Extraction Accuracy and Scale Benchmarks

Every performance claim on this site traces to a test you can run yourself. This guide documents the datasets, the budgets and the reference numbers.

1. What we measure

Correctness first, speed second. Each benchmark asserts exact counts, then logs wall-clock time for humans to compare:

• Extract 100k unique — every address found exactly once, zero lost.
• Dedupe 150k with 50k case duplicates — totals, uniques and removed counts must reconcile exactly.
• Convert 20k-row CSV to JSON — header detection plus full fidelity.
• Analyze, validate and domain-extract 20k — provider tallies, verdict counts and domain tables verified row by row.

2. Reference numbers

Measured September 16, 2026 on Apple Silicon under Node 22. Your browser will differ — the point is the order of magnitude, not the exact digit:

Scenario	Result
Build 100k-line input	5 ms
Parse 100k unique addresses	52 ms
Parse 150k lines, collapse 50k duplicates	70 ms
Convert 20k-row CSV to JSON	33 ms
Analyze 20k list	13 ms
Validate 20k list	19 ms
Domain extraction over 20k	< 1 ms

The engine is regex plus hash maps — throughput scales linearly, and the browser Web Worker keeps the tab responsive while it runs.

3. Reproduce it

npx vitest run tests/performance.test.ts

The same file also serves as the CI scale gate: generous timeouts with exact-count asserts, so slow hardware stays green while regressions in correctness fail loudly.

4. Accuracy methodology

• Fixtures pair every feature with its edge cases: obfuscated forms, plus-tags, Gmail dot variants, multi-label suffixes (.co.uk), quoted CSV fields, BOM markers and malformed rows.
• Export round-trip tests prove TXT, CSV, TSV and JSON exports re-import to the identical address set.
• Real OLE and ZIP fixtures are constructed byte-by-byte in-test, so format parsers are verified against structure, not luck.
• End-to-end browser tests replay the full user path — including a zero-third-party-request guarantee for default processing.

5. What we do not claim

Throughput is not deliverability. A fast extractor can still hand you role accounts, stale addresses and typos that were already wrong in the source — which is why the Email Validator and the list-hygiene guide exist. Speed gets you a list quickly; validation decides whether you may send to it.