Random x402 sellers delivered 74 to 89% of the time. Picked ones: 46 of 46
An AI agent that needs a bitcoin price, a weather forecast or an IP lookup can now pay any of tens of thousands of x402 APIs per call. Most of them answer a 402 correctly. That is not the same as delivering. We measure the gap every week by paying a random sample of endpoints our Trust Index rates "go". And since 23 October we run Fizzl Buy, which picks the seller for an agent instead of leaving it to chance. Here is what both show.
Random sellers: a quarter to a tenth don't deliver
Every Sunday our spot-buy job pays up to 30 random endpoints that the x402 Trust Index rates "go": they answer a valid 402 that a standard client can pay, on Base, at most $0.05 each. Then it records what came back.
| Sample | Bought | Delivered | What went wrong |
|---|---|---|---|
| 18 October 2026 | 27 | 74.1% | the 402 again after paying, settlement failures, 404 and 400 errors |
| 25 October 2026 | 26 | 88.5% | one seller took the payment and sent no answer |
A valid 402 says a seller can be paid. It doesn't say the seller's server will hand over an answer afterwards, or that the answer is what was asked. An agent that picks the first match in a catalog takes that risk on every call.
Picked sellers: what a week of Fizzl Buy shows
Fizzl Buy takes a request in plain words ("BTC price", "weather in Amsterdam tomorrow") for $0.02 and buys the answer from an x402 seller. From 23 to 30 October 2026 it made 46 paid purchases, 42 of them for outside agents, mostly paying in RLUSD on the XRP Ledger, a few in USDC on Base and on Stellar.
- All 46 delivered. Every purchase came back with an answer, and an automatic check confirmed the answer matched the request.
- Seven needed a retry. When the first seller failed, Fizzl Buy paid the next one in line, up to three. The agent paid once.
- Eight requests found no seller that delivered (or the catalog search timed out). Those agents were charged nothing: the x402 payment only settles when an answer goes back.
- 26 different sellers were paid, at a median of $0.001 and never more than $0.01.
The sample is small and the requests were mostly cheap data lookups, and most of them came from one agent fleet. It still shows the difference picking makes: no paid request went unanswered.
What picking means
Fizzl Buy doesn't take the catalog's first hit. It takes the seller our daily Trust Index has seen paying out day after day, at a fair price:
- A payout record, not a listing. Candidates come from the x402 Trust Index, which checks every paid endpoint it can find each day. A seller with 8 payable days out of 8 ranks above one listed yesterday.
- A price cap. Up to $0.01 per seller, so the agent's $0.02 covers a retry.
- Relevance first, then record, then price. Equally good matches go to the cheaper seller.
- A seller that just took a payment without a usable answer moves to the back of the line for a while.
- The answer is checked against the request, and for instructions aimed at the agent (prompt injection).
- No delivery, no charge, and a signed receipt for every purchase.
The free pick shows the same choice without buying: the seller, its record, its price, an example input and where that input goes (URL query for a GET, JSON body otherwise), plus alternatives.
What to look for in a service that buys for your agent
- Does it choose by what sellers actually delivered, or by what they list?
- When the first seller fails, does it try another, and who pays for that?
- Is the agent charged when nothing is delivered?
- Is the answer checked before it reaches the agent, including for injected instructions?
- Can the agent's owner see a receipt per purchase?
How we checked
- Spot-buy samples: up to 30 random endpoints rated "go" in the x402 Trust Index of that day, one per host, paid in USDC on Base with a dedicated wallet, at most $0.05 each; results published on the x402-doctor spot-data branch.
- Fizzl Buy: every paid purchase (routes
buyandx402_buy) in our usage log from 23 to 30 October 2026, with the result each one logged: delivered, the number of sellers tried, the seller's price and the answer check. Four of the 46 were our own and test payments; the rest came from outside agents.
We don't name the sellers: a failed delivery in one sample doesn't make a seller bad.
Questions
Why does an x402 endpoint with a valid 402 sometimes not deliver?
A valid 402 only proves the seller can be paid. After payment the seller's server still has to produce the answer; it can fail, time out, answer 402 again or return an error. In our weekly samples 11 to 26% of random "go" endpoints didn't deliver.
Does my agent pay when Fizzl Buy finds no seller that delivers?
No. The x402 payment only settles when an answer goes back. In the first week eight requests found no seller that delivered, and none of those agents was charged.
How does Fizzl Buy pick a seller?
By relevance to the request, then by the seller's payout record in the x402 Trust Index, then by price, with sellers up to $0.01. It moves sellers that just failed to the back, tries the next one when a seller doesn't deliver, and checks the answer against the request.
Let an agent buy with Fizzl Buy (free pick, $0.02 per purchase, over HTTP or MCP), check a seller yourself with x402 Doctor, and see the daily record in the x402 Trust Index.