Edition No. 2

How long job postings stay listed

Every month we read the careers pages of thousands of US employers and ask one question: how long have these jobs been posted? This is our September 2026 reading, taken on the night of 2026-09-26.

We count the days each posting has been listed on the employer's own site. That is all we measure. It does not tell you whether a job is still open, or whether anyone was hired.

What we found

Listed 90 days or longer, bulk-dated not counted: 13.6%
Listed 90 days or longer, every posting counted: 16.6%
Median time listed: 30 days
Postings counted: 492,340

These figures cover the dated postings at the 1,032 employers we hold fixed from one month to the next, so months can be compared. Some employers upload hundreds of postings at once, all with the same date. That date tells you when the upload happened, and not much about each job. Those postings stay in the total. We simply do not count them as long-listed, which is why the lower figure is the careful one.

Last month we published 18.8% at 90 days or longer, across the 1,037 employers that edition covered. Most of the fall since then is a change in how we count. That same night comes to 16.2% under this edition's rules, so the real movement is a small rise.
Two in three postings carry a title that names no level at all: 65.2% of the 1,324,562 titles we read across every employer, not only the panel.
On 6 of the 10 hiring platforms we read, the board publishes no posting date at all. There we can only say a posting is at least this old.

Ages are counted to 2026-09-26, the night we read, never the day this page was built. How we measured all of this is at the foot of the page.

What was read

This edition comes from a single reading of every active posting, frozen to a file and fingerprinted before any figure was computed. Every number here is derived from that one file.

Active postings in the reading: 1,359,503
Panel employers with a counted posting: 1,012
Of the reading counted in the headline figures: 40.6%

Ages run to the night itself. Edition No. 1's figures were built from a view that refreshed after midnight, so its ages ran a day past the date it printed. This edition counts to an explicit date so that cannot recur.

How long postings have been listed

The headline set: the constant 1,032-employer panel, each posting aged from the date its board shows today. The panel is fixed, so a change here is a change in postings and never a change in which employers we look at.

Postings counted: 492,340
Median time listed, constant panel: 30 days
Lower quartile: 11 days
Upper quartile: 65 days
Listed 30 days or longer: 50.8%
Listed 60 days or longer: 27.2%
Listed 90 days or longer, bulk-dated not counted: 13.6%
Postings listed 90 days or longer, bulk-dated not counted: 67,174

These counts are “age is N days or longer”, so a posting listed exactly 30 days is inside the 30-day figure and the wording has to admit it.

The oldest posting is deliberately not printed. A median absorbs a bad date. A maximum is the bad date, because it is a search for the most extreme value in the set. Edition No. 1 published one and it turned out to rest on a date field that records when a posting record was created, not when the job was advertised. An extremum returns here only with the single row behind it named and checked, and re-checked on every refresh.

The full published set

The same rules applied to every employer we publish, not only the panel. This is the wider picture and the baseline later editions compare against. It is the figure that moves when we publish more employers, so it is reported second and every change to that universe is named below.

Postings counted: 551,492
Employers with a counted posting: 1,283
Median time listed, all published employers: 30 days
Listed 30 days or longer: 51.2%
Listed 60 days or longer: 27.6%
Listed 90 days or longer: 16.9%

Changes to which employers are published

At this reading the published set was 15,943 employers carrying 1,324,562 active postings. That set is not fixed, and when it changes the figures above move for a reason that has nothing to do with hiring. Each change is recorded here:

2026-09-23 237 employers already under review were approved for publication: 1,162 to 1,399 employers carrying a posting (+237).
2026-09-24 a bulk load of verified employers was published, 13,483 of them carrying postings: 1,399 to 14,882 employers carrying a posting (+13,483).
2026-09-26 further verified employers were published. The night of 25 September was not read, so this change is measured across two nights instead of one: 14,882 to 15,943 employers carrying a posting (+1,061).

Each figure here is measured from the frozen readings on either side of the change, never read live: a live count keeps moving, and the published set grew again in the days after the last of these. Every reading records the size of the set it was taken over, so two of them cannot be set side by side without the difference being visible. This page refuses to build if a change of this size is not named here. They are never described as movement in the market. The constant panel in section 2 is unaffected by every one of them, which is why it leads.

The distribution

The constant panel, by how long each posting has been listed. Bucket boundaries are Edition No. 1's, so the two editions can be read side by side.

0 to 7 days 15.7% 77,542

8 to 30 days 34.8% 171,266

31 to 60 days 22.9% 112,761

61 to 90 days 10% 49,091

91 to 180 days 10.2% 50,360

181 to 365 days 5% 24,734

over 365 days 1.3% 6,586

Shares are of the 492,340 counted postings in the constant panel, and sum to 100% up to rounding.

