneath the surface web, a parallel internet is forming. Its users are not people. They are AI agents.

In January, Google launched Chrome Auto Browse, a feature powered by Gemini that turns the browser into an autonomous agent. It scrolls, clicks, types, and navigates without you. OpenAI shipped ChatGPT Atlas. Perplexity launched Comet. These are not search improvements or browser extensions. They are systems that use the web on your behalf, and increasingly, instead of you.

At the same time, enterprises are building infrastructure specifically for agent consumption. Shopify and Google co-developed the Universal Commerce Protocol (UCP), an open standard that lets AI agents browse product catalogs, initiate checkouts, and complete purchases through structured APIs, bypassing the human storefront entirely. Salesforce’s Agentforce now lets AI agents trigger secure API calls across enterprise systems. These are not experiments. They are production infrastructure for a web that humans never touch.

This is the other web. Like Padirac, it is vast, it is growing, and most people have no idea it exists.

The Measurement Problem

Here is why this matters to anyone running a company or sitting on a board.

The economics of the surface web are built on human attention. Advertising works because humans see ads. SEO works because humans click links. Conversion funnels work because humans browse, compare, and decide. Every metric that boards use to evaluate digital performance, traffic, click-through rates, time on site, conversion, measures human behavior.

When agents do the browsing, those metrics become incomplete. Not wrong, exactly. Incomplete. If your customer’s AI agent is comparison-shopping across 50 vendors in parallel, negotiating price and terms through Shopify’s UCP, it never visits your landing page. It never sees your brand campaign. It never enters your conversion funnel. But it might buy your product.

The uncomfortable question is: how would you know? Most companies have no visibility into agent-mediated interactions with their business. They are measuring the surface web and assuming it represents the whole picture. It is like counting the cars on the road above Padirac and concluding there is nothing underneath.

Dawn Song, a computer science professor at UC Berkeley and a leading researcher on AI security, described the shift this way in a recent IEEE Spectrum interview: in the agentic web, a user’s agent can talk directly to a website’s agent, consume thousands of data points in parallel, and negotiate in real time. The entire interaction is different. The interfaces are different. The economics are different.

And the measurement frameworks have not caught up.

The Invisible Layer of Commerce

This is not a future problem. Early agent-mediated commerce is already compressing margins and accelerating comparison shopping. When agents handle routine procurement, they optimize on structured criteria: price, availability, specifications, contract terms. They do not respond to brand storytelling or visual design. They respond to data quality and API accessibility.

Companies that compete on human attention will gradually lose to companies that compete on agent accessibility. Not because attention stops mattering, but because a growing share of commercial activity is moving to a layer where attention is irrelevant.

For boards, this creates a specific blind spot. Most AI governance discussions focus on internal use: how employees use AI tools, how customer-facing chatbots behave, how models are trained and monitored. Those are important questions. But they are all about the surface web.

The other web introduces a different category: AI systems that interact with your business externally, autonomously, and at scale. Your customers’ agents. Your suppliers’ agents. Agents from companies you have never heard of, querying your systems for structured data. The governance frameworks for that layer do not exist yet at most companies.

What Boards Should Be Asking

The EU AI Act’s requirements for high-risk AI systems become enforceable on August 2, 2026. That is less than six months away. Most companies that have started compliance work have discovered something uncomfortable: they do not have a complete inventory of every AI system that touches their business, including the external ones.

You cannot govern what you cannot see. And you cannot measure what you do not know exists. Three questions worth raising in your next board meeting:

Do we have visibility into agent-mediated interactions with our business? If the answer is “we track website traffic and conversion rates,” you are only seeing the surface. Shopify merchants that enabled UCP can now see when a purchase originated from an AI agent versus a human browser. That distinction matters. A company seeing flat website traffic but growing revenue might assume their conversion rate improved. In reality, a growing share of their sales may be happening through agent channels they are not tracking. Without that visibility, your digital performance metrics are telling an incomplete story, and your capital allocation decisions are based on incomplete data.

Is our digital infrastructure accessible to agents, or only to humans? Shopify built UCP. Google co-developed it. Commercetools is redesigning its entire commerce platform around API-first, agent-ready architecture. These companies are building for the other web because they see the traffic patterns shifting. A retailer whose product catalog exists only as a visual website is invisible to every AI agent doing comparison shopping. A B2B supplier without structured, API-accessible product and pricing data will not appear in agent-mediated procurement workflows. This is the equivalent of not having a website in 1999. The cost of inaction compounds quietly.

Who owns the strategy for the other web? At most companies today, nobody does. Digital marketing owns the website. IT owns the APIs. Product owns the customer experience. Legal owns compliance. The other web cuts across all four, and none of them are looking at it as a unified problem. Moderna solved a similar cross-functional gap last year by merging HR and IT under a single executive when AI agents became part of their workforce. The companies that figure out unified ownership of their agent-facing strategy, whether through a new role or a cross-functional mandate, will move faster than those still debating which department it belongs to.

The other web is not a future scenario. It is forming right now, underneath the internet we built for people. Like Padirac, it has been growing quietly, out of sight. Unlike Padirac, it will not wait thousands of years to reshape the surface above it.