The Stalled Post-Layoff Search Log, Diagnosed Stage by Stage to a Restart

This is a teardown of a stalled post-layoff search: six months, a couple hundred applications, almost no interviews, and no clear idea which part is broken. It is for anyone who was laid off and is asking what they are doing wrong. It follows one composite case all the way through its counts so you can run the same arithmetic on your own log. By the end you will have a named failing stage, an honest read of how sure you can be about it, and a four-week restart that changes one thing.

What does a stalled six-month search look like once it is logged?

Once it is logged, a stalled search usually turns out to be one blended callback rate of around 2%, sitting on top of three different searches that nobody ever separated. The case below is a composite built to look like real logs. The counts are illustrative, but the arithmetic and the decisions are the method.

The searcher is a product manager laid off from a fintech company. At week 26 the raw totals are:

• 226 applications sent
• 4 recruiter screens (a recruiter screen is the first phone call with a recruiter, before any hiring-manager interview)
• 2 interview loops (a loop is the full set of interviews with the team)
• 0 offers
• Unemployment benefits end at week 26

What she did along the way, in order:

1. Week 2: added "open to work" to her LinkedIn headline.
2. Weeks 1 to 9: sent 80 applications across three kinds of role: fintech PM, big-tech or consumer PM, and healthcare software PM.
3. Week 10: the overall numbers looked bad, so she rewrote her resume from scratch. That same week she also started sending more applications to consumer roles.
4. Week 14: took a part-time interim job outside her field and listed it at the top of her resume.
5. Weeks 23 to 26: with benefits about to end, she sent 26 applications in three weeks, against a previous average of roughly 9 a week.

At week 26 her own diagnosis was: "My resume is still wrong, and the gap is killing me." Both halves were mostly wrong, and the log shows why. Three of the five moves above were wrong turns: the resume rewrite, listing the interim job, and the volume surge. Each one is examined below.

How much of the problem is the six-month mark itself?

Some of it. But by month six most of the penalty for a long gap has already landed, and what is left of it is mostly outside your control. The audit evidence says callbacks fall fastest in the first eight months and then flatten. So the thing you can still change is which employers see your application, not the date on your gap.

An audit study sends matched fictitious resumes to real job postings and counts the callbacks. Four designs bear on this question:

• Kroft, Lange and Notowidigdo sent resumes to postings in 100 US cities. Callbacks fell significantly as the unemployment spell got longer, and most of the decline came in the first eight months. A later summary puts callbacks after that point flat at about 4%. The authors' explanation is that employers read spell length as a signal of productivity they cannot otherwise see, and they trust that signal less when the labor market is weak.
• Ghayad sent 4,800 fake resumes to 600 job openings. Interview requests dropped sharply once the spell passed six months, whatever the applicant's prior experience.
• Eriksson and Rooth, working in Sweden, found that callbacks only started falling with spell length after nine months.
• Farber and colleagues found no link between callbacks and spell length for experienced college-educated women applying for administrative support jobs. Nunley and colleagues found no effect for young educated workers, although Kroft and colleagues did find one for young workers.

So the duration penalty is real, it is not universal, and at six months it is mostly already priced in. What still varies is who does the screening. Farber found that less selective employers react less to applicant attributes, and Kroft found the spell-length signal counts for less in slack markets. The mechanism is that an employer drowning in applicants uses cheap proxies, and spell length is a cheap proxy. If you aim at clusters where the screen is less crowded, you change the test you are sitting.

How common is a long spell in your occupation?

You are probably not an outlier, and it helps to know how common your situation is. These are the latest BLS readings I captured from Table A-37, which breaks spell length down by occupation. Long-term unemployment means a spell of 27 weeks or more.

Roughly a third of unemployed people in management and professional occupations are past 27 weeks, against 27.0% of all unemployed. These figures change every month, so re-check BLS Table A-12 (all unemployed) and Table A-37 (by occupation) before you quote them. I have left out the all-unemployed mean and median because the figures I have come from a secondary source and are not verified against BLS.