What is counted, and what is left out

Every posting in the reading falls into exactly one row below. The ladder is applied in this order, so a posting excluded by an earlier rule is not counted again by a later one.

	Postings	Share
In the reading	1,359,503	100%
Employer not published on the site	34,941
Continuous-hiring title	4,090
Standing or pipeline listing	15,593
No date published by the board	753,387
Counted in the headline figures	551,492	40.6%

A posting the board gives no date for is a floor, not an age, and is never given one. Separately, 856 postings sit on boards that blocked us during the window. They are in the reading and are counted or excluded by the same rules as any other, and are noted here only because a blocked board cannot retire its postings.

Standing listings, by the reason recorded

Reason	Postings
pipeline family	15,591
retail store role	1
implausible date	1

The comparison sets

This edition changes two things at once: the date a posting is aged from, and which employers are in the set. Reporting only the new number would present the two changes as one movement in the market. All four combinations are published.

Set	Postings	Employers	25th	Median	75th
Constant panel, new rules the date the board shows today · the 1,032-employer panel	492,340	1,012	11 days	30 days	65 days
This edition the date the board shows today · every published employer	551,492	1,283	11 days	30 days	66 days
Constant panel, Edition No. 1 rules the first date the board gave · the 1,032-employer panel	514,532	1,032	12 days	31 days	68 days
Every employer, Edition No. 1 rules the first date the board gave · every published employer	576,589	1,318	12 days	32 days	71 days

The first two rows differ only in which employers are in the set, so the gap between them is the universe and not the market. The first and third differ only in which date a posting is aged from, so the gap between those is the rule change.

Set	30 days or longer	60 days or longer	90 days or longer, as measured / bulk-dated not counted	Postings at 90+, as measured / bulk-dated not counted
Constant panel, new rules	50.8%	27.2%	16.6% / 13.6%	81,768 / 67,174
This edition	51.2%	27.6%	16.9% / 13.8%	93,435 / 76,198
Constant panel, Edition No. 1 rules	52.3%	29.3%	19% / 13.9%	97,760 / 71,674
Every employer, Edition No. 1 rules	52.7%	29.8%	19.4% / 14.1%	111,692 / 81,447

How much of the fall since Edition No. 1 is a change of method

Share of postings listed 90 days or longer. Each bar covers the group named under it.

Edition No. 1, as published 18.8%

476,618 postings at 1,037 employers, counted to 2026-08-31

That night, under this edition's rules 16.2%

460,693 postings at 1,016 of those same employers, counted to the same night

This edition 16.6%

492,340 postings at 1,012 of the 1,032 employers in the panel, counted to 2026-09-26

All three figures are as measured. Edition No. 1's figure is frozen as that edition published it and cannot be restated to set bulk-dated postings aside, so a conservative figure for this edition alone would stand against two that are not. The conservative reading of this edition's figure is 13.6%. The first two bars are one night counted two ways, so the distance between them is the change of rules. The third is a later night on the constant panel.

Edition No. 1, as published and recomputed

	Postings	Median	90 days or longer
Edition No. 1 as published, its 1,037 employers, to 2026-08-31	476,618	28 days	18.8%
The same night, same employers, under this edition's rules	460,693	not recomputed	16.2%

Edition No. 1 is not being reissued. The second row is what the same night becomes under this edition's rules. The difference is the rule change, not a change in the market.

Edition No. 1's population, rebuilt from the event log

Rebuilding Edition No. 1's night from the append-only event log gives 476,077 postings and 89,393 listed 90 days or longer. This reading reproduces those figures, which is the check that nothing has rewritten the history this edition's comparisons rest on.

What titles say about level

Read from the words of each posting's title and nothing else. This is the share of postings whose title names a level. It is not the seniority mix of the jobs, and it cannot be: most titles name no level at all.

Postings read: 1,324,562
Title names no level: 65.2%
Numbered level, not comparable: 0.8%
Postings with a numbered level: 9,962

Nearly two thirds of titles name no level. Any figure that spread those across levels would describe how we read titles, not the market, so they stay in their own row. Levels written as numerals, such as “Consultant 1” or “Officer L1”, are not read either, which makes entry a floor and not a count.

Level named in the title	Postings	Share	Largest employer's share of the level	Share without that employer
Intern	28,501	2.2%	12.3%	1.9%
Entry	51,573	3.9%	33.1%	2.6%
Mid	24,681	1.9%	4.2%	1.8%
Senior	95,088	7.2%	1.1%	7.1%
Staff or principal	19,260	1.5%	9.2%	1.3%
Lead	41,410	3.1%	2.6%	3%
Manager	144,792	10.9%	6.6%	10.3%
Director	35,227	2.7%	0.9%	2.6%
Executive	10,696	0.8%	6.7%	0.8%

Every share is of all 1,324,562 postings read, of which 863,372 (65.2%) name no level. The last two columns are there because a single employer's job ladder can carry a level on its own: the largest contributor to entry holds 33.1% of it.

