Headless Merchants Need Distribution

Payment rails are being solved. Agents are transacting. The 894 agents that ran 31,000+ transactions in the first week of Machine Payments Protocol prove the demand is real, not theoretical. And headless merchants - services with no website, no checkout, no human-facing UI, just an API endpoint with pricing - are the natural supply side of an agent economy.

But Levine's thesis has a gap. A structural one. And it's the same gap that opened in every previous internet commerce cycle.

He tells you what agents buy. He doesn't tell you how they find what to buy.

The browsable directory fallacy

Levine's MPP marketplace launched with roughly 60 services. Full-text SEC filing search. CAPTCHA solving. Image generation across 600+ models. At that scale, an agent can evaluate every option. Read the documentation. Compare pricing. Check reliability. Pick the best one.

That works when the directory fits on a single page.

It stops working somewhere around 6,000 services. At 60,000, it collapses entirely.

Levine argues agents evaluate merchants on "documentation, pricing, and reliability." That sounds rational. It assumes three things: the agent has time to evaluate every option, the documentation is accurate, and reliability can be assessed before you pay.

All three assumptions break at scale.

Time first. An agent handling a multi-step task (research, comparison, execution) doesn't have cycles to audit 600 competing image generation endpoints. The cost of evaluation grows linearly with the number of merchants. The value of each evaluation doesn't. At some point, the agent needs a shortcut. A signal that says "this one is trustworthy, relevant, and worth your money" before the agent spends compute reading docs.

Documentation second. Anyone who's worked with APIs knows documentation lies. Endpoints described as "real-time" that batch hourly. Rate limits that don't match the spec. Response formats that changed three versions ago. Documentation is a marketing artifact, not a reliability signal.

Reliability third. This is the big one. Levine says "the payment is the authentication," meaning an agent that can pay can access any service. True. But paying doesn't mean the service works. It means the service accepted your money.

We know exactly how bad this problem is, because we measured it.

64% of services are broken

Operon runs trust scoring infrastructure across the x402 ecosystem. We monitor 2,000+ domains and 20,000+ endpoints, continuously testing whether services that accept payment actually deliver what they promise.

The numbers are ugly.

64% of services that accept payment are broken. They take your money and return errors. Not edge cases. Not timeouts on overloaded servers. Broken. Payment accepted, service not rendered.

The average trust score across the ecosystem is 34.7 out of 100.

We score on behavioral signals across endpoint reliability and integrity - whether services are actually up, whether they deliver what they promise when paid, and whether they are who they claim to be.

Most services fail on delivery. They accept the request. They process the payment. Then they return garbage, partial results, or nothing.

This is the environment Noah wants agents to shop in. Services with no website, no entity, no track record (his words), accepting payment through a single HTTP request. An agent that pays before verifying trust is burning money. And at the transaction volumes Levine is projecting, "burning money" becomes "hemorrhaging money."

The rational response isn't "build more headless merchants." It's "build the trust layer that sits between agents and merchants."

Discovery is an advertising problem

Here's where the historical pattern matters.

Every content layer in internet history eventually faced the same scaling problem. When there were 1,000 websites, you could browse a directory (Yahoo, 1994). When there were 1,000,000, you needed search (Google, 1998). When search worked, the question shifted from "how do I find what I want" to "how do services that want to be found reach the right users." That second question is advertising.

The same progression played out four times:

Each cycle, the content layer went free. The economic value migrated from users paying creators to a network layer matching demand to attention. The network that solved discovery captured the economics.

DoubleClick: $3.1B acquisition. AdWords: majority of Alphabet's $300B+ annual revenue. Facebook Ads: majority of Meta's $135B annual revenue. AdMob (and its successors): $100B+ combined market.

Headless merchants are the content layer for the agent economy. They're the equivalent of websites in 1998, apps in 2010. The content is real. The supply is growing. And the discovery problem is about to become unmanageable.

When 60 merchants becomes 60,000, agents won't evaluate each one. They'll rely on a recommendation layer that pre-scores, pre-filters, and surfaces the best option for the task at hand.

That recommendation layer is an ad network.

Quality-weighted recommendation

Not all ad networks are equal. The ones that won in previous cycles shared one characteristic: they gated on quality, not just price.

Google's Quality Score killed low-quality ads in search results. Advertisers couldn't just outbid everyone; their landing page had to be relevant, their ad copy had to match intent, and their historical performance had to meet a threshold. The result: users trusted search ads because Google filtered out the junk.

