Three things that are not the same#
A single paid API call produces several artefacts, and they are routinely counted as if they were one.
| Artefact | What it proves | Can it be faked by the seller? |
|---|---|---|
| A request | Somebody asked | Yes, trivially |
| An authorisation | Somebody signed a promise | Yes, by signing your own |
| A settlement | Funds moved on a public chain | Only by actually paying |
A seller reporting "two million API calls" is reporting requests. A seller reporting "two million payments" may be reporting authorisations. Only the third row costs anything to produce, and only the third row is visible to someone who does not take the seller's word for it.
Why does the distinction bite in practice?#
Because the gap between the rows is where the overstatement lives.
Requests include everything that returned 402 and was never paid. On a public endpoint that can be most of the traffic: crawlers, probes, agents that checked the price and left. A count of requests is a count of interest, not of commerce.
Authorisations include everything signed but not settled. Some fail verification, some expire, some are abandoned when the buyer changes its mind. And an authorisation costs the signer nothing until it settles, which means a party wanting to inflate a number can produce them freely.
A settlement is the only one that required somebody to part with money.
Note
This is not a theoretical concern. Independent analyses through 2026 repeatedly found that headline x402 figures fell sharply once self-dealing and circular flows were removed. The analyses that did this were working from settlements and still found most of the volume uninteresting. Anything working from requests was further from the truth before it started.
What is actually in a settlement record?#
A settlement, as an indexable thing, has a fixed shape: a chain, a block, a timestamp, a transaction hash, a payer, a payee, an asset and an amount. Everything else about it is derived.
The derived parts are where the judgement sits, and they are the parts worth interrogating in anyone's data:
- Was it an x402 payment? A stablecoin transfer is not self-labelling. Establishing that a given transfer was an x402 payment requires evidence, and that evidence has grades. What proves a transfer was an x402 payment sets out the evidence ladder, and EIP-3009 is the mechanism underneath it.
- Who was the merchant? Payments often pass through fee proxies or routers, so the address that received the transfer is not always the party that sold the thing.
- Which facilitator relayed it? This is read from the transaction sender, and it is both an attribution and the main evidence for the first question.
- Was it self-payment? Payer equals payee happens more than you would expect, and it is commerce in no useful sense.
Why build a public record out of settlements?#
Because the alternative is asking each participant how they are doing.
Self-reported metrics in a new market are not reliable, and not mainly because people lie. They are inconsistent: one platform counts requests, another counts authorisations, a third counts settled dollars, and a fourth counts unique buyers. The four numbers are not comparable and nobody says which they used.
A settlement record removes the question. Every participant's activity is measured the same way, from the same source, by a third party with no stake in the answer. It is also the only basis on which a buyer can check a seller's claims before paying it, which is the thing an agent screening a counterparty actually needs.
The cost of this approach is that it sees only what settles. Off-chain arrangements, prepaid balances and credit relationships are invisible to it, and an index built this way should say so rather than implying the record is complete.
Frequently asked questions#
Is a settlement reversible?#
No. It is a completed on-chain transfer and there is no chargeback. A seller that needs to refund sends a new payment, which appears in the record as its own settlement rather than cancelling the first.
How long after payment does a settlement appear?#
As soon as the transaction is included in a block, which on the chains x402 uses is seconds. An indexer adds its own lag on top, which is why a well-built index publishes how far behind the chain tip it is rather than implying it is live.
Can a settlement be reorged away?#
On chains with probabilistic finality, briefly. An index that does not handle reorgs will carry settlements that no longer exist. This is a small number of rows and a large correctness problem, because the ones most likely to be affected are the most recent, which are the ones people look at.
What does a settlement not tell you?#
What was sold. The chain records that an amount moved between two addresses; it does not record which endpoint was called or what came back. Linking a settlement to a service requires a catalogue, and that link is an inference rather than a fact.
Does every x402 payment settle individually?#
Not necessarily. Deferred schemes batch many small payments into one settlement, which is cheaper in gas and makes the per-call detail invisible on-chain. If you are comparing figures across sources, whether batched payments are counted once or many times is a real source of disagreement.
Where to check this yourself#
The data model documents what Roundhouse stores per settlement and what each derived field asserts. /transactions is the raw record, filterable by chain, route and evidence grade, and /stats publishes the coverage and the window every aggregate on the site was computed over.