3,575 trial wallets, one call each

Everyone building a pay-per-call API has the same question about their free tier: do the trials turn into payments? It is usually answered with a percentage quoted from a dashboard, without the shape underneath it. We have the database, so we measured the shape.

The population is every signed-wallet trial identity on this marketplace — 3,575 of them, 5,561 trial calls. A trial is five calls across the whole catalog per wallet, drawn with an EIP-191 signature; there is no registration. The measurement is read-only SQL against the live database.

1. What the allowance actually gets used for

calls used      wallets
1               2,480   (69.4%)
5 (exhausted)     511   (14.3%)
0                 426   (11.9%)
2                  94
4                  29
3                  26
>5                  9   (per-endpoint quota era)

Two thirds of the cohort uses one call and never comes back. The obvious reading is that the first call went badly. It did not.

2. The one-call cohort is not a failure cohort

Joining each wallet to its first delivery record, the one-call group looks like this:

first call outcome    wallets
ok                    2,140
refused (caller)         30
refused                   1

98.6% of those wallets got a successful first call — a real answer from a real service — and then stopped. Whatever ends the journey for them happens after the product worked, not before.

Which service they called first makes it sharper. Of the 2,480 one-call wallets, 292 started on x402-time, 267 on x402-mime-type and 102 on x402-gas — 661 wallets (26.7%) on the three endpoints our own quick-start and trial-signature examples name. That is the shape of someone executing the documented example once: the plumbing works, the platform is real, and nothing in the example asks for a second call. The remaining ~1,800 are spread across the long tail, which is the same story told less sharply.

The group with zero calls is different, and mostly not demand at all: 247 failed and 52 caller-refused. The largest rows are missing-parameter 400s from a bulk probe campaign we ran in September — x402-mime-type "ext or filename required", x402-24game-solve "need exactly 4 numbers", x402-dns-record "domain required". Those are our own probes, not external callers.

3. Paid usage and trial usage are almost disjoint

Of the 3,575 trial wallets, five ever made a paid call. All five are in the exhausted cohort, and they are the only ones: zero of the 2,629 wallets that stopped at one to four calls ever paid, and zero of the 426 that never called at all.

Meanwhile the paid ledger holds 720 paid rows across 34 agent_id values (one of them blank). Twenty-nine of those identities never appear as a signed trial identity at all. Whatever brought the paying population in, for most of them it was not the trial funnel.

Small numbers, fragile ratios. Five payers is five payers; the 0.96% rate inside the exhausted cohort is not a forecast, it is what the rows say today. The claim that survives the small sample is the shape: everyone who paid had exhausted the free allowance first, and nobody who left early ever paid.

4. Where the paid volume came from

Pulling IP and User-Agent for the paying wallets is more useful than counting them:

• A directory that pays to verify its listings. One wallet made 25 paid calls, UA Market402PaidProbe/0.3, and 24 of them landed on one endpoint (x402-format-bytes). This is an external agent that pays real USDC to check that a listing is truthful, and it is the single largest source of external paid traffic we can attribute.
• A private integration test. Two different wallets, one IP, one UA (node), six rows, all successful. Same host, same client, two identities — that is one operator testing, not two customers.
• An unattributed address. The largest payer account (308 paid rows, ~1.55 USDC total, so individually sub-cent) is mostly a single IP with no User-Agent at all. We cannot say who runs it, so we do not count it as verified demand.

5. What we changed, and what we would tell another operator

1. Stop reading "trial conversion" as one number. Split it into first-call success and second-call return. Our first-call success rate is ~98.6%, and 669 of the 3,575 wallets (18.7%; 21.2% of those that called at all) ever made a second call. The first number is a product claim; the second is the one that tracks revenue.
2. If nobody outside the exhausted cohort pays, the free allowance is not the problem. Cutting it, extending it, or rewriting the upgrade copy does not touch a cohort that is not there. The work is making the value legible before the fifth call, or accepting that trials are an evaluation surface rather than an acquisition funnel.
3. Paid verification directories are a real channel. A listing that a directory checks by actually paying produces paid calls and a settlement record, which is also what keeps a listing fresh in staged catalogs. That is worth more per unit of effort than a listing that only gets crawled.

Reproduce

-- distribution (case-folded: a raw string key splits some wallets in two)
WITH w AS (SELECT lower(ident) id, SUM(count) used
           FROM trials WHERE kind='wallet' GROUP BY lower(ident))
SELECT used, COUNT(*) FROM w GROUP BY used ORDER BY used;

-- first delivery outcome per wallet, bucketed by trial usage
WITH w AS (SELECT lower(ident) id, SUM(count) used
           FROM trials WHERE kind='wallet' GROUP BY lower(ident)),
firsts AS (SELECT lower(agent_id) id, outcome, cause,
             ROW_NUMBER() OVER (PARTITION BY lower(agent_id)
                                ORDER BY created_at, id) rn
           FROM delivery_outcomes WHERE lower(agent_id) LIKE '0x%')
SELECT CASE WHEN w.used>=5 THEN '5+' ELSE CAST(CAST(w.used AS INT) AS TEXT) END bucket,
       f.outcome, f.cause, COUNT(*)
FROM firsts f JOIN w ON w.id=f.id WHERE f.rn=1
GROUP BY bucket, f.outcome, f.cause ORDER BY bucket;

Caveat on the population: a "wallet identity" is self-reported and case-insensitive. One agent, one script sweeping several endpoints, and one discovery bot that signs once all look like separate wallets here. The 3,575 figure is an upper bound on distinct trial users, not a count of distinct agents.