Ask any agent-payable marketplace how many of its listings actually work — how many would answer if your agent tried to pay one right now — and the honest answer is almost always a shrug. A directory can tell you what a service claims to do and what it claims to cost, but a listing is a promise made at index time, and the gap between a listed endpoint and a reachable one is where agents lose money. We run UseTested, a receipts-anchored review layer for the x402 agent economy, and this week we want to show you the number almost nobody publishes: the live state of our own catalogue, in full, including the parts that flatter nobody.
The uncomfortable math. As of this morning we index 19,671 services. Of those, 6,129 have ever passed a liveness probe and carry a LIVE status; 129 are confirmed DOWN — our probe reached for them and got nothing back; and 13,413 have never been probed at all. That last number is the honest one most directories never print. We haven't got to them yet. It is not a verdict, and we refuse to render it as one: an unprobed service reads UNTESTED on our site, in grey, the visual equal of "nobody knows" — because that's exactly what it means.
Liveness decays, so the label has to decay with it. A probe result is definitive on its day and evidence ever after — but freshness is a property with a half-life, so we made the label carry the decay instead of hiding it. A service checked within the last 24 hours reads ALIVE. Between one and fourteen days it reads LIKELY ALIVE — still evidence, fading. After fourteen days it lapses back to UNTESTED, because a proof that old is history, not news. Right now that decay vocabulary splits our verified set into 1,858 ALIVE and 4,270 LIKELY ALIVE, plus the 129 confirmed dead — and those aren't static badges, they're read-time semantics: the same row in the same database reads differently depending on when you look, which is the only honest way to render a fact that rots.
How the sweep works — and what it finds. Liveness comes from a two-tier sweep. Tier one re-probes roughly the best-measured thousand services in the catalogue every single day — the ones agents actually buy, where a dead endpoint costs real money within hours, not weeks. Tier two cycles through the long tail less often, which is why most of the 4,270 LIKELY ALIVEs sit in the one-to-fourteen-day band rather than at the fresh edge. The daily numbers are quietly brutal: over the last five sweep days we ran 943 to 1,041 checks a day, and on every single day somewhere between 1.75% and 2.76% of the probed cohort was DOWN. Not the same services every day — endpoints come back, deploys finish, bad hours pass — but on any given morning, about one in fifty of our most-watched services doesn't answer. That is the failure rate hiding inside every "listed" status quo, and the reason a directory without liveness is a directory that's wrong by default.
What the dead cost, and who's allowed to say so. A dead endpoint isn't an academic problem in a pay-per-call economy: an agent that pays into an unreachable service can burn the call price learning what a two-second probe would have shown for free. That's why liveness isn't decoration in our rankings — it's a gate. Anything our sweep marked DOWN, and anything with no liveness evidence at all, is excluded from the top-10 category lists and the paid bundle outright; a service's reliability score is weighted by how fresh its proof is. And the rules of the game are deliberately narrow: only our probe can mark a service DOWN — no one can pulse a competitor off the board — while a provider's self-report (a receipt-anchored Pulse call at half a cent) can only ever claim liveness for their own service, and our sweep overrides it. Seventeen providers currently keep their own proof fresh through Pulse; everyone else is on our clock or nobody's.
The honest footnotes. The 129 DOWN verdicts are "unreachable at probe time", not "scam" — a service mid-deploy can be dead for an hour and fine forever, and the daily churn proves the set is fluid. LIKELY ALIVE includes evidence up to fourteen days old; a thirteen-day-old proof is honest but thin, and the label fading grey day by day is the site saying so. The 13,413 unprobed services are a statement about our sweep capacity, not about the services — some of them are excellent, and the moment tier two reaches one, its badge will start telling the truth. And the sweep itself is evidence, not omniscience: a probe that catches a bad hour can be wrong, which is exactly why DOWN verdicts are re-checked rather than permanent, and why a provider with a Pulse receipt can set the record straight between sweeps.
The wider snapshot (2026-11-07, remote D1): 19,671 services indexed across 132 categories; 1,857 carry a graded review — 36 BUY, 1,249 CAUTION, 425 AVOID, 147 INTERIM — behind 2,209 individually receipted calls, every one a settled on-chain transaction. Liveness: 1,858 ALIVE, 4,270 LIKELY ALIVE, 129 DEAD, 13,413 UNTESTED, with 523 review rows refreshed by the deep-sweep writer on 2026-11-04 alone. Freshness is a measurable, published property here — not a marketing word. The catalogue that can tell you what's dead is also the catalogue that can tell you what's alive, and this morning that's 1,858 answers, on the record, at usetested.com/services.