How often a title names no level, by hiring platform

Platform	Postings	Title names no level
Workday	685,300	64.5%
Greenhouse	186,437	57.9%
iCIMS	171,149	76.4%
Oracle	96,032	67.6%
SmartRecruiters	86,620	66.5%
Ashby	41,452	56.1%
BambooHR	21,200	69.2%
Lever	20,834	53.9%
ADP	15,269	70.6%
Workable	269	71.4%

The share varies by more than twenty points between platforms. Part of the level mix above is therefore an artefact of which platforms employers use, not of who they are hiring, and that is why no level figure is published without the unknown share beside it.

Where the postings come from

Which hiring platform served each posting, and what share of each platform's postings carry a date its board publishes.

Platform	Postings	Carrying a board date
Workday	697,222	65.4%
Greenhouse	194,222	12.2%
iCIMS	173,149	52%
Oracle	104,835	0%
SmartRecruiters	86,732	0%
Ashby	42,804	5.9%
BambooHR	22,817	0%
Lever	21,401	0%
ADP	16,052	0%
Workable	269	0%

6 of the 10 platforms supply no posting date at all, so every posting on them is a floor and not an age. They are undated for two different reasons, and the difference matters: most simply publish no date, while one publishes a date this edition declines to use, because it records when a posting record was created, not when the job was advertised. For 12% of its rows that is months earlier. Counting it would age those postings from the wrong day.

Date provenance	Postings	Share
First seen by us, no board date	787,568	57.9%
A date from the board	571,935	42.1%

The completeness check for this night

A night qualifies only if every part of the read reported back and the boards it missed are few and countable. This night passes.

Check	Observed	Threshold	Result
Every part of the read reported back	5	5	pass
Every board we could not read can be named	0	none	pass
Share of boards left unread, %	0.00	5.00	pass
Share of boards whose results failed to save, %	0.00	2.00	pass

4 of 4 gating checks passed. Every night is read and stored whether or not it passes, so the press night can be chosen from records, not from memory. A night that was never read cannot be chosen afterwards.

Of the constant panel, 972 of 1,032 employers were read successfully that night, and 60 were not. A fleet-wide miss rate under the threshold can still hide a single panel employer going stale, so the panel is counted separately from the fleet.

The stale-employer exclusion

An employer with no successful read in the seven nights up to and including this one is excluded from this night's figures, on both sides of the comparison, so the panel stays identical. An employer we did not read cannot retire its postings, so its listings would otherwise age in place and inflate the long-listed share. 11 panel employers are excluded on that rule.

Set	Postings	Employers	90 days or longer
Constant panel, this edition's rules	486,989	1,001	16.4%
Constant panel, Edition No. 1's rules	508,920	1,021	18.8%
Every published employer, this edition's rules	546,141	1,272	16.8%

These figures come from the tool that takes the reading and are passed through, not recomputed here, because the rule reads the crawl record and not the export. That makes them a weaker claim than the figures elsewhere on this page, which were recomputed independently and agreed to the row.

Method notes

The rules this edition is computed under, fixed before the figures were taken. Once this edition publishes, changing any rule below changes the series. Any such change is recorded here. The series is not re-baselined silently.

How we measured this(21 notes)

10.01

A posting's age comes from the date the board shows today

Age is measured from the posted date the job board asserts today. When a board changes a posting's date, we adopt the new one and record the change.

It is the only figure a reader who clicks through can check. If we kept a date the board no longer shows, a reader checking our work would find the board saying something else.

Changed since Edition No. 1: Edition No. 1 was computed under a write-once rule: the first date we ever saw was kept forever. A hand-checked sample of 375 postings, read back against the boards themselves, measured what that cost. Of 188 readable postings open 90 days or longer, 7 carried a board date newer than ours by 95 to 227 days, and none carried an older one. So that edition overstates the age of re-dated postings by an estimated 3.7% of that frame (95% CI 1.8 to 7.5%). Edition No. 1 is not being reissued. The correction is here.

10.02

Cumulative age runs from the first date ever asserted, and a re-post does not reset it

Alongside the board's current date we keep the first date the board ever asserted for a posting. That date is never revised. Any figure about how long an employer has been trying to fill a role across re-posts uses that one.

How old a posting is, and how long the job behind it has been open across re-posts, are different questions. Collapsing them into one column answered neither reliably.

