The question
How long do job adverts stay live, how often do they disappear, and how often does the same role reappear under a new identifier?
Note what this does not ask. It does not ask whether employers are advertising in good faith. Motive is not observable from a job board, and no figure produced by this study can speak to it.
The sample
Thirty-nine employer job boards across three applicant tracking systems: twenty-seven on Greenhouse, seven on Ashby, five on Lever.
Selection rule. Candidate employers were well-known US-headquartered technology companies known to publish through one of those three systems. A board was included if its public interface returned at least one live posting on 27 August 2026. The roster was written to a file and frozen that day; no board has been added or removed since.
This is a convenience sample of 39 US technology employers who publish through Greenhouse, Lever or Ashby. It is not representative of hiring generally, and says nothing about smaller employers, other sectors, or companies on Workday, Taleo or iCIMS.
How the snapshot is taken
Once a day, on a schedule, we call the public job-board interfaces that Greenhouse, Lever and Ashby publish for aggregators to consume. One request per board per day, no authentication, no page scraping, and an identifying user agent so any employer can see who is reading.
For every posting we record the system, the board, the posting’s stable identifier, its title, its location, the publish date the board reports, its link, a normalised title-and-location fingerprint, the first day we saw it, and the day it stopped appearing.
Guard against false disappearances.If a board fails to answer on a given day, none of its postings are marked as absent. Only boards that answered can contribute a disappearance, and every failure is recorded in that day’s file.
Reappearance. If a posting returns after dropping out, the date is appended to its record rather than treated as a new posting. A repost under a new identifier is detected by fingerprint when the data is analysed, not when it is collected.
The observation window
Day 0 is 27 August 2026. Forward observation runs to 26 September 2026, with checkpoint reads at day 7, day 14 and day 30. Publication happens after day 30, not before, and that date was fixed in advance so a result cannot be released early because it looks interesting.
Thirty days is a threshold, not a proof. A duration figure also needs a snapshot on every day of the window and a board that answered on each of them. Where either is missing, the figure stays withheld and the page says so.
What it cannot show
- A posting that stops appearing has stopped appearing, and nothing more. We cannot see the reason: it may have been paused, withdrawn, rewritten, moved to another board, or given to someone. Postings in this study have dropped out of a feed and come back days later, so absence is not closure.
- If a board fails for several consecutive days and a posting disappeared during that gap, the recorded date is the first day the board answered again, not the day it actually went. Board outages are recorded, so affected windows can be excluded.
- A posting’s age is the publish date its board reports. Systems differ in how they set that date, so ages are not strictly comparable between one system and another.
- A long-running posting is not evidence of anything by itself. Some roles are genuinely always open.
Every figure here passed a set of automated checks before this page was built: the sample is large enough, the counts reconcile, the dates are usable, and the snapshot is recent. Those checks block known failures. They cannot prove a figure is correct, and they do not audit the job boards we read.
Rules for anything published
- Report posting behaviour, never employer motive.
- Never average the existing public estimates of this problem. They measure different things and are not interchangeable.
- State the sample limitation in the same breath as any figure.
- Publish the collector and the record, so the numbers can be checked.
- Date every figure. This is a dated observation, not a constant.
- An always-open requisition is not evidence of bad faith. Say so wherever a reader might otherwise infer it.
These rules are enforced in code, not by memory. Every published figure passes a set of automated checks that refuse a page rather than let it state something the record cannot support.