How the number is built

Start with everybody in the age range and the sexes you picked. Cut that down by the share of households that clear your income bar. Cut it again for anything else you’ve added. What’s left is the count you see.

The map and the address report run that same definition through two different engines, which is a good way to end up with two different answers. So a test compares them and fails if they ever disagree. You should never be looking at a map that says one thing and a report that says another.

The census publishes age and sex together, so when you filter on those, the number is a straight count. It doesn’t publish education or homeownership broken out by age and sex at neighborhood level, so those get applied as a share instead. That assumes the trait sits evenly across the age range, and every filter using it is labeled so you can see which numbers rest on the assumption.

The income bar sits in that second category too, because it’s a household measure applied to a count of people, which is true of every model in this category and printed on the face of ours.

Zoomed in, the map draws real census neighborhoods and the math is exact. Zoomed out, it has to lump them together before applying your income bar, and lumping changes the answer. At state-and-county zoom the average gap is about 2%, and the worst county we’ve measured is off by 39%. The map says so on the map when you’re looking at one of those views.

The address report is exact no matter what the map is doing, because it’s calculated fresh across every neighborhood the circle touches. It never reads the shapes drawn on screen.

A hexagon usually straddles several neighborhoods, so we have to decide how much of each one belongs to it. We split by where people live, not by how much ground each one covers. That sounds like a detail and it isn’t: splitting by area puts about a third of the population in the wrong hexagon, with the typical cell off by more than 75%. It makes a map that looks right and is wrong.

The finest hexagons only exist where enough people live to support them. A rural county draws at a coarser size instead of pretending to a precision the data doesn’t have.

We hold 16,160,293 places across the US, counted on 3 September 2026. What counts as a competitor gets decided when you ask, not when the data was built, so adding a category for a new line of work takes a config change instead of a rebuild.

Counting is per category and per place. A location that shows up twice in the same category counts once. Anything flagged permanently closed is dropped.

The honest limit: these are pins in an open dataset, not a verified list of businesses. Some have closed without anyone updating them, some are listed twice, some are filed under the wrong thing, and some are missing. IV hydration is undercounted worse than most. Read the number as a floor, not a census.

One thing we engineer against: if we have no coverage of an area, you get told so. A zero you could read as an empty market is the worst failure this data has.

We compared our three-mile population against a major vendor’s number for 30 stores in one client’s markets. That is the only external check we have. One client, thirty stores, and not a general guarantee of accuracy.

The two tracked each other closely, which is the number most vendors would put on a slide. The more useful finding is that we came in low: 4.4% under on average, 5.2% under at the median, with 26 of the 30 stores below. If you are comparing our number against one you already have, expect ours to run slightly light.

We quote it that way because the flattering version of this stat is a correlation, and a correlation would have hidden the fact that we are systematically low.