10.03

What counts as bulk-dated, and why every 90-days-or-longer figure is published both ways

A posting is bulk-dated when 50 or more postings from the same employer share one identical posted date. Those postings stay in the total. They are never counted as long-listed, so the conservative figure is a floor and not a smaller population. We publish every 90-days-or-longer figure this edition computes both ways, as measured and excluding bulk-dated postings, on the same line. External use leads with the conservative figure. Two sets of figures on this page are not paired. One is Edition No. 1's figures, quoted as that edition published them, not restated. The other is the three columns that also set aside employers whose board was not read on the night. Those come from the capture tool, not from this edition's own arithmetic. The tool does not compute the conservative figure, and the capture is frozen, so pairing those columns would mean re-taking a night that is already fixed.

Thousands of jobs do not go up on one calendar day and then each stay listed for the same number of days. That is one bulk upload carrying one date. The threshold is 50 because the answer barely depends on it. The share listed 90 days or longer with these postings uncounted is 13.1% at 25, 13.6% at 50 and 14% at 100, so a fourfold change in the threshold moves the published figure by nine tenths of a point. A threshold of 10 would catch ordinary same-day posting. 500 would miss obvious bulk events. At 50 there are 76 clusters, few enough for a reviewer to inspect each one.

10.04

Removed means absent from the board's enumerable listing

A posting is removed when it stops appearing in the board's own listing. A detail page that still answers after delisting is not a live posting.

Boards often keep a job's direct URL serving long after the job leaves the listing. Judging by URL reachability would systematically under-report removals.

10.05

Removal counts carry a measured error rate, and it leads the figure

Every removal figure is published with the measured false-removal rate beside it, conservative reading first. As of the 2026-09-01 audit that rate is 7.6% (95% confidence interval 3.7 to 14.9%), from 100 sampled removals of which 92 gave a clear answer and 7 were still live.

A removal count with no error bar invites the reader to treat it as a census. Applied naively to Edition No. 1's 185,659 removals the rate implies roughly 14,100 postings counted as removed that were not. Across the interval that is 6,900 to 27,700. The interval is wide because n=92.

Changed since Edition No. 1: Edition No. 1 published 185,659 removals and a 21-day median with no accuracy figure attached, because none had been measured. The direction of the effect on that median is now known, and it is a slight understatement. A false removal enters the population with a truncated duration, and six of the seven had a true age exceeding their recorded age-at-removal by 3 to 36 days. Their recorded durations (median 19 days) sit just below the population median of 21, so excluding them moves the median up. The magnitude cannot be estimated from seven observations.

The density used for a bound, 12% of removals within three days of the median, was measured on the live removal window on 2026-09-01 (n=187,178), which reproduced the median of 21. It was not recomputed on this edition's frozen 24 July to 30 August population (n=185,659). The two sets overlap heavily and are not the same set, so the one-to-two-day ceiling is indicative for the frozen population and was not measured on it.

Cause is not the multi-board defect: all seven predate its 2026-08-31 guard, which alone would invite that attribution, but the mechanism needs an employer with more than one board, exactly one such employer exists, and none of the seven is it. Zero attributable. Five of the seven were attributed to limited board enumeration. The other two remain unexplained. Individual audit cases are withheld.

The rate bounds nothing else. A single-board false removal that is later re-found emits a reappearance exactly as a genuine re-post does, so this rate implies a false-reappearance channel that the multi-board check above does not cover, because that check rules out one cause and this is a different one. That channel is not bounded.

2,020 of 10,123 reappearances sit on boards flagged capped, but capped and uncapped boards show the same mean retirement-to-redetection gap of 4 days, so cap status is not a boundary. No reappearance in the record exceeds a 30-day gap and 85.6% are within seven days, which is what flicker-then-redetect produces and not an employer re-posting weeks later. 22,489 disappearances had 31+ days of runway to show a long gap without one appearing. Gaps beyond the 40-day event history are unobservable. So the reappearance population must not be published as re-posting behaviour in this edition.

10.06

Part of the July and August removal data comes from records rewritten during the window

Removal counts and removal-duration figures for 24 July to 30 August 2026 carry this limitation. Where a conservative reading exists, we lead with it.

On 10 and 11 August 2026 the stored records for 223 employers were replaced. The rebuild wrote 109,739 rows carrying history that predates the record itself, in some cases by nineteen days. Those employers hold 23.2% of tracked postings and account for 58.9% of the window's removals, two and a half times their weight. A quarter of the data produced well over half the removals we report.

