Match accuracy
Australian address APIs use the same open national address file, so coverage rarely differs. What differs is how well they handle partial and misspelt input. Here is how ours does.
Results by input
Partial input is harder, and the table shows it.
| What was typed | Queries | Right address first | In the top 6 | No result |
|---|---|---|---|---|
| A complete address | 2,000 | 99.7% | 99.8% | 0.0% |
| Without the postcode | 2,000 | 99.7% | 99.8% | 0.0% |
| With the suburb misspelt | 1,974 | 99.7% | 99.8% | 0.0% |
| With the street misspelt | 1,852 | 96.2% | 97.3% | 2.2% |
| Without the suburb or postcode | 2,000 | 98.1% | 99.8% | 0.0% |
| With “unit” before the flat number | 604 | 99.7% | 99.7% | 0.2% |
| With “apt” before the flat number | 604 | 97.8% | 97.8% | 1.8% |
| With the street type part typed | 222 | 86.9% | 96.8% | 1.4% |
| Overall | 11,256 | 98.5% | 99.2% | 0.5% |
Speed
Median 19 ms of server time, 117 ms at the 95th percentile, measured from within Australia. Your own network is on top of that.
Method
Each address is asked in these ways: complete, without the postcode, with a misspelt suburb, with a misspelt street, without the suburb or postcode, with the word unit before the flat number, with the word apt before the flat number and with the street type only part typed. A misspelling is only counted when the word was actually changed. Variants that need a feature (a flat number, a street type longer than four letters) are asked only of the addresses that have it, so each row shows how many addresses it covers. Words of four letters or fewer are not misspelt in this test, so those addresses are left out of the misspelling rows. Mt written for Mount applied to only 28 addresses in this sample, too few to report, so it was not run. A result counts only when it denotes the same physical address as the G-NAF record, matched on unit, number, street, locality and postcode. top1 means it was the first suggestion.
Sampled at random within each state from the raw G-NAF release named above, independently of our own search index. Sample is reproducible from the seed below. Seed 20261023.
The draw is stratified by state, so small states are measured too: ACT 60, NSW 483, NT 60, OT 60, QLD 409, SA 137, TAS 60, VIC 533, WA 198. Small states are over-represented compared with their share of addresses, and the overall figure is not reweighted.
Each row asks one query per address, and the queries column says how many. Rows that need a feature of the address, such as a flat number, cover fewer addresses than the whole sample.
- Latency is server-side and measured from within Australia; it excludes your own network.
- Partial input scores lower than complete input, and the per-variant figures show by how much.
- These are our own measurements of our own service. We publish the method and the sample seed so they can be reproduced.
A result counts only when it is the same physical address as the G-NAF record: the same unit, number, street, suburb and postcode. Close is not counted.
Check it yourself
The numbers on this page are in one file: benchmark.json.
These are our own measurements, and you can reproduce them. The harness is in our repository: it draws its own sample from the G-NAF release you point it at, asks whichever address APIs you configure the same questions, and scores every one of them the same way. Run it against whatever you use today, with your own addresses.
Measured against G-NAF August 2026 on 2026-10-23. G-NAF is republished quarterly and we re-measure with it. Questions about the method: hello@wattleaddr.com.au