Trial counts are the wrong metric. Every marketplace reports them because they look like demand, but at the endpoint level, the number of free samples and the number of paid calls barely correlate. I pulled the per-endpoint numbers from our live marketplace and found the two most-sampled endpoints — over 4,700 trials each — have been paid for zero times. Meanwhile the endpoints that actually earn money are mostly the boring ones nobody leads a pitch with.
This matters because the whole point of a pay-per-call marketplace is the ledger, not the demo counter. Here is the data, endpoint by endpoint.
Sort every endpoint by how many free trials it has consumed, and the top of the list looks like a healthy business:
| Endpoint | Category | Trials | Paid calls |
|---|---|---|---|
| Gas | crypto | 4,715 | 0 |
| Captcha Solve | specialized | 4,626 | 0 |
| Recall | ai | 3,867 | 36 |
| Time | utility | 1,652 | 40 |
| Find | ai | 1,629 | 5 |
| Polymarket | finance | 1,240 | 0 |
The top two rows are pure samples: thousands of free calls, not a single paid one. The trial counter inflates; the ledger doesn't move. Anyone reading "4,700 trials" as "4,700 units of demand" is reading a number that has nothing to do with money.
Sort by real paid calls instead, and the picture inverts. The earners are a small, quiet cluster:
| Endpoint | Category | Paid calls | Trials |
|---|---|---|---|
| Time | utility | 40 | 1,652 |
| Recall | ai | 36 | 3,867 |
| Token Security | security | 14 | 186 |
| Domain Intel | web | 10 | 60 |
| Wallet | identity | 8 | 63 |
| Crypto Price | crypto | 8 | 35 |
| Smart Contract Audit | contract-audit | 4 | 0 (no trial) |
The paid cluster is utility + security/audit + web/domain intel. And notice the last row: the audit endpoint offers no free trial at all, yet it is in the top tier of earners. The only endpoint in the list that cannot be sampled for free is one of the few that actually gets paid.
The split makes sense once you see it as two different behaviors:
So the free samples pile up on the trivially-replaceable endpoints, and the money goes to the hard-to-replace ones. Trial volume is inversely correlated with paid conversion — not because trials are bad, but because what gets sampled freely and what is worth paying for are almost disjoint sets.
The honest metric is paid calls per endpoint, not trials. A design that hands out unlimited free sampling produces a flattering trial counter and an empty ledger. That is the opposite of what a pay-per-call marketplace is supposed to measure.
If the goal is real machine-to-machine commerce, the trial system should be tight enough that the trial counter means something. A small, shared allowance gets an agent to the moment of "do I pay for this?" without letting the demo counter masquerade as revenue. The data here is the argument: the endpoints that convert are the ones where the free sample ends and a real job begins.
We are acting on this. Starting September 1, the free-credit grant goes away and the trial allowance tightens — the endpoint-level numbers say those giveaways were converting to almost nothing while making the marketplace look busier than it was.