Filtering out what is not worth it

Listings exist that are not live vacancies: roles already filled internally, postings kept open to collect CVs, positions advertised to satisfy a process requirement, and outright fraud. Candidates spend real hours on all of them.

Detecting these is more valuable than improving match scores by a few points, because the effort saved is total rather than marginal.

It is also the part of the search a candidate cannot do for themselves. Nobody applying to one role can tell that it has been reposted four times in six months, and a platform seeing the whole market is the only participant with that view.

No single signal is conclusive, and several together are informative. Reposting the same role repeatedly, an unusually long time open with no change, a requirement list nobody could satisfy, vagueness about what the person would actually do, and an employer with no record of responding.

Weight them, do not treat any as decisive. Some genuine roles are hard to fill and stay open for months, and a system that condemns those is removing real opportunities.

Comparison against the norm for the role type matters more than absolute values. Four weeks open is unremarkable for a specialist position and notable for a high-volume one, and a threshold that ignores the category flags the wrong things in both directions.

• Reposted repeatedly with no substantive change
• Open far longer than typical for the role type
• Requirements that no realistic candidate profile satisfies
• No description of the actual work, only of the ideal person
• Employer with a poor observed response record
• A salary range so wide it communicates nothing

If your platform sees applications and responses, you know which employers reply and how quickly. That is information candidates cannot get anywhere else and it predicts outcomes better than anything in the posting text.

Use it carefully. Aggregate over enough applications to be meaningful, account for role type and volume, and avoid publishing something that amounts to an unverified accusation about a company.

Statistical honesty matters here because the claim is about a real organisation. Three applications with no reply is not evidence, and a system that presents it as a response rate is producing a defamatory-sounding number out of noise — so set a minimum sample and show a confidence, or say nothing.

Silently filtering postings out is a strong action on an uncertain judgement. A candidate who would have applied to a real role never learns it existed, and cannot appeal a decision they cannot see.

Show the posting with the concern stated: this has been reposted four times in six months, this employer rarely responds. The candidate decides, with information they did not previously have.

State the observation rather than the conclusion. "Reposted four times since March" is a fact the candidate can weigh; "probably a ghost job" is an inference presented as one, and the first is both more useful and more defensible.

Fraudulent postings — advance-fee schemes, identity harvesting, money laundering recruitment — have distinct patterns and cause direct harm rather than wasted effort.

Treat them differently: block rather than annotate, and be willing to accept false positives, because the cost of one fraudulent posting reaching a candidate far exceeds the cost of holding a legitimate one for review.

The patterns are specific enough to encode directly. Requests for payment, requests for bank or identity documents before any interview, contact moved immediately to a consumer messaging app, and a package far above market for the stated requirements each warrant blocking on their own rather than contributing to a weighted score.

Make review the recovery path so blocking is safe. A legitimate employer caught by these rules should be able to reach a person quickly, which is what makes an aggressive threshold acceptable rather than arbitrary.