Choosing a Web Search API

Start with the job

Agents that need one verified fact want an answer endpoint. Apps that run their own reading step want links and full text. Chat and voice products want a low-latency lite mode. Research tools that run many steps can use deep modes that synthesize results.

Name a runner-up for each job before you test. Parallel suits excerpt-based agents. Exa suits neural search over links. Brave suits teams that want an independent index. Tavily suits teams that prefer credit-based billing. Also ask who will read the results. A chatbot that shows users cited answers needs different output than a pipeline that feeds another model. Decide the reader first, then the shape of the output.

Test before you commit

Write 50 questions from your product and record the correct source for each one. Run every provider with the same mode, result count and model. Score the answers, the sources, the latency and the cost per correct answer.

Check the freshness settings too. Confirm that a live fetch works, that a cached mode behaves as documented, and what the response does on timeout. Include a few hard queries, such as pages that contradict each other or claims that need a date. A provider that handles them well will save review time later.

Decide, then keep checking

Pick the provider with the lowest cost per correct answer that still meets your latency target. Re-run the test each quarter, because prices, indexes and modes change. Keep your pricing and benchmark notes dated so you can see when a number went stale. Write the decision down with the test results and the date. When a colleague asks why you chose a provider, the record answers the question without a fresh test.

Budget, failure and data terms

Plan for failure. Every provider has outages and rate limits. Set a timeout, a retry limit and a fallback, such as a cached answer or a second provider, for each call path. Check the rate limits in the docs, since a burst of traffic can hit them even when your monthly spend is small.

Also check the data terms: how long queries and results are kept, and whether the provider uses them for training. Read the terms with your legal team before you send private customer questions through any search API.

Checklist

• Name the job: cited answer, raw links, neural search, or chat lookup.
• Shortlist two or three providers that fit that job.
• Write 50 test questions with known sources.
• Measure quality, latency, freshness and cost per correct answer.
• Re-test each quarter and date every price you record.

Questions & answers

Sources for this entry

• Hyperresearch Answers
• Parallel Search API best practices
• Exa search API reference
• Brave: what sets the Brave Search API apart

Vendor descriptions are attributed to the vendor. Prices change; confirm with each vendor.

Continue this reading path

• Selection & fit
• Best Web Search API
• Resource directory

Cite this entry