Changed since Edition No. 1: Edition No. 1 published 185,659 removals and a 21-day median over this window without the limitation, which had not been measured then. The direction of causation is unestablished. The removals may have prompted the rebuilds, or a common cause may have produced both. Either way the limitation holds. We offer no corrected count. The affected removals cannot be separated from ordinary ones, and a number that looks cleaned up without being cleaned up is worse than the warning. The rebuilds left the underlying observation log alone. Pre-rebuild events were written on the days they describe and none was deleted, so the damage is to the accounting. The observations are intact. Two earlier versions of this note gave 272 employers on 08-10 alone, then 570 across five days. Both were wrong. The signature had been measured without excluding a one-day gap, which a walk crossing UTC midnight produces legitimately and in bulk. The measured figures are two days and 223 employers. A nightly tripwire now watches for the same signature.

10.07

A posting with no board date gives a floor, never an age

Where a board gives no machine-readable date we count from the day we first saw the posting. We label that "listed N+ days", never "open N days".

It is a minimum. The posting is at least that old and probably older, and the label has to admit it.

10.08

The reappearance distribution is published as observed, with three limits, and carries no behavioural claim

Reappearance gaps are published as an observed distribution only. No claim is made that they measure employer re-posting behaviour, and no re-posting rate is derived from them.

The distribution is right-censored twice and contaminated once, and none of the three can be corrected for. (1) The window. The event history spans 40 days, so a gap longer than that cannot appear at all. (2) An identity discontinuity, and we have measured how far it reaches. On 10 and 11 August 2026, records for 223 employers were replaced, and a posting whose record is replaced cannot reappear under its former identity. It enters as new instead. 75.2% of the reappearances in this distribution are at those employers and 46.7% are on rebuilt records, so the limit is not marginal. Its reach, however, is bounded and small. The event log was never rewritten. Pre-rebuild events were written on the days they describe and none was deleted, so the observations on the near side of the boundary are genuine observations, not reconstructions, and at most 77 returns were logged as new instead of as reappearances. What the rebuild did produce is 108,908 appearance events that are individually correct and collectively misleading. Each key genuinely was new, and what they describe is a re-keying, not hiring. Those do not enter this distribution. (3) False removals. A posting beyond a board's enumeration cap is retired without leaving, and re-detected when the board's window shifts, emitting a reappearance that no employer caused. 8,813 postings currently sit beyond their board's enumeration.

Changed since Edition No. 1: Edition No. 1 made no reappearance claim, so nothing is retracted. What changed is that a reappearance analysis was scoped for this edition and is now explicitly out of scope as behaviour. The observed shape: of 10,123 postings that reappeared on or before 2026-08-31, the mean gap between retirement and re-detection is 4 days, 85.6% are within seven days, and none exceeds 30. Retirement itself requires three missed observations and cannot complete in under three days. That shape is what detection flicker produces. An employer re-posting weeks later would produce a long tail, and none is present, though limit (1) means a long tail could not have been seen past 40 days anyway. 22,489 retirements did have 31 or more days of subsequent history without producing one.

10.09

Figures are checked against the boards before publication

Before each edition we draw a random sample from the published population and commit it publicly before any of it is checked. A second implementation, sharing no code with the collector, reads the sample back from the boards. We publish the match rate and every miss.

A recorded seed proves only that the draw can be repeated. Seeds can be retried until one flatters the sample. Committing the draw before the answers exist is what makes the rate worth quoting.

10.10

Days the completeness check does not grade as a complete fleet read are listed, each with its cause

A completeness check has graded every day's read since 18 August 2026. Six days are not graded complete, and every figure in this edition that spans them is computed over the days that were observed. No value is interpolated. Two had no fleet read. On 23 August the collector was upgraded and split into shards to stay inside its host's run limits, and the dead day was the transition. On 31 August the shard planner counted boards where the runner counts employers, the first employer with two boards appeared in the plan twice, and every shard was refused. Four more are not confirmed complete. On two of them boards went unread: 24 August (1,120 of 1,124 boards read, the first sharded night) and 26 August (1,123 of 1,124). On the other two every board was read and the run's own record is incomplete: 28 August (across three dispatches) and 1 September (a host outage lost the run records of 26 boards). Days before 18 August are ungraded, because the check did not exist.

A removal is detected on the first observed day after the posting is gone, so a day with no read moves detections later and lengthens the apparent lifetime of anything that came down on it. The effect is small and runs in one direction: a missed day can only lengthen an apparent lifetime, never shorten it. A day not confirmed complete costs the series nothing from the boards that were read. It is listed because the record cannot show it complete, and a series with quiet gaps reads as a quiet market.

