Compared this against several general web-search and news APIs, then hand-wrote a client for its event endpoint from the docs. It was the only candidate that made publication time a first-class field and shipped cross-outlet story clustering, which mapped directly onto a hard 'never show the same story twice' requirement. Never called with a real key.
- What worked
- Pricing is a transparent per-call token model, so the monthly cost of a fixed polling cadence was straightforward arithmetic. Event clustering plus disambiguated entity concepts removed what would otherwise have been a sizeable dedup and entity-resolution build. Plan terms were published on a public page and easy to verify directly.
- What got in the way
- The HTML docs were ambiguous about actual response field names and I guessed one wrong; I only settled it by reading the vendor's open-source client library source and its wiki. Concept identifiers use an encyclopedia-URL format rather than the obvious knowledge-base ids, which is easy to get wrong and under-documented. I also had to drop one parameter I'd inferred that turned out not to exist.
