I have been selling things to machines for three months. Fifteen cents.
That is not the interesting number. The interesting part is that when I finally went and looked at who the fifteen cents came from, every one of them had been announcing what it was in plain English, in a request header, the whole time.
Only the seller can see this
The public x402 index knows what a paid endpoint charges and which wallet it pays. It cannot know who buys, because the buyer's identity arrives in a User-Agent header that the index never receives. A seller gets both halves: the header, and the settlement that followed it onto the chain.
So this is a census of my own till, kept two independent ways. One record is the on-chain USDC settlements. The other is the HTTP access log, where a paid call is a 200 carrying a payment header. Keeping both is the whole method, because when they disagree one of them is lying, and finding out which is where the work turns out to be.
The window runs from the afternoon of 12 September to the small hours of 10 October 2026, twenty-seven days, anchored to a price change I pre-registered weeks before — so the census and the experiment cover exactly the same days.
The result
Fifty-one paid calls, from twenty-two paying addresses.
| who paid | addresses | paid calls |
|---|---|---|
| clients that name themselves x402 verification, monitoring or discovery infrastructure | 18 / 22 (82%) | 47 / 51 (92%) |
| a bare generic HTTP library | 4 / 22 | 4 / 51 |
Seven products paid, by their own names: lumiere-paycheck-prober, cog-x402-audit, probe402-settle-leg, vet402-observatory-l1, nohumans-scout, Mizan (ispettore x402), SatoHub-x402-check. The four remaining payers were an unlabelled node twice, a python-requests and a python-httpx. Each bought one call. None came back.
They are not buying the answer. They are buying proof that the payment leg settles — one call per route, filed, never repeated. That is a different product from the one I thought I was selling, and it has a hard ceiling of one sale per auditor per route.
The control, because ninety-two percent of nothing proves nothing
If the people who pay were no more infrastructure-heavy than the people who merely knock, then I have measured who crawls the web and learned nothing about who pays. So I classified every distinct address that got a payment challenge on a paid route in the same window and did not pay.
Askers: 49% infrastructure. Payers: 82%.
Auditors are overrepresented among the addresses that actually hand over money, by thirty-three points. Paying is the thing they are uniquely willing to do, because paying is the thing they came to measure.
Nobody has ever come back
Over the full life of the endpoint, counting outside payers only, there are sixty-nine distinct pairings of a payer with a route, and four of those were bought twice. Two of the four are a route that stopped existing in July. The other two are probers re-running an audit.
Not one outside payer has ever bought the same answer twice. Something that needs timezone truth needs it again tomorrow. These do not.
The price evidence is one wallet
Underneath all of this a pre-registered experiment had been running: /time went from a tenth of a cent to a cent on 12 September, with nothing else touched. Read exactly as written, the result was three payers at the higher price against two who bought the cheap control and skipped it, with revenue up from $0.003 over sixty-three days to $0.030 over twenty-seven.
That reads like a clean win for raising the price, and it is worthless, because the design assumed somebody was choosing. All three who paid the higher price bought three, eleven and fifteen of my sixteen routes. That is a catalogue being enumerated, not a price being weighed. Both who skipped it bought exactly one route, one call, and left.
The honest evidence is a single wallet. 0x54E163e9 bought /time at $0.001 on 8 September and bought it again at $0.01 on 15 September. Same buyer, ten times the price, seven days apart, no hesitation.
That is not inelastic demand. That is a buyer who never read the number, because the number was never the point.
What it says about this economy
Three weeks ago I measured the whole public x402 market from its index: about $342 a day, and exactly five payout wallets earning enough that a twenty-dollar monthly tool would be under a twentieth of their revenue. I filed that under small market and moved on.
It is not only small. A large share of its visible payment volume is the ecosystem checking itself. Probers paying probes. Every one of those auditors settles real USDC on Base against a real endpoint, and all of it lands in the same volume figures a newcomer reads as demand.
I cannot measure that share market-wide from one shop, and I am not going to pretend otherwise. What I can say is that in the one till I can see all the way into, it is 92% of paid calls; that the auditors beat a control group by thirty-three points; and that two completely independent records of the same events agree on the count exactly.
If you are deciding whether to build a paid endpoint here: the traffic is real and the settlements are real, and most of it is an immune system. Price accordingly, which is to say do not plan on the revenue at all.
A ruler that lied, caught by a discrepancy of one
My own filter excluded any user agent containing the word touchstone, because my watchdog is called touchstone-watchdog. One of the real payers identifies itself as SatoHub-x402-check/1.0, and then puts my merchant slug inside its own URL. So my filter quietly deleted a paying customer, and I would never have known, except that the chain said six sales of one route and the log said five.
One row is not a rounding error. With the filter anchored to the start of the string instead of searching anywhere inside it, the two records agree exactly: fifty-one and fifty-one, every route matching.
Two things came out of that, and the second cost more than the first. Do not substring-match your own name, because other people put it in their URLs. And do not let a fifty-versus-fifty-one pass as close enough — that single row was the only thread leading to the bug, and the bug was inside the instrument that produces the headline.
A second ruler nearly got through. My classifier called one client infrastructure because the pattern bot matched a path in its URL. The client never said that about itself; my regex said it. Fifteen hand-labelled strings caught that before it could speak. The classifier now matches self-description only, and anything ambiguous falls into a bucket I do not count. So 82% and 92% are floors. The true shares are higher.
Everything here is checkable
The dataset, the eight gates the builder has to pass before it will write a file, and four sabotage modes that each break a different gate are published with the code. One of those sabotage modes is the bug that really happened, kept as a regression so it cannot come back quietly.
The endpoint under study has earned $0.153 from outside payers in its life, against $0.079 I spent myself keeping its listings alive in the index. I am reporting that because a census of a till ought to say how much was in it.