Geocoding
Address to coordinate, and back
AvailableForward and reverse for all fifty states, with a confidence field that tells you when we are not sure. Try it below — the box talks to the live index, not a canned example.
Try it
This calls the production geocoder. Autocomplete is on — start typing and pick a suggestion, or switch it off to see what the geocoder makes of a half-typed string on its own. The last example has no commas on purpose, to show what the parser does with an ambiguous state code.
And backwards
Give it a point and it returns the nearest addresses, with the distance to each.
88.3% of real coordinates return an address, at 0.95 storage reads each. Every provider here offers reverse geocoding and none differentiates on it. It is table stakes, so it is the same price as forward rather than a separate SKU.
What is behind it
123M
US addresses
1,853,763
street names indexed
26,134
cities
28 ms
typical response
Coverage here is national. Property measurements are not.
Geocoding, reverse and typeahead are national. Property measurements are not — those are Illinois today and expand by state. The two products in this stack have different footprints and it would be easy to assume otherwise. The coverage page is about the property side.
How accurate is it
Measured against 1,995 addresses, reported by outcome. These are not interchangeable — a rooftop point and a guess interpolated along a street are different products to anyone integrating against them.
| Outcome | Share | What it means |
|---|---|---|
| Exact — a rooftop point | 76.19% | Median error 3.6 ft. p90 102 ft, p99 709 ft. |
| Interpolated along the street | 4.71% | Worked out from Census TIGER street ranges. The only tier available in Hawaii and New Hampshire, which have no address points we can use. |
| Nearest known address | 2.86% | The closest address we do hold, labelled as such. |
| Street level only | 0.5% | The street was found, the house number was not. |
| Refused as ambiguous | 2.21% | Candidates returned, no coordinate chosen. Deliberate: picking anyway was correct up to eight candidates and wrong every time past twenty-five. |
| No match | 13.53% | We had nothing to offer. |
76.19% land on a rooftop. That is the number worth quoting, and it is also an average that describes nobody — read the next section before you use it.
Where it is weaker
The national figure is the unweighted mean of 7 equal-sized geographies. Coverage swings 24.2 points across them.
| Geography | Rooftop | No match | Median error |
|---|---|---|---|
| Urban | 87% | 5.6% | 3.9 ft |
| Major urban | 85.3% | 8.8% | 4.6 ft |
| Small urban | 78.9% | 11.2% | 4.6 ft |
| Rural fringe | 78.6% | 11.6% | 3.3 ft |
| Rural | 72.6% | 15.8% | 3.3 ft |
| Consolidated city-county | 68.1% | 14.4% | 15.7 ft |
| Remote rural | 62.8% | 27.4% | 3.3 ft |
Coverage moves 24.2 points; precision barely moves at all. The median error is 3.3 ft–15.7 ftthroughout. Where we are weak we refuse rather than answer imprecisely, which is why our coverage number reads lower than a system that always returns something.
The confidence field is the part worth integrating against
Our largest errors sit almost entirely in the not-confirmed bucket, by design. That is the thing to integrate against: read the confidence field and our worst case drops by a factor of six. A caller who trusts it sees a worst case of 8.4 mi; ignore it and take every answer and the worst case is 49.8 mi. Of 1,520 rooftop answers, 13 are more than 820 ft out. Each of those thirteen was re-checked against the US Census geocoder — an independent third-party reference, not a field survey. It put four of the errors with us, four with the reference itself, and was inconclusive or silent on five.
What we will not claim
That we are more accurate than Google or Mapbox.
We measured it and the test was biased: our answer key comes from OpenStreetMap, and so does Mapbox's data. Judged by a neutral third party, Mapbox's apparent lead over Google falls from 1.51× to 1.14× and reverses on part of the sample. The referee's own median disagreement with our key is 177 ft — wider than most of the gaps being argued about. Anyone quoting a provider ranking from a test like this is quoting the test.
That the addresses we miss are the hard ones.
On the addresses we refuse or fail, Google and Mapbox answer 100% of them, in every geography. They are not hard addresses. That is our gap, and it is why the free tier is a hundred thousand calls — we would rather you find it than be told it is not there.
That the typeahead is cheap for us to run.
It reads storage on nearly every keystroke. Giving it away at the free tier is a pricing decision we can afford, not a property of how it is built.
One comparison that holds without anyone's answer key
Every provider's largest errors sit overwhelmingly on answers that provider labelled coarse itself — Google 25 of its 30 worst, Mapbox 12 of 12, Smarty 20 of 20. Read the precision field, whoever you buy from.
You keep what you geocode
A coordinate, and the address at it, are facts about the world. US law recognises no copyright in facts, so keeping them needs nobody's permission — ours included. That is why storage is not a tier feature here: there is no version of this product where we could take it back.
Google's Maps Platform terms require cached geocoding results to be deleted within 30 days. Ours are yours permanently, so you never re-buy an answer you already have.
One restriction follows the data regardless: the address corpus may not be used to build or sell a mailing list. That comes from the USDOT National Address Database and applies to use rather than redistribution, so storing the data does not lift it. It is set out in section 6 of the terms.
Pricing
One call is one unit, forward or reverse, in any mix.
| Tier | Monthly | Included | Beyond | Storage |
|---|---|---|---|---|
| Free | $0 | 100,000 | requests stop | Permanent |
| Growth | $99 | 2,000,000 | $0.25 / 1k | Permanent |
| Scale | $499 | 15,000,000 | $0.15 / 1k | Permanent |
| Enterprise | Contact us | Unlimited | — | Permanent |
On Free there is no overage: requests stop at the allowance and return an error rather than billing a card we do not have.
These come back with the geocode, at no extra call
- Census tract and block group
- School district
- State and federal legislative district
- FEMA flood zone
- Time zone
- County and place FIPS
Most providers bill each append as its own lookup, so four fields on a million addresses becomes five million billable calls. These are one call.
Completing an address as someone types
The typeahead is a separate product on its own meter, because a session is not a request. See the typeahead page, which has its own live demo and its own accuracy numbers — it is noticeably weaker outside cities and says so.
Other products
Property profile
One address in, every measurement we hold for that house.
Property search
The same measurements, asked of a whole population at once.
Aerial imagery
30cm imagery and spectral overlays, as tiles on your own map.
LiDAR
What a laser measures that a photograph cannot.
Mapping
An OpenStreetMap basemap with no egress fees.
Typeahead (autocomplete)
The dropdown that completes an address as someone types.