guide · Evidence checked 2026-09-02
A cheap search call can produce an expensive answer
A practical, evidence-led decision guide. Product capabilities and limits are separated from anything that would require hands-on testing.
Price per thousand searches is rarely the final cost. One accepted answer can require query rewriting, multiple searches, page extraction, model tokens, retries and a human source check.
Use this monthly model:
`total = search calls + extracted pages + input tokens + output tokens + reranking + retries + storage/egress + reviewer labour`
Then divide by accepted answers, not attempted queries.
Example: 100,000 user questions at 1.4 searches each produce 140,000 search calls. At $5 per 1,000, retrieval is $700. If 30% require five extracted pages at $1 per 1,000 pages, extraction adds $150. Those two known layers total $850 before model tokens, retries and review. The example is arithmetic, not a quote; insert the current rates for You.com, Brave Search, Kagi, Tavily or Perplexity.
Measure p50 and p95 calls per task, cache-hit rate and rejection rate. A provider with a higher unit price can be cheaper if it reduces failed answers; a cheap endpoint can be expensive when weak coverage triggers repeated searches.
Primary action: export one week's call trace, replace every estimate with an observed distribution, and recalculate after human validation.