One more piece of context: Farber found that applicants aged 50 and over get significantly fewer callbacks. If that applies to you, part of your low rate may come from age, not the gap. That changes what a fix can achieve, and it is covered under failure modes below.

Which benchmarks should you hold your own funnel against?

Use Ashby's employer-side application-to-interview rate as the outer bar: 3.6% for technical roles and 4.7% for business roles. Use your own best cluster as the inner bar. Neither is enough on its own, because the employer average includes unqualified and spam applicants. Your funnel is the sequence of stages from application to screen to loop to offer.

How to read the table:

• What the Ashby rate tells you: roughly where an average applicant to an average posting ends up. If your best-targeted cluster falls well below 3.6%, something in that cluster is not working.
• Where it misleads you: it is averaged over everyone who applied, including people who were never qualified. A focused searcher sitting at 4% may actually be underperforming. That is why your own best cluster is the second bar.
• What the audit rates are: context, not a benchmark. An audit callback is any positive response to an identical fake resume, so do not compare it directly with your screen rate.
• What is not established: I found no public breakdown of these rates by seniority. Ashby and Gem also disagree on the late stages. Ashby has post-onsite and offer stages converting at 95% and 81%, while Gem puts onsite-to-offer at 39%. Do not build a diagnosis on the late-stage benchmarks.

The shape of the market points to one stage. In Ashby's data, applications per hire rose from about 100 to 291 over the period they track. Greenhouse measured 203 applications per hire across more than 6,000 North American employers. That much volume swamps the first screen. For a stalled searcher, the failing stage is most likely application to screen, and coaching for later interview rounds rarely fixes that.

The table also contains the one documented lever with a multiplier. Referred candidates pass the recruiter screen at 52% against 35% overall, about 1.5 times as often. The reason is that a referral swaps the cheap proxy screen for a person vouching for you.

The teardown: splitting 200 resolved applications by cluster and channel

Splitting the log by cluster turned a 2% blended rate into one cluster converting at 5.7% and two converting at zero. That changed the diagnosis from "fix the resume" to "stop aiming at two of my three targets".

Remove what has not resolved yet

The 26 applications from weeks 23 to 26 have not had time to get a response, so they would count as zeros if left in. I remove them and work with the 200 resolved applications. Leaving out the last two to three weeks is a desk convention, not a published standard. Use it anyway, because the alternative inflates your failure count right when you are deciding what to change.

Split by cluster

The rule-of-three ceiling is the highest true rate still consistent with the zeros you saw, at 95% confidence. The next section explains it.

Split by channel

Inside cluster A, 64 cold applications produced 2 screens (3.1%), and 6 referred applications produced 2 screens. Cluster B had 74 cold applications and 6 that came through recruiters, all with zero screens. Cluster C was all cold, 50 applications with zero screens. The referral cell is far too small to prove anything on its own, but it points the same way as the Ashby multiplier.

Check the later stages

Screen to loop was 2 of 4, or 50%, against roughly 35% overall. Loop to offer was 0 of 2, against an onsite pass rate of 24%. Two loops cannot tell you anything about interview skill. Interview coaching is not where the problem is.

Replay wrong turn 1: the resume rewrite

The blended rate fell after the rewrite, so the rewrite looks like it made things worse. Inside cluster A, though, the rate went from 5% to 6.7%. The drop in the blended rate came from the change in where she applied. She moved volume into B and C in the same week she changed the resume. Two variables changed at once, so nothing can be credited or blamed. The blended rate made it look as though the new resume had failed. Split by cluster, it shows the resume was never the main problem.

A blended rate averages a cluster that works with clusters that do not. Split it before you rewrite anything.

The case's verdict: cluster A sits in the top-right, above the Ashby band and her best cluster by a distance. Cluster B sits near the bottom-right. Its ceiling of 3.75% is just above the 3.6% technical bar, so on its own it cannot be declared below benchmark, but next to cluster A it is clearly the weaker bet. Cluster C is bottom-left at 50 applications. The failing stage is application to screen, outside cluster A.

How many applications does a verdict need?

