Sources & raw data · Video 04
Everyone says companies post jobs they never intend to fill. That is a motive, and motives cannot be measured. Here is what can: how long an advert stays online, and whether it still leads anywhere. Method, control group, the guess we got wrong, and a correction to our own previous video.
Taken 2 September 2026. One job board served us 1,980 listings tagged “Entry Level” by the employer who posted them, all presented as live. We followed the link on 482 of them, in two independent draws, capped so no single employer could carry the result.
| Advert posted | Listings on the board | Tested | Still open | Wilson 95% |
|---|---|---|---|---|
| 0–30 days ago | 101 | 87 | 100% | 95.8–100 |
| 31–90 days | 23 | 36 | 100% | 90.4–100 |
| 91–180 days | 26 | 43 | 100% | 91.8–100 |
| 181–365 days | 251 | 46 | 34.8% | 22.7–49.2 |
| 366–540 days | 1,230 | 135 | 11.9% | 7.4–18.4 |
| over 540 days | 349 | 135 | 16.3% | 11.0–23.4 |
Everything under six months opened — every single one, in both draws. Past a year, roughly one in eight still works.
The first fourteen old adverts we tested all failed. Fourteen out of fourteen. That is exactly the kind of result that should stop you rather than excite you, because it has a boring explanation that would invalidate everything: if the site had changed its URL scheme, every page would be dead, including the new ones.
So the young listings are the control. 166 recent adverts, 166 opened normally. The links work. It is the old adverts that are gone.
On 1 September we recorded 1,000 listings. The next day we asked the feed about each one, individually. Two had been removed. Two out of a thousand, in twenty-four hours.
That is a very small number of events, and one day is not a trend. But the order of magnitude matches the first measurement: at that pace, clearing half a board takes about a year — which is the median age we found (385 days per listing, 450 per employer).
| Google Jobs | The Muse | |
|---|---|---|
| Listings | 444 | 1,980 |
| Distinct employers | 306 | 62 |
| Median age | 12 days | 385 days |
| Oldest advert seen | 31 days | 601 days |
| Older than a month | 0.8% | 94.9% |
| Older than a year | 0% | 79.7% |
Not one advert older than thirty-one days in the entire Google sample. That is not a different market — it is a freshness window, applied silently and documented nowhere. One board shows you a moving job market, the other shows you a warehouse. Neither is the truth.
Widened sample: 985 listings, 575 employers, 20 job titles written by hand, from warehouse work to nursing to engineering. They arrive through 95 different platforms, 81 of which serve fewer than ten listings each.
| Platform of origin | Listings | Share | Show a salary | Carry a date |
|---|---|---|---|---|
| 177 | 18.0% | 27.1% | 73.4% | |
| Talents By Vaia | 167 | 17.0% | 34.1% | 1.2% |
| BeBee | 129 | 13.1% | 17.8% | 81.4% |
| ZipRecruiter | 68 | 6.9% | 48.5% | 32.4% |
| College Recruiter | 65 | 6.6% | 0% | 100% |
| Indeed | 47 | 4.8% | 85.1% | 57.4% |
| JobLeads | 38 | 3.9% | 100% | 100% |
| Glassdoor | 2 | 0.2% | — | — |
The search suggestions people type are “ghost jobs indeed” and “ghost jobs linkedin”. Indeed is under five percent of this sample, and Glassdoor is two listings out of nine hundred and eighty-five. The second largest supplier is a name most people have never heard.
Our previous video set this aside on suspicion: 26% of listings showed a salary on one source and 99% on another, which looked like it measured the platforms rather than the market. Now it is tested. We shuffled the labels a thousand times and compared the observed spread with what chance produces.
| What we explain | By the platform | By the profession |
|---|---|---|
| Shows a salary | V = 0.546 (p = 0.001) | V = 0.211 (p = 0.003) |
| Carries a date | V = 0.730 (p = 0.001) | V = 0.274 (p = 0.001) |
The platform explains it about two and a half times better than the job does. JobLeads shows a salary on 38 listings out of 38 [91–100]. College Recruiter on 0 out of 65 [0–6]. Same country, same day, same job titles. The national average of 30.4% describes neither, and describes nobody.
We expected to find the same job repeated across many platforms, so that one opening looks like five to a candidate. It is a good story and it would have suited our argument.
It is not there. Of 985 listings, 882 are distinct, and two appear on more than one platform. Google Jobs deduplicates, and does it well.
