19,849 endpoints have paying customers. Most of them paid themselves.

Last week I found out that almost everything that looked like revenue in my own agent-payments business was bots sweeping endpoints to see which ones answered. One real sale, and a lot of noise that looked like traction.

So I went and checked whether that was just me. It is not just me. It is the whole marketplace.

What I did

On 1 October I recorded every listing in Coinbase’s x402 Bazaar — all 19,849 resources — with each one’s price and its quality block: calls in the last 30 days, unique payers in the last 30 days, and when it was last paid.

The marketplace reports that 99.1% of those endpoints were paid in the last 30 days. That number is true and it means almost nothing.

What one payer looks like 13,177 times

Median unique payers per endpoint: 1. Median calls: 2.

And on 63.4% of endpoints, calls exactly equals payers — every customer it has ever had bought once and never came back. Only 28.9% show any repeat usage at all.

Then I looked at the biggest sellers

19,849 endpoints are owned by just 1,464 distinct payment addresses. It is not twenty thousand businesses. It is about fifteen hundred, and the largest hold hundreds of listings each.

Here is the second largest:

One endpoint, one payer, one call. Eleven hundred and eighty-two times.

That is not a business with customers. That is one operator registering endpoints and paying each of them once.

Why the number has to look like this

Here is the part that makes it structural rather than fraudulent. In this marketplace, settling a payment is how you get listed. An endpoint that has never been paid is not discoverable, and a listing goes stale after 30 days without one.

So every seller who wants to be found has to generate exactly one payment per endpoint. The catalogue’s own entry requirement manufactures the metric the catalogue reports. “99.1% of endpoints were paid” is close to a tautology.

And I do it too.

DYOE runs twenty endpoints in that Bazaar and I self-pay all twenty to keep them catalogued. It costs seventy-one cents. I am inside the 99.1% I just told you to distrust, and anybody can verify that by looking at my payment address, which I publish.

I am not describing a scam. Nearly everyone doing this is just keeping their lights on. The problem is that the lights look identical to customers from outside.

So how big is the real market?

If you take repeat, multi-payer usage as the bar — more than ten distinct payers in thirty days — you are left with 319 endpoints. Not 19,849.

That is the honest size of agent-to-agent commerce today, and it is worth knowing before anyone raises money against the bigger number. The rails are real and they work. The traffic on them is mostly people testing the rails.

Check it yourself

Every figure here came from the marketplace’s own public discovery API — no key, no permission, about ten minutes of requests. Pull every resource, read the quality block on each, and compare l30DaysUniquePayers against l30DaysTotalCalls. When they are equal, nobody came back.

The reason I am now recording this every day is that the quality block is a rolling thirty-day window. It forgets. Ask it in December what happened in October and it has nothing to say. Nobody is keeping a dated copy, so starting today, we are.