1. Collect real queries
Take 50 to 100 queries from your logs or your users, across the topics and languages you care about. Avoid queries you wrote to make one provider look good.2. Run every query at each effort level
Run each query atlow, medium and high with the same search_window, and
store the top 10 results.
3. Judge relevance
For each result, ask one question: does this article answer the query? Judge one result at a time, with a person or an LLM judge, and check a sample of the LLM verdicts by hand.4. Compare
- Precision at 10: the share of the top 10 results that are relevant.
- Distinct events in the top 10: how many different stories the results cover.
- Freshness: how old the relevant results are.
- Latency and cost per query.

