Research note · Census
The state of x402 supply: what’s actually real
2026-08-24 · 9 min
The question that keeps coming up in the x402 working groups is some version of “what's real?” — thousands of services are listed, and nobody can tell which ones answer, deliver, or earn. We measured it. As of 2026-08-24: three out of four probed catalog listings are dead, one seller in fourteen shows observed on-chain traction, and ten sellers carry roughly ninety percent of everything we can attribute. This note publishes the numbers, the bases behind them, and the honesty box that goes with them.
The x402 catalog looks enormous from the outside. Our census indexes 138,158 endpoints across 5,645 sellers, assembled from public catalogs and the 402 offers the endpoints themselves serve. Numbers like that get quoted as evidence of a thriving machine-payments economy. But a catalog entry is a claim, not a fact: somebody once said this URL sells something. The whole point of this census is to check claims against things we can observe — our own probes, our own paid purchases, and settlement indexed from the chains themselves.
A listing is a sentence somebody wrote. An answered probe, a delivered purchase, a settled payment — those are events that happened. The gap between the two is the actual state of x402 supply.
Three out of four probed listings are dead
We probe catalogued endpoints continuously — a real HTTP request to the catalogued URL, looking for a live response or a valid 402 payment challenge. In the 48 hours before measurement we probed 80,452 endpoints. 61,578 of them — 76.5% — were dead: no valid response at the address the catalog advertises. Another 57,706 catalogued endpoints are still in the probe backlog and are counted neither way.
catalogued endpoints 138,158 (5,645 sellers) probed in last 48h 80,452 answered alive 18,874 23.5% of probed dead at catalogued URL 61,578 76.5% of probed not yet probed 57,706 (excluded from the rate)
An agent choosing from raw catalogs therefore has roughly a one-in-four chance that the endpoint even answers — before any question of whether it does what it claims. This is not an accusation; catalogs accumulate and services churn. It is simply what the catalog is, measured.
Observed traction is rare — and very concentrated
Liveness is the lowest bar. The next question is whether anyone actually pays a seller. Our provider score assigns a tier only from observed evidence — indexed settlement, probes, paid tests — never from a seller’s own claims. Of 5,682 scored sellers, 423 (7.4%) reach any traction tier above Listed. The other 92.5% are Listed or Unrated: present in catalogs, no corroborating on-chain evidence yet.
Corroborated 18
Established 38
Emerging 367
Listed 3,414
Unrated 1,845
------
with observed traction (Emerging+) 423 / 5,682 = 7.4%And within that thin band of real activity, the money concentrates hard. Of the $29.9M in settlement we can attribute to catalogued sellers, the top ten sellers carry 89.6%; the top fifty carry 98.6%. “Thousands of merchants” is a listings number. The business, today, is a few dozen services.
More than a quarter of earning sellers depend on one wallet
Volume leaderboards cannot show you who the volume comes from. Payer-level analysis can. Among the 294 sellers with at least $100 of settled volume, the median seller’s largest buyer accounts for 23.6% of its revenue — healthy enough. But 84 sellers (28.6%) take more than half their revenue from a single wallet, and 31 of them (10.5%) take more than ninety percent. For anyone deciding whether a seller’s traction is durable, this is the decisive number — and it is invisible in every volume ranking.
sellers in basis 294 median top-payer share 23.6% top payer > 50% of revenue 84 (28.6%) top payer > 90% of revenue 31 (10.5%)
One listing in eight is a copy of another
Resellers wrap upstream APIs and list the wrappers — sometimes hundreds at a time. As a crude proxy we counted endpoints whose description text is an exact match with at least nine others: 16,716 endpoints (12.1% of the catalog) fall into 302 duplicate groups. Some duplication is legitimate (multi-chain listings of one service); much of it is the same upstream repackaged. Discovery systems that cannot collapse these will keep recommending the same thing under many names.
The instrument that matters most is the smallest
Everything above measures whether endpoints answer and whether money moves. Neither proves an endpoint delivers. For that we pay endpoints with real USDC and grade what comes back — settlement transaction and response hash published. That program is deliberately small so far: seven paid tests (six delivered, one failed). Seven tests support no ecosystem-level delivery-rate claim, and we make none. We publish the count anyway because paid testing is the only signal in this list that no self-declared metadata can ever substitute for — and because scaling it is the point.
How we know — and how to check us
Every number above is observed by Agents Trust: settlement indexed from the chains, probes we ran, purchases we paid for. Nothing comes from a seller’s self-reported metadata — on the site, sellers’ own claims are labelled as such and never counted. The scoring method is public on the methodology page, the per-seller evidence is on each service page with per-signal provenance tags, and the aggregate reads behind this post are a public API — no key required.
If you run an index, a bazaar, or a gateway: these are the checks you are probably rebuilding privately. We would rather run them once, neutrally, and compare results. Corrections and challenges to any number here are welcome on the public feedback tracker.