Benchmarks · Additional research
Which web search API retrieves the right source most often, and what does it cost?
Retrieval accuracy, latency and price for the search APIs that sit underneath retrieval-augmented systems, alongside a reliability disclosure scorecard covering price stability, supply independence, documentation, compliance and procurement.
Current edition
September 2026
- Providers measured
- 7
- Accuracy spread
- 75.9% – 91.1%
- Reliability dimensions
- 5
- Verified
- 2026-09-03
Headline findings
What the September 2026 edition found.
Measured 2026-09-03 · re-verification due 2026-12-02
Exact-source retrieval ranged from 75.9% to 91.1%. Exa returned the reference URL for 911 of the 1,000 questions; Keenable for 759 of 1,000. Each of the 7 providers was sent the same 1,000 SimpleQA questions; 2 of 7,000 calls failed and are scored as misses, so every rate is over all 1,000.
Providers whose intervals overlap are not distinguishable at this sample size. Grouped by overlapping 95% Wilson intervals: Tier 1, Exa, Brave and You.com; Tier 2, Linkup, Serper and Tavily; Tier 3, Keenable. The groups do not separate cleanly: You.com's interval (86.8–90.7%) overlaps Linkup's (82.4–86.9%).
Quality and speed are close to uncorrelated. Keenable has the lowest mean latency, 404 ms, and ranks last on retrieval; Exa, the best retriever, is fourth of seven on speed at 1,705 ms. Latency is the mean over successful calls in a separate pass with one request in flight.
Published prices ranged from $1.00 to $7.50 per 1,000 searches. Serper is the lowest at $1.00, a low-confidence figure; Tavily the highest at $7.50. These are vendor list prices, verified 2026-09-03, not measurements.
Scored disclosures, not measurements. The reliability scorecard scores 6 vendors 0–4 on 5 dimensions from their published documents: 29 of 30 cells carry a score, and 1 has too little history to score. No vendor scores above 2 of 4 on documentation and commitment. A low score means poor disclosure, not poor practice. Verified 2026-09-03.
Editions
Every run of this benchmark.
| Edition | Measured | Re-verification due | Data |
|---|---|---|---|
| September 2026 | 2026-09-03 | 2026-12-02 | Data |
Custom benchmarks
Run the same test on your own ICP.
Bring your ICP and the data providers you use or are considering. We scope a custom evaluation around your decision and show you the evidence: who each provider finds, who your current source misses, and how accurate the records are.