How accurate is Australian address autocomplete?

17 July 2026 · 5 min read

Every address API says it’s accurate. The honest way to talk about accuracy is to say how you measured it, on what, and where it falls down. Here’s how we do it for WattleAddr.

The method

We take a random sample of real G-NAF addresses, turn each back into the kind of text a person would type — complete addresses, addresses with a mistyped suburb, addresses with no postcode, unit addresses — and score whether the correct record comes back as the top suggestion, and whether it’s anywhere in the first handful. The method and results live in our public repository so the numbers can be checked, not just asserted.

The numbers

• A complete Australian address returns the correct record as the top suggestion about 94% of the time, and within the top six about 98% of the time.
• Partial and misspelt input scores lower — which is exactly where the interesting engineering is.

The weak spot: a mistyped suburb

The hardest real-world case is a misspelt suburb, because a mistyped final token can otherwise wipe out the whole result set. We handle it with a fallback that peels off a trailing token and retries, so “23 Wakool Ave Deerpark” still returns the street even when the suburb is wrong. That turned misspelt-suburb queries from an empty dropdown into a useful shortlist. Misspelt streets are harder still, and honestly, we don’t fix all of them yet.

Why we publish the weak spots

Because a benchmark that only reports the good number isn’t a benchmark, it’s marketing. If you’re evaluating an address API, ask for the method and the failure cases, not just a headline percentage. Ours are in the open.