10.11

Seniority is reported only as the share of postings whose title names a level

A posting's level is read from the words of its title and nothing else: intern, entry, mid (only where the title says so, as in “II” or “intermediate”), senior, lead, staff or principal as a technical level, manager, director, or executive. A title that names no level is counted as unknown, and never assigned a default such as mid. Levels numbered III and above, or given as a range, are shown as their own row and never folded into a level, because the same number means different things on different employers’ ladders. Every level figure is printed beside the unknown share of the same postings, at the same prominence. Each one also shows the share of that level contributed by its single largest employer, and the level’s share recomputed without that employer. Unknown shares are also published by job-board platform. The figure is never described as the seniority mix of jobs.

About two thirds of titles name no level. Spreading those postings across levels would report the classifier, not the market. The unknown share varies by platform, from roughly half to three in four, so part of the level mix is an artefact of which platforms employers use. A single job ladder can dominate a level. Two limits are stated with their direction. Entry is a floor: levels written as Arabic numerals (“Consultant 1”, “Officer L1”) are not read and count as unknown, so entry is undercounted and the true share is at least the published one. Validation was near-blind, not blind. A reader labelled 300 titles: 100 the classifier called executive, 60 look-alikes it did not, and 140 at random, at most one per employer per group. The reader never saw the classifier’s answers, but the level definitions they worked from were written by the classifier’s author with its rules in mind. On that sample executive was right 99 times in 99. The reader and the classifier agreed on 294 of 300 titles, and on the 140 random titles they agreed on whether a level was named 137 times.

10.12

From 4 October some postings gain a date they did not have, and this edition is captured before then

From 4 October 2026 a second collector supplies employer-published posting dates for Ashby, Greenhouse, iCIMS and Workday postings at the employers it writes. Those postings carried no date before, so they sat outside every age figure. From that date the ones with an employer date join them. We take only the dates an employer states. A board’s “posted N days ago” arithmetic and its censored floors are left out. One platform is excluded altogether, because its date records when a posting was created, not when it was advertised. This edition is captured before 4 October and none of its figures moves.

A later page that counts more dated postings at those employers is measuring a bigger population, not a changed market. The constant panel this edition leads with is unaffected either way. None of its 1,032 employers is written by that collector. We measured that on 27 September against the list of employers the collector writes, and the same query returned an employer known to be on the list, so the zero is not a query that failed to see. The list changes whenever a board moves, so we run the check again before comparing any figure across 4 October.

10.13

The night of 25 September was not captured, and no reading was taken late to fill it

A night's figures come from a single reading of every board, taken after that night's checks and before anything else writes. On 25 September nothing was read. Another process began writing at 04:55 the next morning and was still writing when the reading was due, so the reading was refused. The night is absent from this edition. No figure is estimated for it.

A reading records what is on the boards at the moment it is taken, and labels it with the night. By the time the other process finished, 116,858 postings had been seen for the first time on the 26th. That is 9% of everything listed. A reading taken then and labelled the 25th would have given those postings an age measured to a night before they existed. Every earlier reading contains none of this. The three before it carry zero postings first seen after their night, out of 595,396, 606,711 and 1,085,467.

Changed since Edition No. 1: Edition No. 1 had no such check and no record of the nights it missed. A refused night now leaves a note saying so. The reading is also taken earlier in the evening, an hour after the night's checks instead of eight, so less can happen in between.

10.14

Each edition's population is frozen as a fixed list of postings, because a query that reselects them drifts

At capture we fix an edition's population as an explicit list of the postings in it, stored with a checksum. Every later figure is read from that list. We do not re-derive it by re-running the query that selected it. Edition No. 1's population was pinned this way on 26 September 2026 and has not changed since.

Edition No. 1's population was defined by a query. The columns it read are ones the daily reading of job boards keeps updating: whether a posting carries a date, when we first saw it, whether its title marks it as continuous hiring, and which employer it belongs to. So re-running the query gave a slightly different population every night. It returned 476,625 postings on 23 September and 476,077 on the 26th, about 0.1% lower, with no rule changed and nothing rewritten. A figure that is meant to be frozen cannot be defined by a question asked again later.

Changed since Edition No. 1: Edition No. 1 kept its figures but not the list of postings behind them, so the list had to be rebuilt from the records afterwards. The rebuild moved. Its published figures stand exactly as published. What is pinned is the reconstruction used for comparison, at 476,077 postings. The 0.1% between the two readings is drift, not a correction.

10.15

The constant panel is 1,032 employers: four never had a correct board, and two entries were one employer

The fixed group of employers compared across editions is 1,032. Four come out of both sides of every comparison. The only job board attached to each of them turned out to belong to another company with a similar name, or to be a test board and not a careers site: Strategy, Axon Enterprise, Quanta Services and Aon. Two entries for Zebra Technologies were one employer listed twice, and are now counted once.

