How AI Agents Will Change the Internet

Almost everything about the modern web assumes a human on the other end of the connection. Layout assumes eyes. Advertising assumes attention. Free tiers assume that some fraction of visitors will eventually pay. Rate limits assume nobody clicks faster than a person can click. Sessions assume someone stays logged in.

Every one of those assumptions fails when the visitor is software with a goal, no eyes, no patience and no interest in the surrounding page. That is not a distant scenario - a meaningful share of traffic to documentation, pricing pages and structured data already comes from agents, and the share is going one direction.

1. The page stops being the product

For thirty years the unit of the web has been the page: a bundle of content, navigation, branding and monetisation delivered together. An agent wants one field out of that bundle and discards the rest. It does not see the header. It does not read the related articles. It will not be retargeted.

What replaces the page is the answer with a citation: a structured response an agent can consume, with a reference back to where it came from. Sites that publish only rendered HTML will be read anyway, badly, by browser agents. Sites that publish a described interface will be read accurately and cheaply - and will be in a position to charge for it.

2. Advertising loses its anchor

Display advertising is a bet that attention can be measured and sold. Agents provide no attention. They will not be persuaded by a banner, will not remember a brand, and will actively strip anything that is not the requested information.

This does not kill advertising, but it splits the web's economics in two. Human-facing surfaces keep the attention model. Machine-facing surfaces need a different one, and the only candidate that scales is metered access: a small fee per structured request, paid by the calling agent. That is precisely why machine payments are a load-bearing part of this story rather than a crypto detour - the business model of the agentic web depends on being able to charge a tenth of a cent without a checkout.

3. Search becomes a wholesale market

Searching used to be a retail activity: a person types, scans ten results, picks one. An agent's version is wholesale - issue forty queries, read everything, discard 95 % of it and synthesise the rest. The economics of being result number seven change completely when nobody is scrolling.

Two things follow. Ranking matters less than being parseable and being correct, because an agent that can extract a clean answer from position seven will use it. And traffic stops correlating with value: a page can inform a thousand decisions while recording a hundred visits, because one agent read it once and cached it.

4. Identity becomes mandatory, and pseudonymity gets harder

When the visitor is software, "who is this" is not a marketing question but an operational one. Is this agent acting for a customer with an account? Is it allowed to hold inventory while it decides? Is it the same agent that hammered the endpoint yesterday? Sites will need to distinguish agents from each other and to establish whose authority each one carries - which is the practical driver behind agent identity work, well ahead of any philosophical interest in self-sovereign identity.

5. Abuse changes shape

Scraping used to be a nuisance measured in bandwidth. Agentic traffic is different: it is purposeful, it retries intelligently, it works around obstacles, and it is indistinguishable from a legitimate customer's assistant. Blocking it wholesale means blocking customers. Allowing it wholesale means being read for free by every competitor's model.

The equilibrium most sites will land on is neither: identify the agent, price the access, and enforce with a payment requirement rather than a CAPTCHA. A CAPTCHA asks "are you human", which is rapidly becoming the wrong question. A payment header asks "are you authorised and willing to pay", which is answerable by both humans and machines.

6. Interfaces become a fallback

Not immediately, and not everywhere. But the direction is clear: for transactional tasks - booking, ordering, comparing, rescheduling - the interface is overhead that exists because there was no other way to express intent. Once intent can be stated in a sentence and executed by software, the form is a legacy compatibility layer.

Interfaces survive where the interaction is the value: creative tools, exploration, anything where a person wants to see options rather than receive an outcome. They erode where the interaction is friction. Most e-commerce is friction.

What this does not mean

• The human web does not disappear. It becomes one audience among two. Publishing for both is the normal case, not a transitional phase.
• Agents do not become the majority of visitors everywhere. They concentrate in structured, transactional and reference content first, and may never dominate the rest.
• Nothing here requires better models. Every change described is a consequence of tool use and reliability, not of intelligence. The models are already sufficient; the plumbing is not.

The realistic timeline

The technical parts arrive faster than the institutional ones, as always. Machine-readable interfaces and agent traffic are here now. Metered access and agent identity are in the standards-and-pilots phase and will be unevenly adopted for years. Liability for autonomous action is unresolved, and it is the constraint that will actually set the pace - see autonomous commerce for where that bites hardest.

The useful posture is neither breathless nor dismissive: assume agents will read everything you publish, assume they will not see your design, and decide deliberately whether you want to be readable, chargeable, or closed. Those are the three options, and most sites are currently in the first by accident.