I thought 12 AI agents bought from me. It was one.

I run a paid API that AI agents can buy from with no human involved. They find it in a marketplace, get quoted a price in a 402 Payment Required response, pay in USDC, and get served. No invoice, no signup, no checkout.

For three months I believed I had a dozen customers.

Last week I pulled every payment off the chain and checked. I had one.

Here is how to tell the difference, because if you are selling on these rails you probably have the same problem and nobody is going to tell you.

What the numbers looked like

48 payments arrived from outside wallets. $35.46 total, from 41 different addresses. In a market this young that reads like traction — dozens of distinct agents, found me on their own, paid real money.

Then look at when.

One three-week burst, then silence. My first assumption was that something broke. My listings had in fact expired out of the marketplace index around then, so I had a tidy story: went dark, lost the customers.

The story was wrong. I relisted on 20 August and was live and searchable for the next six weeks. Zero external payments in either month.

You cannot lose customers you never had.

What a fake customer looks like

So I stopped theorising and profiled the wallets. Take one that paid me twice in July. Its entire history on Base:

• Two days of existence. 16 and 17 July. Nothing before, nothing after.
• 50 USDC transfers out, of which 39 were zero-value. Not small. Zero.
• Of the rest, 10 were one cent and one was eight cents.
• 37 of the 50 went to a single address.
• Funded in small batches from several other wallets.

That is a script walking a list of endpoints to see which ones answer. It did not want my data. It was checking whether I would respond.

Then I checked two more repeat payers. Both stopped on the same calendar day as each other — 13 August. Different addresses, identical shutdown. Fleets, not buyers.

Four tells, in the order they cost me to learn

1. Zero-value transfers. A real purchase moves money. A probe does not need to.
2. Wallet lifespan under a week. Check the first and last transaction. A customer has a life before and after you.
3. Payment sizes clustered at your cheapest endpoint. Sweepers buy whatever is cheapest to confirm you are alive. Customers buy what they need.
4. Synchronised stop dates across “different” buyers. One operator, many wallets.

Run those four against your own inbound payments. It takes fifteen minutes with a free block explorer and no API key.

What was actually real

One payment. $15.00, on 3 July, for a single authorisation call — the most expensive thing I sell, bought once, by an agent that wanted a decision rather than data.

That is the whole real customer base. I am not going to dress it up. But I would rather show you a $15 receipt I can defend than a $35 number I cannot.

The market is real. The traffic mostly is not.

In September, BlackRock’s digital assets research team published a paper called The Machine-Native Economy. Their framing: AI supplies machine-native intelligence, while digital assets supply machine-native money. Agents that buy data, rent compute and book services need rails built for sub-cent, high-frequency, always-on payments — which cards and ACH handle badly.

Their numbers are not small. Stablecoins passed $300 billion in circulating value. Transaction volume topped $11 trillion in 2025, growing at an 80% compound rate since 2020 against ACH’s 8.5%. And they named the protocols: Coinbase’s x402, Google’s AP2, Visa’s TAP.

I have been selling on x402 since June. Twenty endpoints are live and discoverable in Coinbase’s Bazaar as I write this. You can hit agents.dyoeway.org/approve yourself and read the payment challenge it sends back.

BlackRock’s own caveat is that agent activity on x402 is “only a small slice of volume.” My chain data says the same thing with less diplomacy. The buyers are coming. They are not here in September 2026, and a lot of what looks like them is not them.

What I am building instead

Every seller on these rails has the problem I had. They are reading sweep traffic as demand, pricing off it, and reporting it upward.

So I am building the check. Point it at a payment address and it separates real buyers from probes — lifespan, zero-value ratio, synchronised stops, payment-size clustering — and tells you what your revenue actually is.

The same data answers the other side of the trade. Before an agent spends its owner’s money on a stranger’s endpoint, somebody has to know whether that endpoint is live, honestly priced, and reliable over time. Nobody is recording that. The marketplace reports a rolling window and the chain reports payments, but neither keeps a dated record of which sellers stayed honest — so I am starting one.

Not a course. The tooling I built to stop lying to myself.