Testing location data with edge cases, not just happy paths
An address in a well mapped city center tells you almost nothing about how your system handles a rural route, a disputed border, or a query near the poles. Test the hard cases deliberately.
A geocoder that has never been checked against a known point is just guessing with confidence. The only honest way to measure accuracy is to start with addresses whose real coordinates are already known, from surveyed benchmarks, official parcel records, or verified building locations, and compare those against what the API returns.
The comparison itself is simple. You take the returned latitude and longitude, measure the distance to the known true point, and log it. Do this across a large and varied set of addresses, not just the ones near a head office, and patterns appear quickly. A geocoder might land within a few meters for a suburban house number and be off by a city block for a rural route address, and both of those results are true at once.
What matters more than any single distance figure is how the error correlates with the precision field in the response. A result marked house should cluster tightly around the true point. A result marked street should be farther off on average, because it was matched to a road segment rather than a specific building. A result marked postcode or city should be understood as a much coarser estimate from the start. If the error does not track the precision label, the label is not doing its job.
This is also why coverage and accuracy have to be measured separately. Coverage asks whether an address can be found at all. Accuracy asks how close the answer is once it is found. A country can have excellent coverage and mediocre accuracy, or the reverse, and reporting only one number hides which problem you actually have.
We do not publish a single blended accuracy percentage for My Geocode, because a number like that flattens exactly the variation that matters. An address in a country with detailed building level data and one in a country where postal authorities only track localities are not comparable, and averaging them together produces a figure that describes neither. What we can tell you is how to run this kind of test yourself: pick reference addresses relevant to where your users actually are, check the returned precision and confidence together, and set your own thresholds based on what you see for your footprint.
If you are evaluating any geocoding provider, not just ours, ask for reference-point testing rather than a headline percentage. A vendor who is comfortable with their data will show you the method. Try a batch of your own addresses against the forward geocoding endpoint and look at how precision and distance move together for your specific mix of locations, since that is the number that actually predicts how the API will behave in your product.