Seven checkboxes, and the one score the Ghost Job Detector will not give you
Key measurements
| Measure | Value | Source and date |
|---|---|---|
| Signals in the test | 7, each worth exactly 1 point | Sep 7, 2026 · ghost-job-signals.ts, the module the server page and the client scorer both import |
| Verdict bands and their cutoffs | 3 bands. 0 through 2 apply normally, 3 through 4 apply with caution, 5 through 7 skip it | Sep 7, 2026 · ghost-job-signals.ts, the module the server page and the client scorer both import |
| Lowest score the interface can produce | 1, because the reveal button stays disabled until at least one signal is checked | Sep 7, 2026 · GhostScorer.tsx, the client component the detector renders |
| Analytics events the tool fires | 2, one on reveal and one on copy, each carrying only the score | Sep 7, 2026 · GhostScorer.tsx, the client component the detector renders |
| Score addresses the page will accept | 8, every whole number from 0 through 7, and anything else falls back to the plain page | Sep 7, 2026 · page.tsx, the server page that parses the score parameter |
There is a posting you have seen before. Same title, same company, same three paragraphs about a fast moving team.
The date says posted two days ago. You are fairly sure you read it in March.
That posting is what the Ghost Job Detector was built for. Seven checkboxes and a verdict. You read the listing, you toggle the signals you can see in it, and it tells you to apply, to apply with caution, or to skip it.
No model. No fetcher. No field where you paste a URL.
What does it feel like to use?
Four boxes, about a minute. Age. Repost cycling. No named human. No salary band.
Press the button and the count animates up to four out of seven in amber, under a warning triangle that pulses. Apply with caution. Cap your time and pair the application with outreach to a named human.
Tick a fifth and the mood changes. The circle holds still, the icon becomes an envelope, and the line under it reads: it does not matter how perfect the role sounds, the hour goes to a referral.
That is the entire product. A minute in, and you have your evening back.
Where do the seven checks live?
In one file with no client directive at the top, which is the whole trick. The server page and the scorer import the same arrays, so the eyebrow that says seven signals is counting the list the buttons render.
The seven, in order.
- Posting age over 30 days.
- Repost cycling, where a role is deleted and relisted so the date reads fresh.
- No salary band, in a state whose posting law covers that employer.
- No careers page match, which takes sixty seconds to check and on its own still scores as apply normally.
- No named human. No recruiter attached, no hiring manager activity, applications into a generic portal.
- Evergreen copy. We are always looking for talented engineers is pipeline language, not vacancy language.
- Zero response history, meaning you applied here before and heard nothing back.
That last one has a hole in it. If you have never applied to this company, you cannot answer it, so you leave it off. The count cannot tell a signal you cleared from one you could not see.
The labels came from the full ghost jobs breakdown, word for word. A test reads that post and fails if any label has stopped appearing in it. It catches a rename. It does not check that the explanations still agree.
Two surfaces making the same claim in slightly different words is how a product starts lying by accident.
Why is every signal worth one point?
The math is a reduce. Each checked signal is one point. Zero to two says apply normally, three to four says caution, five to seven says skip it.
No weights. That was the argument I had with myself for a week.
A missing band where the law requires one is stronger evidence than a posting being a month old. Weighting it would have been a few lines of code.
I did not do it, and not because weights are dishonest. Disclosed weights are perfectly inspectable. It is that I could not justify any particular number, so I would have been picking by feel and presenting the result as arithmetic.
Equal weight is a judgment too. So are the two thresholds, which came out of the article that established the test rather than out of anything I measured.
I kept the rule short enough to run with the tool closed. Seven yes or no questions and two lines.
The cost points against me. The score treats a weak signal and a strong one as equals, and what comes back is a count, labeled as a count.
What is the one score it will not give you?
Zero.
The reveal button stays disabled until you have checked at least one signal, so the reachable range is one through seven. The band table starts at zero. The interface does not.
The clean posting, the one case where this thing should be happiest to say apply normally, is the case it will not answer. I did not see it until I went back through the component to write this piece. It is still there.
The row of seven pips under the checkboxes has the same shape of problem. I had been calling them progress without noticing they are the score with the digits removed. They fill as you check signals. They know nothing about how much of the list you have read.
