Ben Thompson's recent article, 'The Agentic Web and Original Sin', talks about the proliferation of agents on the internet and how this will dramatically skew incentives for content creators.

The article theorizes that the old, ad-based revenue models for creators will break. Without a replacement, content creation becomes siloed in walled gardens (Facebook, TikTok etc). Creative energy on the open web dries up, and what’s left is a bleak, increasingly vacant internet.

This is already beginning. If you find yourself using ChatGPT instead of Googling, you’ve probably felt the shift. As AI agents continue to change how users access information, fewer people will visit websites directly, and ads, which agents don’t see or click, stop being effective.

And here's the shift we're entering:

I spent some time diving into how the open web actually sells ad space, and the complexity is wild. DSPs, SSPs, and Ad Exchanges, these invisible layers react to nearly every click you make. They profile you, match your behavior with advertiser demand, and run a real-time auction for your attention. All in about 200 milliseconds.

It’s taken decades to build this infrastructure. But ironically, just as it reached peak efficiency, AI agents arrived that don’t click anything. The system was designed for humans, not machines.

It reminded me of a tweet from Paul Graham, where he described how Mercedes V8 engines, after decades of refinement, reached a kind of mechanical perfection, just as they were about to be replaced. The open web’s ad infrastructure feels the same: after decades of refinement (from banner ads in the late '90s to sub-second auctions today), it achieved mechanical perfection, right as it was becoming obsolete.

​

Ben raises two points that are key to understanding what the world looks like today for the open web and ad revenue:

It’s widely understood that Google has been propping up the long tail of the open web by directing users to these sites through search. The hypothesis is: once Google loses the incentive to keep doing that, users stop visiting and ad revenue for those sites trends toward zero.

​

Why does it matter if the long tail gets cut?

​

This question easily drifts into ideological territory i.e. debates about the internet’s open, decentralized promise. Personally, I think that promise alone is worth preserving. But we can't assume all actors on the web have the same ideological alignment.

So, in this section, I’ll try to steelman both sides: why the long tail does and doesn't matter.

​

The long tail doesn't matter:

A strong case can be made that much of the long tail is filled with outdated or duplicated content. Centralized platforms are often better at curating high-signal material (e.g. Instagram and YouTube), which helps explain why users increasingly gravitate toward them. Because the content quality is generally low, the ad revenue from these sites was likely small. Google probably didn’t earn much from them.

This creates a natural flywheel: as users consolidate on platforms, more creators follow, and the cycle reinforces itself. If browsing the internet becomes the job of agents instead of humans, the need for websites fades further. Agents don’t want visual content or rich interfaces; they want structured, machine-readable APIs. Walled gardens are best positioned to deliver that kind of infrastructure.

​

The long tail does matter:

Creators on the open web aren’t beholden to the rules and incentives of walled gardens. Innovation and freedom of expression thrive outside algorithmic monocultures. Independent journalists, researchers, and creators can publish without gatekeepers. This openness is essential for progress in science and journalism.

On the infrastructure side, a decentralized web is more resilient to censorship and single points of failure. The reasoning is similar to the case for cryptocurrency: open systems with built-in verification and no need to trust a central authority.

Ironically, AI agents are powerful today largely because the open web served as a rich and diverse training ground. If the long tail disappears, future AI models risk being trained on a narrower and more homogenous dataset dominated by walled gardens.

If you find yourself leaning toward the view that the long tail matters, keep reading. Next, we’ll look at how Ben’s proposed solution might help preserve it.

​

Replacing the Ad Economy: From Attention to Transaction

​

In the early days of Ethereum (circa 2016), among all the other hubris, I remember a recurring narrative: blockchains could be used by AI agents for payments. Imagine fully autonomous self-driving cars that could pay for their own maintenance, refueling, and even accept payments from passengers.

Back then, the breakthrough wasn’t the AI-based self-driving car itself, it was the fact that the car had its own bank account. There was a belief that it was the digital wallet that unlocked the last hurdle to full autonomy.

I mention this because when I read Ben’s article, the idea immediately clicked. The basic concept of giving an AI agent a wallet doesn't feel outlandish like it did back in 2016. This time, it felt obvious, even practical given the problem space we now find ourselves in. Additionally, advancements in blockchain scaling over recent years has finally made microtransactions viable.

​

To recap, the idea outlined in Ben's post is depicted in the diagram below:

​

But this raises a few obvious questions.

Why use a blockchain at all? Isn’t that overkill when we already have payment networks like Visa or Stripe? And even if you could get agents to pay per article, what stops them from just sharing it for free once it’s unlocked?

If we’re really going to replace the ad economy on the open web with agent-based payments, we need to explore these edge cases and make sure crypto isn’t just a solution in search of a problem.

​

Why use a blockchain?

​

Favorable Fee Structures

Traditional ways of making payments on the internet aren't optimized for microtransactions. The fee structure alone prevents them from being a viable option. The table below shows each mechanisms viability for supporting microtransactions:

​

​

The table shows that the only cost-effective payment rails for microtransactions today are Ethereum Layer 2s and potentially other chains like Solana, though one would then need to argue for their decentralization.