Remove them from the later period only and it reads as postings disappearing. That is a claim about hiring, and we do not make it. Leave them in and four employers stay in the group on the strength of boards we no longer believe are theirs. None of the four has any other posting in the group, so nothing else put them there. Counting one employer twice would weight it double.

Changed since Edition No. 1: Edition No. 1 published 1,037. The five entries removed here held 27 postings between them, 0.006% of that edition. The correction on Edition No. 1 says which of its figures move: the counts. No percentage and no percentile changes.

10.16

Both sides of the comparison use corrected employer attribution, so our own fixes are not movement

Where we have since established that a board does not belong to the employer we counted it under, its postings come out of both periods of every comparison, not only the later one. The comparison baseline therefore excludes 27 postings from five boards. Four belong to other companies with similar names. One holds a single test requisition on a board that is not that employer’s careers site. Edition No. 1 stands exactly as published. The check that we can still reproduce it rebuilds its population with the attribution it had on the day it was captured, which is a different question from what we would publish today.

Correcting who a board belongs to changes a count. Applied to the later period alone, the correction looks like postings appearing or disappearing, which is a claim about hiring that we are not making. Size is not the issue. 27 postings is 0.006% of Edition No. 1, and it moves no percentage or percentile that edition printed: the median, both quartiles and all three thresholds are identical either way. The issue is that a reader cannot tell a correction from a movement unless we say which it is.

Changed since Edition No. 1: Edition No. 1 had no mechanism for this. Its population is rebuilt from the event log using today’s attribution, so an identity correction silently moved a figure that was meant to be frozen. That happened on 2026-09-24. The check that rebuilds Edition No. 1 on every capture caught it. No one had seen it by eye.

10.17

Edition No. 1 recomputed under this edition’s rules comes first, so a change of method cannot read as a change in the market

Four sets of figures are printed, in this order. First, Edition No. 1 as published beside Edition No. 1 recomputed under this edition’s rules: same employers, same postings, same date, only the rules differ. Under this edition’s rules the share of postings listed 90 days or longer on 31 August was 16.2%. Edition No. 1 published 18.8%. That gap is method alone, and it comes first because the market had no part in it. Second, Edition No. 1’s rules applied to Edition No. 1’s employers today, set against Edition No. 1 as published. Only time differs, so this is the comparison that describes change. Third, this edition’s rules applied to every employer we now read. That is the baseline later editions compare against. Fourth, the difference between the second and third, split into the part due to the rules and the part due to the change in which employers are measured. We never call that difference movement in the market. Edition No. 1’s published figures do not change. The recomputation is labelled as a recomputation wherever it appears.

Since Edition No. 1 we have changed how some postings are counted (the next three notes) and added employers. Either change moves the headline figures on its own, and the rule changes lower them. Printed only beside the old figure, a lower share of long-listed postings would read as an improvement in hiring when it is a change in counting. We can recompute Edition No. 1 because its population was rebuilt from the records we keep and matched every published figure.

Changed since Edition No. 1: Edition No. 1 kept its figures but no record of which postings produced them. We rebuilt the population afterwards from the day-by-day record of which postings we saw appear and disappear, then checked it against every published figure. It gives the same 1,037 employers, agrees on postings to within 7 of 476,618, and reproduces every published share, overall and by platform. The employer list is now frozen, so the comparison cannot drift as employers join or leave the index. An employer that has since left the index stays on the list, because it was measured then. An employer on the list that no longer posts contributes nothing and is not replaced. Figures under Edition No. 1’s rules deliberately keep values this edition no longer treats as posting dates. They are there so the comparison measures time, and they are labelled that way wherever they appear.

10.18

Employers we could not read in the week before a measurement are left out of it, on both sides of the comparison

An employer is left out of a night’s figures when we have no successful read of any of its job boards on that night or on any of the six nights before it. In the comparison with Edition No. 1 we apply the rule on both dates. An employer that fails it on either date is left out of both, so the two sides still cover the same employers. We print the employers left out, and how many postings each held, beside the figures.

A posting is recorded as removed only when a read of its board finds it gone. When a board cannot be read, its postings stay recorded as open and keep ageing, whether or not they are still up. Fixing the reader just before publication would not help. The first successful read would record every posting that came down during the gap as removed on that one night, the same distortion described in the note on September’s late recording. Leaving the employer out is the honest option. Naming who was left out is what makes the exclusion checkable.

