TrusteeSaleData collects notices of trustee’s sale published in newspapers of record across 15 US deed-of-trust states, extracts structured fields from each notice, and discards any value that cannot be found verbatim in the source text.
Data as of
Every record begins as a legal notice published in a newspaper of record. We collect those notices through Column, the public-notice platform used by a growing number of US state press associations, and we keep the notice text exactly as published.
We do not filter on the publisher’s category label. Notice types are applied inconsistently from state to state: in a representative 30-day pull, 632 of 2,368 genuine trustee’s sales — more than a quarter — were filed under some other type, including “Legal Notice” and “Pending Categorization”. Trusting the label loses a quarter of the corpus, so we match against the body of the notice instead.
| Stage | When | What it does |
|---|---|---|
| Scrape | Daily, 09:00 UTC | Pulls a rolling three-day window of new notices. The overlap is deliberate — papers publish late. |
| Parse | Daily, 09:10 UTC | Extracts structured fields from the day’s new notices, then drains the older backlog newest-first. |
| Rollup | After each parse | Recomputes every count on the site and re-evaluates which pages are published. |
| Backup | Daily, 09:30 UTC | Captures notices and derived records together, so a restore is internally consistent. |
A notice published today is generally searchable here within a day. The newest notice we hold was published September 1, 2026.
Every extracted value is searched for in the source notice. Anything that cannot be located there is dropped rather than published.
Extraction is done by a language model, which is good at reading legal prose and bad at admitting when it has guessed. So the model’s output is not trusted: after each notice is parsed, a verification pass looks for every extracted value in the original text and discards the ones it cannot find. Each record carries the resulting quality score.
This is why a blank field here means the notice was silent. Many notices genuinely omit the parcel number, the opening bid or the unpaid balance. Where that happens we show nothing, rather than substituting a nearby number that happens to look like the right kind of figure.
Two fields are never taken from the model at all. The state and county on every record come from the publication’s own metadata, because a notice’s text frequently names several jurisdictions — the property’s, the trustee’s, and the recording office’s — and picking between them is exactly the kind of judgement that produces confident errors.
Formatting is decided by rules, not by the model. Dates, money, phone numbers and street suffixes are normalised deterministically, because formatting is the thing a model does inconsistently across thousands of rows.
We hold notices from 155 counties. We do not publish 155 county pages.
A county page is worth a reader’s time when it has enough upcoming sales to describe a market rather than list two addresses. So a county enters the index at eight upcoming sales and leaves it only when it drops below five. Today 5 of 155 counties clear that bar.
The gap between eight and five is deliberate. Trustee sale inventory is a rolling window — sales are held and leave the count every week — so a county sitting on a single threshold would enter and leave our sitemap every few days. A page must move meaningfully before its status changes.
Counties below the threshold still have pages. They render, they carry their real numbers, and they are linked from their state page. They are simply not submitted to search engines, and they carry a noindex directive until they qualify. Promotion is automatic and requires no republishing.
Fifteen states, not fifty. We cover deed-of-trust states, where foreclosure runs through a trustee and produces a published notice: Arizona, Nebraska, California, Utah, Missouri, Idaho, Montana, Virginia, Alaska, West Virginia, Tennessee, Washington, Nevada, North Carolina, Oregon. Judicial-foreclosure states produce court filings instead of trustee’s notices, and are out of scope.
Not every newspaper. Coverage begins with papers that publish through Column. A notice published only in a paper outside that network will not appear here, and we have no way to know how many those are in any given county.
A parse backlog. 2,234 notices we hold have not yet been parsed into structured records. They are processed newest-first, so the backlog is older material, and every page states its own parsed share rather than implying completeness.
No property data beyond the notice. Notices do not state bedrooms, square footage, condition or market value, so neither do we. Nothing on this site is enriched from a third-party valuation source.
Sales change after publication. A scheduled sale can be postponed, rescheduled or cancelled at any time, sometimes hours before the auction. We record postponements where a later notice states them, but the authoritative answer on the day is the trustee’s sale line, not this site.
TrusteeSaleData is not a government agency, not a county recorder, not a trustee, and not a real-estate brokerage. We are not affiliated with any trustee firm, lender or servicer named in a notice, and naming one is not an endorsement or a claim of any relationship. Nothing here is legal or financial advice.
Questions about the data, or a correction: how to reach us.