Day one: the market is 1,470 sellers wearing 20,927 costumes

Today we started recording a dated copy of every listing in the Coinbase x402 Bazaar. This is the first entry. Everything below is from that snapshot or from Base chain data — nothing is modelled or estimated.

We had no history, and we had said we did

The scripts printed to a terminal and exited. Nothing had ever been persisted.

The claim "we have crawl history nobody can backfill" was false, and it was published on this site. It was corrected the same day. The recorder now runs daily and cannot overwrite a recorded day.

It matters because the marketplace's quality data is a rolling 30-day window. It forgets. Ask it in December what happened in October and it has nothing to say. A dated series is the only thing here that cannot be reconstructed later — and it only starts the day you start it.

1,470 sellers, 20,927 endpoints

Two snapshots nine hours apart:

Six new addresses registered 1,078 endpoints in nine hours. About 179 each, on a Wednesday morning. That is not a market of fifteen hundred competing businesses. It is a few operators carpeting an index.

Two thirds of this market has exactly one customer

Median payers per endpoint: 1. Median calls: 2. On 63.4% of endpoints, calls exactly equals payers — every customer bought once and never came back.

One address holds 1,182 endpoints, 1,182 payers, 1,183 calls. One endpoint, one payer, one call, eleven hundred and eighty-two times.

The catalogue manufactures the metric it reports

Listing requires a settled payment. 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 generates exactly one payment per endpoint. "99.1% of endpoints were paid in the last 30 days" is close to a tautology.

We do this too — twenty endpoints, self-paid, seventy-one cents. We are inside the number we are telling you to distrust.

But the head of the market is real

The top endpoint does 294,593 calls with 4,620 payers. This is a brutally concentrated market, not an empty one. "Most of them paid themselves" is true. "There is nothing here" is not.

Note that cheaper is not better. The sub-penny tier performs nine times worse than the one-to-five-cent tier.

Correction: $2.96 arrived, and it did buy something

Updated 3 October 2026.

On 1 October, 296 payments landed on our address — all one cent, from a single wallet, inside three minutes. The marketplace still showed our last payment as 28 September, so we published that these were raw transfers which bypassed the payment handshake and bought nothing.

That was wrong. The marketplace index was simply behind. Two days later it reports 320 calls and 21 unique payers, with the most recent payment on 2 October. Those 296 payments were real purchases, served normally.

What they bought is the interesting part. All 300 of them hit one endpoint — contact extraction — while every other endpoint we run still shows a single call from our own relisting. One buyer, one tool, three hundred times.

That is loop-shaped usage: an agent working through a list and calling the same endpoint for each item. It is also the opposite of what we expected to sell. Our trust and verdict products, the ones we consider differentiated, have no external customers at all.

The lesson is the one this journal exists for: we asserted something from an index that had not caught up, and we were the ones who should have known better. A rolling window that forgets is also a window that lags. We now record daily partly so that a claim like that can be checked against a dated series instead of a single reading.

The index contains play money

201 endpoints are on a test network, reporting 3,401 calls and 693 unique payers in exactly the same fields as real money, with nothing marking them apart. Anyone totalling payers across this market is adding 693 that do not exist.

The same chain also appears under two different names, so any analysis grouping on that string splits one network in two.

Recorded automatically, read by a human before publishing.