Beyond fees, think back to the autonomous vehicle example. Giving an AI agent or inanimate object a digital wallet is the closest thing we have to a native financial primitive for software. Stripe’s founder, Patrick Collison, makes a similar point, that crypto rails are less encumbered by the bureaucracy and outdated nature of traditional finance:

​

Simple onboarding

Another major reason to use blockchain rails is that it's trivially simple to give an AI agent a crypto wallet, while giving that same agent a traditional bank account is a regulatory and logistical nightmare. Who owns the account? Who does the KYC? Shoehorning AI agents into legacy financial frameworks feels like an unnecessary burden on progress.

​

Low Stakes

One of the main reasons people hesitate to embrace blockchain infrastructure is its irreversible nature. Once you send a transaction, it is final. There is no "undo" button and no bank to recover the funds if they are sent to the wrong address. While UX and middleware have improved to reduce this risk, it is still a valid concern.

I think the low-stakes nature of microtransactions, coupled with the high-stakes nature of irreversible blockchains, makes it easier for people to accept this as a use case. People simply aren't as worried about misplacing fractions of a penny.

​

Without going through every detail, here are a few more reasons why digital wallets are well suited to AI agents:
​

• Global by default​
  One wallet works across jurisdictions, borderless payments.
• Massive scalability​
  You can spin up thousands of wallets instantly for thousands of agents. This is impossible with bank accounts.

​

The Pay Once, Free Forever Problem

​

If an AI agent unlocks a piece of content, what’s stopping that same agent from redistributing it freely across the entire web?

This is a real and difficult problem. But it isn’t new. The internet has been wrestling with it since its inception.

Why does Spotify exist, even though anyone with enough effort can pirate music? Why is the lowest price for a CryptoPunk $94,000 at the time of writing, when anyone can copy and paste the image?

If it’s technically possible to get these digital items for free, why do so many people still choose to pay?

Let’s stick with the Spotify example, it’s the more relatable of the two. (I’m not here to convince you a JPEG should be worth over $90k… maybe in a future post.)

If you search around, you’ll find that the main reason people stick with Spotify over free, illegal alternatives is simple: convenience. Spotify turned a clunky, fragmented experience into a smooth, reliable portal to your favorite music.

Now, in the case of amoral AI agents, they don't care about doing the right thing, they simply have a task to complete. These agents could technically find the content for free elsewhere, but they’d likely prefer to pay a few cents to get it instantly, in context, and without friction.

If humans remain in control and embed their own morality or legal constraints into these agents, then, especially under the threat of copyright infringement, the agents will be incentivized to 'do the right thing' and pay the fraction of a penny required to legally access the content.

​

Technical Deterrents

As we know, solving the “pay once, free forever” problem is difficult. However, there are technical strategies we can use to reduce the likelihood of this behavior occurring at scale.

​

Watermarking​
All content returned to agents should be watermarked. This means embedding a digital signature that includes the content, the requesting agent, and the originating domain (e.g. chatgpt.com). This creates accountability. If the content leaks, the source can be traced. Over time, reputation systems would naturally emerge around these identifiers, and bad behavior (like redistribution) could be punished with bans or throttling.

​

Staking/Bonds

As part of an AI agent’s registration with a content service, the agent could be required to stake tokens or approve a larger allowance (for the publisher) to signal good behavior. If the agent is later found to have leaked content, the content service, following pre-agreed rules, could claim a penalty from the agent’s smart account.

For example, the agent could issue the service a session key (e.g. Porto accounts) with limited rights to pull a predefined penalty amount, but only if they can cryptographically prove the content was leaked by that agent (via watermarks). This creates economic consequences for bad behavior and aligns incentives toward compliance.

​

I see this more as an engineering problem to be solved, not an existential reason why microtransactions can’t replace ad-based revenue models.

​

Who is working on this?

The project most actively working on this new technology is Coinbase's x402 project:

Their whitepaper describes the architecture of their protocol. It's worth mentioning that this protocol is not focused on providing a solution for ad revenue on the web, but instead, provides the plumbing to do so.

While the project is advertised as a payments standard for the web, much of the solution relies on using Coinbase's x402 facilitator. The facilitator is a centralized server that facilitates the buyer's purchase request and ensures it settles onchain.

The idea is that you can write your own facilitator, but using Coinbase’s CDP is easier. It seems the facilitator adds a layer of indirection, likely to avoid being opinionated about which chain handles settlement. A more pessimistic take is that centralizing this part of the system positions Coinbase to potentially dominate the market if the standard gains adoption.

That said, the preprocessing of purchase requests could be deferred to the RPC layer or handled directly onchain at the smart contract level. This would eliminate the need for a centralized server (i.e. facilitator), reduce setup complexity, and potentially accelerate adoption.

Building upon our previous diagrams, the resulting architecture may end up looking something more like this:

Where do we go from here?
​

As AI agents begin to traverse the web autonomously, the economics of digital content must evolve with them. Microtransactions offer a viable alternative to surveillance-based ad models but only if the infrastructure is lightweight, decentralized, and usable. Whether or not it’s through protocols like x402, the foundation is being laid for a more open, agent-friendly internet. The next step is turning these primitives into real, developer-friendly products that make paying for content as seamless as consuming it.