Sources and methodology

What counts as an item

An item is one headline fetched from a named outlet’s public feed, stored once by slug with its publication time in UTC, its summary, and its source link. The corpus begins in late May of this year. Items that fail the relevance gate are kept in the database with a suppression reason but never displayed and never counted; every query on the site filters them out first.

The suppression gate

Feeds carry residential, consumer-finance and general-business stories alongside commercial real estate. A rules-based gate scores each headline and summary for commercial real estate anchors and marks off-topic rows as suppressed. The gate is conservative in one direction: a borderline story is more likely to be kept than a relevant one dropped, which is why the reports show an explicit “no transaction cue” row rather than a cleaner sample.

Market and sector inference

Market and sector tags are inferred from the headline and summary by regular expression, first match wins, using the same lists that build the markets and sectors hubs. A headline that matches nothing is left untagged and shown as such. Tags are never back-filled by a model, so an untagged row is a limit of the rules, not a judgement about the story. Some tags are known to be broad: the Washington market, for instance, currently collects Washington state headlines as well as the District. Because the tagger also reads the excerpt, a story can be tagged to a firm’s home city rather than the property’s. The monthly market reports therefore drop any row whose headline names a different state, market or well-known city than the one it is tagged to, and print how many rows that removed.

Deal extraction

A pure, synchronous extractor (src/lib/deal-extract.mjs) reads each headline and returns a deal type, a USD value, a square footage, a unit count and an acreage where each is stated. Deal type is decided by precedence: a leasing cue makes it a lease; a money figure within a few words of loan, financing, refinancing or mortgage makes it a financing; a sale, acquisition or purchase cue makes it a sale; a financing cue anywhere in the headline with no sale cue also makes it a financing; groundbreaking, construction or topping-out language makes it a development; any other transaction cue is “other”; and a headline with no cue is “none” and yields no value.

A value is withheld when the headline states more than one distinct money figure (the amount is ambiguous), when the currency is not USD (a letter before the dollar sign, or a euro, pound, CAD or AUD marker), when the figure is a per-unit, per-key, per-month or per-square-foot rate, when the headline has no commercial real estate anchor, or when the amount exceeds the plausible ceiling for a single transaction. A bare dollar figure with no transaction cue is never treated as a deal value.

De-duplication

Outlets often report the same transaction within days of one another. In the priced-sales tables, two rows in the same market whose values fall within one percent and whose publication times fall within seven days are treated as one deal: the row with the larger stated value (the earlier one on a tie) is shown and the others are listed beneath it as “also reported by”, each still linking to its own story. Item counts elsewhere are not de-duplicated, because they measure coverage, not transactions.

Reporting threshold

A monthly market report is offered to search engines only when the month holds at least 60 items and 8 disclosed-price sales. Months below the threshold are still published with a visible notice, so the archive is complete, but the sample is too thin to read as a signal. Reports are produced for New York, San Francisco, Chicago, Los Angeles and Washington, DC.

No model in the numeric path

No figure on this site is produced by a language model. Every number on a report page is the output of one named query in src/lib/figures.ts, and every extracted figure is displayed beside the headline it came from, linking to that headline on this site and to the original story at its source. Where the site carries written commentary, it is written by a person, contains no numbers, and is labelled as such.

What is not published

Summed deal volume, per-firm deal credit, market share, cap rates, price-per-square-foot medians and year-over-year comparisons are not published. The corpus is a record of what named outlets chose to report, not a census of transactions, and those measures would present a sample as if it were the market. Firm mentions are shown as counts of headlines naming the firm, with no role attached.