From leasing websites to rental data you can compare.

The expensive way to do this is to point a language model at every leasing page and pay for the reading. The cheap way is to notice that most communities do not have a bespoke website — they run on a leasing platform, and each platform serves its floor plans from an endpoint that returns the same shape every time. So the pipeline tries 19 deterministic adapters first, one per platform, and only falls through to the model when a site is genuinely bespoke. That tail is about 29% of runs — 13,456 reads over the last two months — and it goes out as one overnight batch at half price rather than one request at a time.

The adapters are also the accuracy story, not just the cost one. An adapter reading a JSON endpoint either returns the floor plan or fails loudly; it cannot hallucinate a two-bedroom. The model only ever sees the sites where there was nothing structured to read, and its output goes through the same validation every adapter's does.

Markets do not all run weekly. A metro's cadence is a column in a rotation table — weekly, fortnightly, or monthly — and every staleness check derives its window from that same column, so a fortnightly market cannot be reported stale for being fortnightly. Daily assertions check the things that would otherwise rot quietly: coverage rates per market, snapshot freshness, geocode gaps, duplicate communities arriving under two names.

Two different numbers, and conflating them is the easiest way to overstate a database like this. Tracked means the community is in the universe with a confirmed leasing URL. Live plans means the last scrape actually came back with floor plans and rents. A metro seeded last month has the first number and is still working on the second — Dallas–Fort Worth is mid-first-pass as I write this, which is why its two columns are so far apart.

A rent comp that ignores concessions is the most common way a rent assumption goes wrong, and it is wrong in the direction that makes a deal look better. Six weeks free on a twelve-month lease is not a marketing flourish; it is an eleven percent haircut to the rent you will actually collect. So the scraper reads the concession banner as well as the plan grid, parses it into weeks, and stores an asking rent and a net-effective rent side by side. 32,460 of 66,124 live plans are carrying one right now.

The spread between those two numbers is the part you cannot get from a listings site, and it varies enormously by metro. That variation is the whole point: in the metro at the top of this list, seven in ten communities are buying occupancy, and that is telling you something a median asking rent never will.

Here is the honest limit of this thing. Fleet-wide collection only reached full coverage in August, so there is no rent index here and I am not going to draw one. What the history can support is direction of travel: of the 7,481 communities seen at least twice, at least two weeks apart, over an average gap of 40 days, 47% cut a plan rent and 42% raised one. Ask me again in six months and this section becomes a chart.

Counts are abstract, so here is what the rows actually look like — a stable random draw of up to 20 communities per metro, weighted to include the build-for-rent tail, which is small and is the part I care most about. Only communities that have been read and carry a priced floor plan are drawn. Average rent is the mean advertised asking rent across a community's floor plans at its last read, using the low end of each plan's range.

Unit counts come from public records: the county property record, a Florida apartment licence, or the City of Austin's building permits. Year built comes from the same records where one exists, otherwise from third-party market data. Hover a figure to see which. A dash in Units means no public record covers that community yet; build-for-rent homes sit on many separate lots and rarely have one.