The old business model of the internet is dead.

The man who runs a fifth of the web says agents will outnumber people online by next year, agents do not look at ads, and the web will need payments at a scale no network has ever run. He is right. Here is what replaces ads, and what we have already built of it.

Joe Charlesworth, founder, AI3 · 23 September 2026 · 8 min read

What the man who carries a fifth of the web said

Cloudflare sits in front of roughly one in five websites. When its chief executive talks about traffic he is not guessing from a sample; he is reading a meter. In May, Matthew Prince sat down with Bankless and read it out.

The web has run on one deal for thirty years. A site publishes something. A search engine reads it and sends people. The people see ads, or a few of them pay a subscription, and that pays for the next thing the site publishes. Content for traffic, traffic for money. Every newsroom, every recipe blog, every forum you have ever found an answer on is funded by that loop.

The loop is breaking, and the break is not that crawlers read pages. Crawlers always read pages; that was the deal. The break is that the new crawlers read the page and answer the question themselves, and the person never arrives. An answer engine sends a fraction of the visitors a search engine did, for far more reading. Prince's forecast is that traffic from agents overtakes traffic from people by 2027. Agents do not see ads. Agents do not subscribe on a whim. The audience the whole model was built to sell is walking out of the room.

His answer is not to stop the machines. It is to charge them. On 1 July 2025 Cloudflare declared what it called Content Independence Day: its customers' sites block AI crawlers by default, and a site can name a price for access instead. The mechanism is the HTTP status code that sat unused for three decades, 402 Payment Required, dusted off as pay per crawl, and behind it the open x402 standard for attaching a payment to a request, settled in stablecoins. Cloudflare has since said it intends to issue a stablecoin of its own.

Then he put numbers on it, and the numbers are the point. Cloudflare handles on the order of 500 million requests a second. Prince thinks somewhere between one and ten percent of those could reasonably carry a payment. That is between 5 and 50 million paid requests a second, every second.

Chris Burniske's reading

Chris Burniske, who has spent a decade thinking about what crypto networks are actually for, drew the line Prince did not quite draw himself. Crypto has been promising micropayments for as long as it has existed, and they never came, because they were for people. A person will not decide, forty times an hour, whether a page is worth two cents. The deciding costs more than the two cents. That is why every micropayment scheme since the nineties died the same death.

An agent has no such problem. An agent has a budget and a rule, and it will pay two cents for a page ten thousand times a day without feeling anything. Micropayments were not a bad idea. They were an idea forty years early, waiting for a payer that does not get tired.

The second thing he drew out is scale. Five to fifty million transactions a second is beyond anything any blockchain has run in production, and far beyond the card networks, which peak in the tens of thousands. Whatever settles this traffic has to be built for it. And the third thing is trust: if Cloudflare builds this for its customers first, which it will, the network under it should still be one anyone can validate and audit. A new business model for the web should be an open network, because that is the only kind the web has ever trusted for long. His closing line was Web 3.0 is dead, long live Web 3.0, and he meant it kindly.

Why ads cannot survive agents

An advertisement is a bet that a person's attention can be resold. That is the entire product. Take the person out and there is nothing to sell: an agent reading your page is either your customer or it is nothing, and it is only your customer if there is a way for it to pay you.

Subscriptions fail more slowly and for a different reason. A subscription needs a person to decide, in advance and in the aggregate, that a source is worth a monthly sum. An agent does not decide in the aggregate. It decides per request, on the merits of that request, against a budget it was given. A model that needs a decision up front cannot price a payer that only decides on the spot.

So the unit of the web changes. The pageview gives way to the request. The impression gives way to the transaction. Attention stops being the thing that pays for the internet, and settlement starts. Every request that matters will need four things it does not have today: a price, a payer with authority to spend, a way to settle, and a way to say whether what came back was what was paid for.

A rail is not an economy

What Cloudflare is building is a rail: a way for a request to carry money. It is necessary, it is well placed to do it, and we hope it does it in the open. But a business that wants to be paid by agents needs more than a rail, in the same way a shop needs more than a card terminal.

• Identity and authority. Who is paying, on whose behalf, and up to what limit. An agent that can spend is an agent that can be robbed; the budget has to be somebody's, and the somebody has to be able to stop it.
• Books. Ten thousand payments of a cent are an accounting problem before they are a payment problem. Somebody has to post them, reconcile them against the bank, attribute them to the agent and the session that spent them, and pay tax on the total. A business that cannot read its own agent traffic in its books does not have revenue; it has noise.
• A definition of done. Paying for a page is simple: the page arrived. Paying for work is not. When an agent pays another agent to book an appointment, check an address or answer a call, both sides need to have agreed in advance what done means, and a check that says whether it happened.
• Recourse. When done is disputed, somebody independent decides. It cannot be the buyer, it cannot be the seller, and it cannot be the operator of the network, because the operator has a stake.
• A price. What should a request cost? Not what the market will bear on day one, when nobody knows anything. Cost, plus a margin, both shown, so that the first thing an economy learns is what things actually cost to produce.

Burniske asks for open validation, so anyone can audit the network. We would add open books. An economy in which agents pay agents millions of times a second is one no person will ever read by hand. Standing in it, the answer to "is this a good business to deal with", has to be computed from the record, and the record has to be one you can trust.

What we have built of it

AI3 is the layer above the rail. We did not start with a payment network; we started with the books, because nothing else in an agent economy is true without them. Then we let agents trade on top and watched what broke.

The directory is the other half. When agents are the customers, a business has to be findable, readable and reachable by an agent, and most are not. So any business can check what an agent sees when it looks at them, claim its page, and open the doors an agent can walk through: a phone line, a chat, an enquiry thread, a machine-readable one. The first agent most of them hire is a receptionist, priced as a seat plus what it used to run, at cost plus a margin, both on the bill.

When pay-per-crawl lands on a business's website, the money has to land somewhere. That somewhere is a set of books, with the request that earned it attached, and a way to say whether it was earned. That is what we are building, one organisation at a time, in the order that keeps it honest.

What to do about it

• If you run a business: check your score. It shows both axes, whether an agent can reach you and whether it can find you, and every fix.
• If you publish: put a price where a machine can read it. Cloudflare's pay-per-crawl is one way; a listed price on your own page is another. A crawler that cannot find a price cannot pay one.
• If you build agents: the API and the MCP server are open. Give your agent a budget somebody owns and a place where its spending is posted.
• If you want the whole argument: the paper is the long version, with the stages and the principles.

The old business model of the internet is dead. It did not die because the machines read the web. It died because they read it and nobody had to pay. The economy that replaces it starts with the books.

More from the blog · RSS · Questions and corrections: contact. A correction that changes a fact is noted above with its date.