Diagnose a Thin Pipeline Before Widening It, Then Assess Everyone

What do you do about a role almost nobody applies to?

Separate the two causes before spending anything. Low volume used to have one meaning, that too few people saw the posting, and the standard remedy list was written for that world. It now has two, and they point in opposite directions. A posting nobody saw is a reach problem with cheap fixes. A role few people can do is a market fact, and no amount of reach will change it.

The reason the old list stopped working is that the base rate moved for some roles and not for others. Ashby's benchmark of over 109 million applications reports the average recruiter processing 291 applications per hire against roughly 100 in early 2021, with the share of applications reaching an interview down to between 3.6% and 4.7% depending on role type ¹. That sample is one vendor's customer base, weighted to venture-backed technology employers hiring knowledge roles, and it says nothing whatever about hourly, clinical, field or public-sector work ¹. Which is exactly the point. Silence is not evidence in either direction, so nine applications is a question about the market for that skill before it is a verdict on your posting.

So the first move costs nothing and takes an afternoon:

1. Write down your application count and the days the posting has been open. 2. Find three live postings for the same work, at roughly the same pay, in the same geography. 3. Compare. Everything else follows from whether you are in range.

How do you tell low reach from real scarcity?

The only reliable test is comparison: get the application count on three live postings for the same work at the same pay, and see where yours sits. Several job boards publish an applicant count, a peer at another employer will usually answer a direct question, and a recruiter who works that market can tell you in one call. Your posting is either drawing what comparable roles draw or it is not, and that single fact routes the whole response.

If comparable postings draw normal volume and yours does not, the leak is on your side and there are four places to look, in cost order. The application form comes first. Appcast's 2025 benchmark, built on over 281 million clicks and 25.6 million applies across more than 1,300 United States employers, puts the median apply rate on long-apply ATS ads at 6.1% against a 19.37% average on one-click easy-apply ads ². Those two rates are not a controlled comparison: easy-apply traffic comes from different boards, different ads and different candidate intent, so the roughly threefold gap is an upper bound on what shortening a form would buy. Both figures also count completed applications per click on paid job ads, which is not the traffic arriving at your own careers page ². Then the title, which is the query candidates actually search. Then the pay, and whether you posted it at all. Then the boards.

If nothing comparable draws volume either, stop editing. Every fix on the standard list adds reach into a population that does not contain the skill, and reach into the wrong population arrives as noise you then have to read. The honest description of that situation is that you are competing for a small number of people who are mostly employed, which is a sourcing and compensation problem rather than a marketing one.

One caution about timing. Judge volume against days open, because a posting three days old and a posting three weeks old are not the same evidence.

Assess everyone who applies, and write back

Under about twenty applicants, the cheapest good decision available is to give all of them a real piece of the work and a written result. The screening budget you are not spending on volume is sitting unused, and depth is the one thing a thin pipeline can buy that a full one cannot. A handful of well-evidenced candidates beats a pile of unread ones on every dimension except the one nobody should be optimising for.

The method is not exotic. Take a task from the actual job, cut it to sixty or ninety minutes, write the answer key before you send it to anyone, and score every submission against the same key. A job-shaped exercise predicts job performance about as well as the best-evidenced interview method: the 2022 meta-analytic estimate for work samples is .33 against .42 for structured interviews, close enough that their credibility intervals overlap heavily and the paper reads the pattern as coherent, not a demotion ³. Two caveats belong with that. Almost all of the underlying studies tested people already doing the job, and none of them tested an unpaid multi-hour take-home, which is a different animal.

Then write back to everybody, including the people you decline. Two or three sentences naming what their submission showed costs a few minutes each at this volume and is the only part of the process the candidate can use. In a scarce market the person you decline this quarter is a person you may be approaching next year, and what to keep on them before that role opens decides whether the approach starts warm.

The wider case for putting the evidence stage before the document screen is made in whether it is cheaper to assess everyone up front, and the version where the resume screen goes entirely is in dropping the resume screen for a work sample. A thin pipeline is where both arguments are easiest to run, because the arithmetic that makes them expensive at volume does not apply.

Cut the requirements nobody validated

A thin pool is the honest moment to delete a requirement you cannot defend, because a degree line or a years-in-role floor removes a much larger fraction of a small pool than of a large one. It is also the moment the cost is visible: with nine applicants, one unjustified line can be the difference between a shortlist and nothing.

Candidates read requirement lists as literal eligibility rules, and one survey put a number on it. Asked why they had not applied for a job whose qualifications they did not fully meet, over a thousand professionals gave the same top answer, twice as common as any other for both women and men: they did not think they would be hired since they did not meet the qualifications, and did not want to waste the effort. That was the top reason for 41% of women and 46% of men, while the least common answer of all was doubting they could do the job well ⁴. That is a self-report survey published as a blog post, with the sampling frame unstated and no control condition, and it measures what people say about a decision already taken. The piece is usually remembered as the origin of the confidence-gap story. Its own numbers cut the other way.

So run the list line by line and ask one question of each: could you test this in the exercise you are about to send? If yes, keep it and test it. If no, it is not a requirement, it is a preference, and it belongs in the duties section where it deters nobody.

The requirements most worth deleting are the ones that were never chosen. A degree line copied from a decade-old template, a years-of-experience floor inherited from a level nobody uses any more, and a tool name that changed last year all filter real candidates for reasons no one in the building would defend out loud. In a market where the pool is small, that is the most expensive kind of line to leave in place.

See what gets scored >