Why do the rows lock?
When you press reveal, every row locks. You cannot walk a signal back after you have seen the verdict.
The cost is plain. Fix one wrong answer and your only move is reset, which clears all seven.
There is no running number while you answer. If your system asks for reduced motion the count snaps instead of animating, and every animation in the component is gated on that same check.
The one link out to Orbyt Jobs renders on the caution and skip verdicts only. A clean result has nothing to sell you, so it says nothing.
What does it refuse to do?
It does not read the posting. Fetching one could give you the date and tell you whether a range is printed. It could not tell you whether the role exists on the company's own careers site, whether a human is attached, or whether you ever heard back.
Two of those three are about you, not the posting. Automating half a checklist and hiding the seam is not a better tool.
It stores nothing, and it is not silent either. Two analytics events fire, on reveal and on copy, each carrying the total by itself. No posting text, no company name, no account. At seven of seven the total does describe every box you ticked.
And it does not tell you a posting is fake. It cannot. A ghost job is a real listing from a real company that is not going to hire anyone from it, and seven signals are evidence, never proof of intent.
The headline the page is built around is that roughly one in five active postings is a ghost. That figure comes from postings measured on one hiring platform, not from a census of the internet, and the breakdown carries the denominators.
The law is moving, slowly. Since the start of 2026 Ontario has required employers with 25 or more staff to state in a public posting whether it is for a real vacancy. There is no federal equivalent in the United States.
What happens when you share a score?
Copying writes a line of text plus a link carrying your number. There are eight such links, one per score.
Open one and the social card becomes a generated image of the score and the band, and the title becomes the score. Those variants ship noindex follow, and the canonical stays pinned to the clean address. A score parameter that is not a whole number from zero to seven falls back to the plain page.
A cold visit gets the designed static art instead. A dynamic card built for a number nobody has is worse than a picture somebody drew.
Why is there a game about this too?
Before the detector there was a maze game called Ghosted, where the job board is the labyrinth. Four recruiters chase you. The Ghost comes straight at you, the Closer cuts ahead of where you are facing, the Algorithm pincers, and the Flake does whatever it likes.
Grab a warm intro and they panic and you can close them out. Get cornered with no follow ups left and the screen reads GHOSTED.
Same subject, opposite register. The game is the feeling. The detector is the arithmetic.
You need both. A checklist does not capture what it is like to be on the receiving end of this. The game lives in the arcade, and the rest of these build notes are on the blog.
Related reading:
- The Ghost Job Detector seven checks and a verdict, with no account and no URL field
- How to spot ghost jobs where the seven signals came from, with the sourcing the tool leaves off
- Ghosted the maze game version of the same subject, if you would rather feel it than count it
- Free tools the cover letter generator, the salary widget and the unemployment calculator live here too
- The job search guide what to do with the hour a skipped posting hands back
Methodology
I read every number here out of three files in the Ghost Job Detector directory. ghost-job-signals.ts holds the seven signal objects and the three scoring bands, GhostScorer.tsx holds the reveal button with its disabled condition and the two analytics calls, and page.tsx holds the parser that decides which score values the page will accept. I counted the signal objects and the band objects by hand and took the cutoffs straight out of the band array rather than out of any prose describing it. The repository is private, so the artifact a reader can open for themselves is the live detector page, which renders the same arrays.
Limitations
These five numbers describe what the code counts. They do not tell you whether the count is right. Nobody has run a posting through this tool, waited, and come back to say whether the job turned out to be real, so there is no evidence that four out of seven predicts anything, and the tool stores nothing, which means there is no record of a single run to learn from. The seven signals and the two cutoffs came out of an article I wrote rather than out of a dataset, and giving every signal one point is a judgment I made because I could not defend any other number. The one in five headline is not mine either. It was measured on one hiring platform, so it describes that platform's postings and not the internet.
Sources
Ghost job score questions? Answers.
Related research
Follow the work.
I write these while the Machine runs. Get the next one wherever you already read.
- Subscribe on Substack (opens in a new tab)
- RSS feed
- Bluesky (opens in a new tab)
- Threads (opens in a new tab)
Add Orbyt Labs as a preferred source on Google (opens in a new tab)