More than most people think. Zero responses in n applications only tells you the true rate is below 3/n, and telling two rates apart takes hundreds of applications per arm (one arm is one version of the search you are comparing). Treat any cell under about 60 applications as inconclusive.

Use these rules when reading your own log:

• Rule of three: zero callbacks in n applications puts the true rate below 3/n with 95% confidence. So 100 applications with no callbacks means your rate is below 3%, not that it is zero.
• Chance zeros: at a 5% base rate, a run of 20 applications with no callbacks happens about 36% of the time by chance (0.95 to the power of 20). A run of 20 zeros is not evidence that something is broken.
• Power to compare: these sizes come from the standard two-proportion power formula, two-sided alpha of 0.05 and power of 0.80. Nothing is published for job seekers specifically, so treat them as approximate.

What this means for a couple hundred applications: split across three clusters, that is roughly 67 per cell. That is enough to rule out rates of about 4.5% and above in a zero cell, and not enough to prove one cluster beats another. Two conclusions follow. First, you act on how the evidence points, because waiting for statistical proof would take another year. Second, the restart should use fewer clusters with more volume each, so the next read is sharper than this one.

Step-by-step: from raw log to a four-week restart

The procedure turns your history into a matrix of conversions, names one failing stage, and changes one variable for four weeks while everything else stays fixed. Expect the log rebuild to take about 3 to 4 hours and each of the next two steps about an hour. These are desk estimates, not sourced figures.

The log needs one row per application and one terminal status per row. Copy this header into a spreadsheet:

date_applied,company,role_title,posting_age_days,cluster,channel,furthest_stage,terminal_status,resume_version,interim_role_listed,notes
2024-01-08,Example Co,Senior Product Manager,3,A-fintech-PM,referral,screen,rejected,v2,no,referred by former manager

The resume_version and interim_role_listed columns are there for a reason. They let you check after the fact whether two variables changed at once.

For step 2, keep cluster tagging honest. A cluster where you are only half qualified will show zeros that are really about fit, not channel or resume. Refolk scores how well you actually fit each posting, which gives you a defensible way to tell those apart before you count a cluster as dead.

Sources disagree on what comes first. One guide puts a two to three day pause first, then tracking your last ten to fifteen applications. Another puts outside eyes on your materials first. I put the log first. Asking someone to review your materials assumes the materials are the problem, and in this case they were not. Fifteen applications are too few to separate clusters.

The case restart

The one variable is cluster mix: all volume goes to cluster A, B is dropped, and C is paused. Held fixed: the post-rewrite resume, the channel mix (about 9% referred, matching cluster A's history), and the interim job staying listed.

Over four weeks she sent 36 applications, all in cluster A: 33 cold and 3 referred. Illustrative result: 2 screens, or 5.6%. By the power table that proves nothing. It does show three things. Screens kept arriving at cluster A's historical rate. Volume stopped going to a cluster that had returned 80 zeros. And she now gets about one screen every two weeks instead of four in 23 weeks. The verdict is keep, and the next test is referral share.

If the variable you choose is the resume itself, the treatment arm is a resume tailored to each posting. Refolk tailors the resume to every posting you apply to, which lets you run that arm at full volume inside one cluster instead of dropping to a handful of hand-edited applications.

For a sense of cost: Krueger and Mueller found the average US unemployed worker spends about 41 minutes a weekday on job search. A 3 to 4 hour log rebuild is about a week of that time, spent once.

How the diagnosis goes wrong

Most failed diagnoses come from reading a blended or tiny number as a verdict, or from changing two things at once. Each failure mode below gives what it looks like and the check that catches it.

