Austin's construction moved a kilometre south. New houses moved furthest.

The short version

Between 2018 and 2025 the geographic centre of Austin’s building permits moved 1.23 kilometres south.

That is not a large distance. Austin’s permits are spread across roughly ten kilometres in each direction, so a kilometre is about an eighth of the city’s own spread — a drift, not a migration. We are not going to tell you the centre of construction shifted.

The interesting part is which permits moved.

New construction moved 2.6 times further than repair work. That is what a city expanding at its edge looks like from the permit record: new houses go where there is land, and everything else follows the buildings that already exist.

Why this is a finding and not an artifact

The obvious way to be wrong here is to measure your own data instead of the city. Three things could produce a moving centre with nothing actually moving, and each was checked before the claim was made.

Was the centre pulled by a changing mix of permit types? A city that issues more downtown alterations and fewer edge subdivisions has a centre that drifts without a single building changing location. So the drift was measured inside each permit type separately — and every one of them moved south. The 1.23 km is not composition.

Was the corpus growing? More data each year, especially newly-added outer areas, moves a centre on its own. Austin’s permit volume runs between 52,000 and 71,000 a year across the window with no step change, and the drift is measured against that flat base. Two candidate cities were excluded for failing this: Philadelphia’s record jumps 3.4× between 2018 and 2019, and San Antonio’s begins in 2025.

Were the coordinates real? This is the one that would have been invisible. When an address cannot be matched, some systems drop the permit on a default point — a city hall, a centroid — which manufactures a dense cluster that looks exactly like real concentration. Across 20.0 million permits there are zero such placeholder coordinates and 84 out-of-bounds points. 94.8% of the geocoding sits at a confidence of 0.95 or higher.

What this does not tell you

• It is one city. We measured four and can defend one. Boston and Minneapolis have the corpus for an overall drift but 92 and 430 new-house permits respectively in 2018, and 108 and 148 in 2025 — enough to compute a centre, nowhere near enough to trust one. We are not going to publish a number we would not act on.
• A permit is not a building. Some are never built, some are amended, and issuance timing varies.
• A centre of gravity is a summary, and summaries hide things. Austin’s drift is consistent across permit types, which is what makes it believable, but a single point cannot tell you whether growth is one large development or a thousand small ones.
• Nothing here is per-capita. We hold permits, not population. Every figure is a count of permits, not a rate.

A candidate we removed, and why we are telling you

Columbus, Ohio was the strongest candidate in the country by our own selection test. Its permit volume varies by only 9% across eight years — the flattest record of the 105 jurisdictions we hold, and flatness is exactly what a claim like this needs.

It was flat because it was broken.

In 2019 Columbus moved the word “new” from one field of its permit records to another. Our classifier read the old field, so from 2020 the city appears to issue no new houses at all — while 8,173 of them were quietly filed under alterations. The permits were never lost; they were mislabelled.

And that is why it screened so well. The error moved permits from one category to a neighbouring one, so the yearly total never wavered. The measure we used to pick a trustworthy city was the measure the fault preserved — Columbus did not merely survive the screen, it topped it, because a record whose errors are internally conserved looks steadier than an honest one.

We found it while choosing cities for this report. It is fixed, the correction is published, and Columbus is excluded here because the fix landed after these figures were measured.

Method

Source. 20.0 million building permits from 105 US and Canadian jurisdictions, as published by each city. Austin’s figures come from 498,866 permits carrying coordinates at confidence ≥ 0.90, 2018–2025.

Dating. Each permit is dated by its issue date, or its application date where the city publishes no issue date — applied identically to every series here.

Position. Cities publish coordinates directly or we geocode the address. Only permits at confidence ≥ 0.90 are included. The “centre” is the arithmetic mean of latitude and longitude for a city-year; the “spread” is one standard deviation of the same, converted to kilometres at that latitude.

On significance, and why we do not quote it. With 30,000–65,000 permits per city-year, a centre is measurable to about 40 metres. Every drift in this report is therefore many standard errors from zero — including Minneapolis’s 0.22 km, which we regard as no movement at all. Statistical significance is not the question at these sample sizes; size relative to the city is. Austin’s 1.23 km is 12.7% of one standard deviation. That is the number we would defend.

Reproducibility. Every figure in this piece appears in a figure block held with the source, including the ones we rejected and why.