Changed since Edition No. 1: Edition No. 1 applied no such rule. On Edition No. 1’s date it would have left out one employer, Comcast, whose boards had last been read successfully on 21 August, nine days earlier. Its 855 postings were counted as open without having been seen that week. Leaving them out changes nothing Edition No. 1 printed: the share listed 90 days or longer moves from 18.76% to 18.79%, the shares listed 30 and 60 days or longer move by less than a twentieth of a point, and the median stays at 28 days. Every figure is the same at the precision it was published, so Edition No. 1 carries no correction for it. In this edition’s comparison Comcast is left out of both sides, with the other employers the rule excludes.

10.19

Standing families of identical requisitions a year or older are left out of age figures

A posting is left out of age statistics and rankings when it has been listed a year or longer and either (a) or (b) is true. (a) It belongs to a family of five or more near-identical titles at the same employer, and either five or more of that family share its posted date or the family’s median age is itself a year or longer. (b) Ten or more of the same employer’s postings share its posted date. A posting left out this way stays listed on its employer’s own page. The criteria were written down before they were first run.

Some employers keep one requisition per location listed indefinitely: hundreds of copies of the same role, many carrying the same date. Each copy is a real posting with a real date. Together they are one standing pipeline counted hundreds of times, and at the old end of the distribution they outweigh everything else. On 17 September, 67 of the 100 oldest US postings came from a single employer’s standing requisitions.

Changed since Edition No. 1: Edition No. 1 counted them. Applied to the postings Edition No. 1 measured, the rule as it stands today leaves out 13,502 postings at 64 employers, every one of them listed a year or longer. Figures under Edition No. 1’s rules keep them. Figures under this edition’s rules do not.

10.20

Lever’s date records when a posting’s record was created, so Lever postings carry no age

Postings on boards run on the Lever platform are treated as undated. They are listed from the day we first saw them, labelled as a minimum, and they do not enter age statistics or rankings.

Every date source is tested the same way: for postings that appear on a board we have already been reading for at least two weeks, compare the date the board claims with the day we first saw the posting. A publication date lands within a day of first sight for nine postings in ten, because we see a new posting the day it goes up. Workday, iCIMS, Greenhouse and Ashby all pass at one day. On Lever, one posting in ten claims a date 138 days or more before the day we first saw it: the field dates the posting’s record, which can be created years before the job is published.

Changed since Edition No. 1: Edition No. 1 used Lever’s dates for 2,429 postings at 21 employers, of which 60.4% were shown as listed 90 days or longer, the highest of any platform and an artefact of the field. Figures under Edition No. 1’s rules keep those values so the comparison measures time. Figures under this edition’s rules do not. The original values are kept, not deleted. Lever postings first seen after 17 September have no stored value, so figures under Edition No. 1’s rules cannot include them. The number left out is printed beside those figures.

10.21

Two weeks of changes on the largest boards were recorded late, on 18 and 20 September

From about 3 to 17 September we read the largest boards but could not save their changes. The backlog was recorded on 18 September (new postings) and 20 September (removals). No figure in this edition treats either day as a change in the market. We do not quote day-to-day changes across 17 to 20 September. A removal recorded on 19 or 20 September for a posting last seen before 18 September is left out of removal-duration figures and reported separately. It ended at an unknown point between the last sighting and 18 September. We never treat the recording date as its end. We quote levels only from nights the completeness check passes.

The daily count of postings rose by 21,954 on 18 September and fell by 46,264 on 20 September. The fall is exactly the backlog: 58,069 removals, less 10,672 new postings and 1,133 returns, all recorded between those two readings. Four employers account for 52% of it. For one of them the last sighting was 30 August, so its recorded removal date is up to three weeks late. No daily figures exist for 19 or 21 September. The completeness check refused both nights, and a daily figure cannot be taken afterwards.

How to reproduce these figures

The reading is a single file, taken in one read-only snapshot and fingerprinted before anything was computed from it. Anyone holding the file can recompute every figure above and get these numbers.

Reading night: 2026-09-26
Reading sha256: c7a7041feb03b48663d36033ba40e9fa7a0665d4e84e52468bdecab7fc779df9
Generating query sha256: 604c96031ec290394ee4208f7ba53ba4102d2a0d2048a254e24d6176a8caf47d
Reading tool sha256: 70507d0fd8581c6e8b04571e6aece7f67f258fc12d321789dcd09cb080a9239e
Constant panel sha256: 68b6183cad2ad31f35b269fb732eefb934f9e8c011ba526d59ccfb730841b8bd
Reading bytes: 133,131,236

The figures above were recomputed from the reading independently of the tool that took it, and 21 of them were compared against the tool's own numbers row for row before this page was built. Two implementations reading one file and agreeing is what makes the rules above a description of what was actually computed.

How these rules were tested against the job boards is set out in How we checked. The previously published edition is Edition No. 1, whose aggregate figures remain unchanged. Individual historical examples are withheld.