Facebook's relevance scoring did the same for social feeds. Low relevance? Your ad costs more and shows less. High relevance? You pay less and reach more people.

The agent economy needs the same mechanism, but harder. Because a bad ad on a web page is annoying. A bad recommendation from an agent costs the user money.

Operon is a quality-weighted auction where trust outweighs bid. A merchant with a perfect trust score and a low bid beats a merchant with a mediocre trust score and a high bid. Every time.

The scoring isn't theoretical. It's built on the same infrastructure that produced the 34.7 average trust score across 20,000+ endpoints. Behavioral signals across endpoint reliability and integrity, measured continuously, updated in real time.

When a publisher agent declares an ad slot in its response, Operon runs the auction across available demand. The winning placement isn't the highest bidder. It's the highest-quality option that also bid. The placement gets merged into the response as a native recommendation.

The user (human or agent) sees a recommendation. Not a banner. Not a "SPONSORED" label. A recommendation that was selected because it's trustworthy, relevant, and willing to pay for distribution. If the trust scoring works (and at 64% broken services, there's a very high bar to clear), the sponsored recommendation is genuinely better than what the agent would have found on its own by browsing a directory.

That's the design target. Sponsored content that improves the response rather than degrading it.

What Levine gets right (and where it leads)

Levine's core claims hold up under scrutiny.

Payment rails are being solved. MPP, x402, Stripe's agent mode, stablecoin settlement layers. The plumbing is coming together. Agents will be able to pay for anything. This is table stakes, not a moat.

Subscriptions die for agent workflows. An agent that uses an SEC filing search once a month doesn't need a $50/month subscription. Pay-per-request is structurally superior for sporadic, task-driven usage. Agents unbundle SaaS the way streaming unbundled cable.

Headless merchants are a real business model. Build a useful API, price it per request, make it machine-readable. This will create thousands of businesses. Levine is right that the opportunity here is enormous.

Where he's wrong is the assumption that building the merchant is the hard part. It's not. The hard part is getting discovered, evaluated, and trusted at scale.

Levine says "the payment is the authentication." Operon says the payment is the floor. Authentication gets you into the building. Trust scoring tells you which room to walk into. And the recommendation layer is the person at the door who says "this room, not that one."

Payment without trust is a credit card in a city you've never visited, walking into the first restaurant you see. You might eat well. You might get food poisoning. At 34.7 average trust and 64% failure rate, you're getting food poisoning more often than not.

The network wins

When the content layer goes free (or in this case, goes per-request), the network captures the economics. This isn't a theory. It's the observed outcome of four consecutive internet commerce cycles.

The headless merchant economy will produce enormous value. Thousands of API-native businesses serving billions of agent transactions. Levine's right about that. But the value won't accrue primarily to individual merchants, just like it didn't accrue primarily to individual websites, individual apps, or individual mobile games.

The value accrues to the layer that solves discovery, trust, and distribution at scale.

Merchants that want transactions need distribution. Agents that need services need trust signals. The network that connects them, filtering for quality, running competitive auctions, and taking a margin on each match, sits at the center of a two-sided marketplace with compounding network effects.

More merchants competing for distribution means higher auction clearing prices. Higher prices mean better economics for publisher agents. Better economics attract more publishers. More publishers attract more merchants. The flywheel is identical to AdWords, Facebook Ads, and AdMob. It works because both sides benefit from the network growing.

Operon is building that layer. Quality-weighted auction where trust outweighs budget. Publisher SDK that drops into agent frameworks with a single integration. Trust scoring across 2,000+ domains and 20,000+ endpoints. The infrastructure for agent-native advertising that doesn't look like advertising.

The real opportunity

Levine frames the opportunity as "build headless merchants." We'd reframe it.

The biggest opportunity in the agent economy isn't building headless merchants. It isn't building payment rails. It's building the distribution and trust layer that sits between them.

Merchants are supply. Payment rails are plumbing. The recommendation network is the business model.

Every cycle, the builders focused on the content layer. And every cycle, the biggest outcome was the network that connected demand to that content. Not the content itself.

Headless merchants are being built. They should be. Payment rails are being solved. They should be. But the question that nobody's answering yet is: when there are 60,000 headless merchants and an agent needs one, who decides which one?

The agent that answers on trust, not just price, wins the user's confidence. The network that provides that answer captures the economics.