The other correction against us: 40.6% of listings carry no date at all, and they are not scattered — 99% of one large platform's listings have none, and that platform alone accounts for half of all undated listings. So a missing date is a platform's habit rather than a sign of age. Our twelve-day median is a median of what is dated.
Checking this sample made us re-check the previous one, and it had the same weakness. Those 1,000 listings came from 53 employers, one of which — Walmart — posted 37.7% of them, the same advert repeated store by store.
| Sample | State a number of years |
|---|---|
| As published in video 03 | 5.3% |
| Deduplicated by ad text | 5.5% |
| Deduplicated and excluding Walmart | 9.7% |
The conclusion of that video holds in every split — median one year, maximum two to three, none asking for five. And the instability it already reported now has a cause we did not know while filming: which employers happen to be in the window. The full correction is on that video's page.
Thirty-six of the adverts we tested, six per age band, one per employer, in the order we drew them. Click any of them. The recent ones open; most of the old ones return a page that no longer exists — which is the whole finding, and it takes about a minute to disprove if we are wrong.
| Age band | Days old | When we tested | Advert |
|---|---|---|---|
| 0–30 d | 0 | open | listing |
| 0–30 d | 4 | open | listing |
| 0–30 d | 4 | open | listing |
| 0–30 d | 8 | open | listing |
| 0–30 d | 14 | open | listing |
| 0–30 d | 27 | open | listing |
| 31–90 d | 36 | open | listing |
| 31–90 d | 36 | open | listing |
| 31–90 d | 38 | open | listing |
| 31–90 d | 43 | open | listing |
| 31–90 d | 59 | open | listing |
| 31–90 d | 67 | open | listing |
| 91–180 d | 128 | open | listing |
| 91–180 d | 137 | open | listing |
| 91–180 d | 168 | open | listing |
| 91–180 d | 168 | open | listing |
| 91–180 d | 168 | open | listing |
| 91–180 d | 168 | open | listing |
| 181–365 d | 278 | open | listing |
| 181–365 d | 283 | dead | listing |
| 181–365 d | 291 | dead | listing |
| 181–365 d | 331 | dead | listing |
| 181–365 d | 331 | dead | listing |
| 181–365 d | 354 | open | listing |
| 366–540 d | 386 | dead | listing |
| 366–540 d | 412 | dead | listing |
| 366–540 d | 449 | dead | listing |
| 366–540 d | 456 | dead | listing |
| 366–540 d | 501 | open | listing |
| 366–540 d | 507 | dead | listing |
| 540+ d | 541 | dead | listing |
| 540+ d | 542 | dead | listing |
| 540+ d | 544 | open | listing |
| 540+ d | 547 | dead | listing |
| 540+ d | 548 | dead | listing |
| 540+ d | 580 | dead | listing |
Statuses recorded 2 September 2026. An advert can be taken down or restored after that date — if one disagrees with our column today, that is new information, not an error in the table.
Job data comes from The Muse public listings feed, where the employer tags the level, and from Google Jobs results for the platform comparison.
We do not republish the ad texts. The Muse terms prohibit cloning their content, and a frozen copy would only let you re-read ours. We publish the scripts instead: run them and you get your own sample, from the day you run it, and you can contradict us with fresh data. That is a stronger form of the same promise.
recherche/etape3_ghost.py
(instantane, survie, pages)recherche/etape3_plateformes.pyrecherche/13-ghost-jobs.mdIt does not say employers are lying. We measured how long adverts stay online. A dead link does not tell you why it died — filled, cancelled and expired look identical from outside. Nothing here establishes intent, and intent is what the word “ghost job” usually implies.
It does not say the job market is dead. We measured what one platform serves to a candidate. Sliding from that to the state of employment is exactly the move we criticise.
One board, one country, and 2,000 of 90,000. The feed claims 90,057 listings and lets you page through 2,000 — page 100 returns an error, and that limit is documented nowhere. We do not get to choose which 2,000 we see. Sixty-two employers, one of them two thirds of the sample, which is why the age figures are also weighted per employer.
The platform comparison is a single day. A day is not a year, and a freshness window could be changed tomorrow.
The file is downloadable. The 302 listings we opened one by one, with each age and status, are on the raw data page — free, no address needed.
One email when a new measurement is published, with the raw file attached to it. No course, no newsletter, no advice — the data, and what we found in it.
We send one email asking you to confirm; ignore it and the address is never used. One click to leave, on every email. Stored by Brevo in the EU, never sold, never shared. What we do with it.