1. Pooling clusters. Looks like: a 2% overall rate that leads you to rewrite your resume. Reality: one cluster at 6% and another at 0%, where the fix is to drop the dead cluster. Check: compute rates per cluster before touching your materials.
2. Benchmark mismatch. Looks like: "I'm at 4%, inside the 3.6% to 4.7% band, so I'm fine." Reality: that band includes spam and unqualified applicants, so a targeted searcher at 4% may be underperforming. Check: compare against your own best cluster as well as Ashby.
3. Small-n verdicts. Looks like: 20 zeros, so the cluster is dead. Reality: at a 5% base rate, 20 zeros happen 36% of the time by chance. Check: apply the rule of three before calling any stage broken.
4. Right-censoring. Looks like: your most recent weeks are your worst. Reality: the last two to three weeks of applications have not had time to resolve. Check: exclude them from the rates (a desk convention, not a published one).
5. "Any job helps." Looks like: listing the interim role to cover the gap. Reality: Farber found that taking an interim job significantly lowers callbacks: 9.16% overall against 9.82% with no interim job. The effect was weaker at less selective employers, and this is the only US design that tested it. Check: run the same cluster with and without the role listed.
6. Two variables at once. Looks like: the case's week 10, where a new resume and a new cluster mix went out together. Reality: the lift or drop belongs to neither. Check: the resume_version column and a written list of held variables.
7. Trusting profile signals. Looks like: "I turned on open to work, so recruiters can see me." Reality: the data below. Check: do not count headline changes as part of the restart.
8. Mixing up age and duration. Looks like: hiding the gap to fix low callbacks at 50 or older. Reality: Farber found workers aged 50 and over get significantly fewer callbacks, so fixing the gap may not move the rate. Check: for those applicants, test less selective clusters, where Farber found the age effect substantially weaker.

Why the open-to-work headline does not help

In Refolk's index of professional profiles, very few product managers put the phrase in their headline, and the ones who do are not reliably looking.

Profile counts and headline matches come from Refolk's index, and the shares and multiples are derived from them. Since several of the matches are people already employed at large companies, a recruiter cannot read the phrase as "available". And since almost nobody uses it, it does not make you visible either. The case's week-2 headline change was harmless, but it was not a lever.

What to do after the four weeks, and how to keep the log current

After the restart, make one of three decisions, write it into the log, and start the next queued variable. Keep the log running the whole time, because it doubles as your work-search record when your benefits office reviews it.

The three decisions:

• Keep the change if the target cluster held or beat its own prior rate.
• Revert it if the cluster's rate fell compared with its own history, not compared with the old blended rate.
• Next: start the next queued test.

The case's test queue, in order:

1. Referral share in cluster A: 52% against 35% screen passthrough is the strongest documented multiplier.
2. Interim role listed or not, same cluster, same channel mix.
3. Resume tailored or not, only if cluster A stops converting.

Benefits and the RESEA meeting

Benefit length varies by state and changes over time. Most US states pay up to 26 weeks, 16 pay fewer, and one pays more. At my last check, Florida paid up to 12 weeks and Georgia up to 14, and Extended Benefits were not triggered in any state. Re-check the CBPP weeks-of-UI tracker and your state's labor department before you plan around a date.

RESEA (Reemployment Services and Eligibility Assessment) selects claimants most likely to exhaust their benefits. The meeting includes an eligibility assessment and a review of your work search. Timing is set by each state. Vermont selects within the first five weeks of a claim, and Massachusetts schedules the initial meeting by week 3 and a review by week 5. A clean log answers the work-search review directly.

The evidence on how search effort changes over a spell conflicts. Krueger and Mueller found effort rising before benefits run out. Their later weekly survey found search time falling sharply over the spell, with no rise when extended benefits ended. Marinescu and Skandalis, using administrative data, found search effort up 51% in the year before exhaustion and still 42% above baseline a year after. Whichever pattern fits you, extra effort only helps if it goes into a cluster that converts.

The weekly review

Week: __ of 4
Variable being tested: __________
Held fixed: resume version __ | channel mix __ | interim role listed: yes/no | clusters: __
Target cluster this week: __ applications, __ screens, __ loops
Running total since restart: __ applications, __ screens (rate __%)
Target cluster's own prior rate: __%
Unresolved rows excluded: __
Anything that broke the hold (and what I did about it): __________
Decision due: keep / revert / next variable

Re-check the market data monthly as well. BLS Tables A-12 and A-37 update with each jobs report. A rising long-term share in your occupation means more competition for the same screens. It is not a reason to blame your own materials.