Trial counts are the wrong metric. Every API 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 a live pay-per-call marketplace for AI agents 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.
Trial volume ≠ demand
Sort every endpoint by free trials 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. Reading "4,700 trials" as "4,700 units of demand" is reading a number that has nothing to do with money.
What actually gets paid
Sort by real paid calls instead, and the picture inverts:
| 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 the last row is the interesting one: the audit endpoint offers no free trial at all, yet it's in the top tier of earners.
Sampling vs. buying are different behaviors
The split makes sense once you see them as two different behaviors:
- Sampling is free and reflexive. An agent will trial gas prices or a captcha solver because it can, not because it needs to pay for them — both are trivial to self-serve or find free elsewhere.
- Buying only happens when the endpoint does a job the agent can't do for itself: verify a token, audit a contract, resolve domain intelligence.
So free samples pile up on the trivially-replaceable endpoints, and 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.
What this means for trial design
If the goal is real usage (not demo theater), the trial allowance should be tight enough that the trial counter means something: a small shared allowance gets a caller to the moment of "do I pay for this?" without letting the demo counter masquerade as revenue.
The endpoints that convert are the ones where the free sample ends and a real job begins. Everything else is just traffic.
Data from the live service catalog of minia2a.uk, pulled August 25, 2026. "Paid calls" counts real pay-per-call usage, not the free trial allowance.
