We Read 31 Years of US Job-Search Data. One Method Collapsed.

In 1995, if you were out of work in America, the most common thing you did about it was walk up to an employer and ask. Nearly two thirds of jobseekers did that. Fewer than half sent out a resume.

Thirty years later those two lines have swapped places, and the swap has a date.

The US government has asked unemployed jobseekers what they have done to find work, every month, since the questions were redesigned in 1994. Seven methods, published as annual averages, broken out by sex and age. It is the longest-running official measurement of US job-hunting behaviour we are aware of. BLS itself last analysed this series in 2000; its more recent work on job search, in 2020, uses a different dataset entirely. We found no analysis of the annual series for a general audience in the intervening twenty-six years.

We pulled all thirty-one published years — 1995 through 2025 — and re-read them. One method collapsed. Nothing else really moved.

We also checked our own site against what we found, and had to change seven things on five pages. That section is further down and it is not decorative.

The year the resume overtook the handshake

Two lines matter more than the other five.

In 1995, 65.3% of unemployed jobseekers said they had contacted an employer directly, against 45.1% who had sent out a resume or filled out an application — a gap of twenty points. Direct contact led every year for fifteen years.

In 2010 the lines crossed. Resumes and applications reached 55.2%; direct contact fell to 53.6%. They have never crossed back.

By 2025 the gap has reversed and widened: 56.1% sending resumes or applications, 50.8% contacting an employer directly.

There is a detail that makes the comparison unusually clean. In 1995 there were 6.375 million unemployed jobseekers. In 2025 there were 6.405 million. Two labour markets thirty years apart, half a percent apart in size. Whatever changed between them, it is not that one year was a recession and the other a boom.

Only one method actually collapsed

Across thirty-one years, contacting an employer directly fell 14.5 percentage points. No other method moved that far in either direction.

The obvious objection is that job search looks different in a bad economy, and the period contains three NBER recessions. So we checked it.

Direct contact has the weakest relationship with the business cycle of any method and the steepest trend. Every other line rises when unemployment rises; this one barely notices.

To be sure, we compared like with like: taking only the sixteen years when the jobseeker count sat at or below the series median — that is, only the healthier labour markets — the decline survives almost unchanged, at −5.8 points per decade.

This is the finding. Not that job hunting went digital, which everyone already believes, but that the specific behaviour that died was showing up and asking.

Everything else is the business cycle

Put all seven methods on one chart and the shape of the period is obvious.

The lower cluster bulges in 2008–2013 and subsides. That bulge is not a change in how people look for work. It is what a recession looks like when you measure behaviour: more people, trying more things.

The clearest way to see it is the measure the tables publish and nobody quotes — the average number of methods a jobseeker used.

Search effort tracks the number of jobseekers at r = +0.89. It rose to 2.04 methods per person in 2009 and fell back as the market recovered. And over the whole thirty-one years its trend is zero: 1.80 in 1995, 1.8 in 2025.

Through the arrival of the commercial internet, job boards, LinkedIn, one-click apply and AI-written applications, the average jobseeker does not use more methods than they did in 1995. They use the same number, and they use more of them when the market is bad.

Networking rose, but less than you have been told

“Contacted friends or relatives” went from 17.3% in 1995 to a peak of 28.8% in 2010, and back to 23.7% in 2025.

Read carelessly, that is the networking revolution. Read against the table above, most of it is the recession: the series correlates with the jobseeker count at +0.67, and its peak sits exactly on the employment trough. A real secular component does survive when you control for the cycle — about +3.3 points per decade in the healthier years — so the rise is real. It is just much smaller than the shape of the raw line suggests.

It is also worth stating what this series does not say. Roughly a quarter of unemployed jobseekers use personal contacts. That is a measurement of who tries a method, not of which method produces jobs — and the famous claim that most jobs are filled through networking is not measured here or, as far as we can find, anywhere with a primary source. We looked, because it is on our own site. More on that below.

Thirty-one years, thirty-one out of thirty-one

The published tables split every method by sex, for every year in the series. The result is the most consistent thing in the dataset.

Not a trend. Not a majority of years. Every single year, in all three methods, without one reversal — through three recessions — including the pandemic — and the entire history of the consumer internet.

Why this happens is not established by this dataset, and we are not going to pretend otherwise. Occupational composition is a live explanation: women and men are not evenly distributed across occupations, and some occupations hire almost exclusively through formal application while others run on personal contact. Differences in network access are a live explanation. So is the possibility that men and women describe the same activity differently to an interviewer — which matters more than usual here, for a reason we explain in the next section. This table cannot separate those, and neither can we.

What we can say is that the pattern is not noise. Three separate methods, thirty-one consecutive years, no exceptions.

The resume is a young person’s instrument

Age detail exists only from 2011, because the earlier table broke jobseekers down by reason for unemployment instead. Within those fifteen years the gradient is stable and steep. In 2025:

Fifteen points separate the youngest jobseekers from the oldest on formal application, and the personal-contact line runs the other way. Both directions hold in all fifteen years with published age detail — the youngest band is above the oldest on resumes in every year, and the 55–64 band is above the youngest on personal contacts in every year. The gradient is not strictly monotonic band by band: that holds in only five of the fifteen years, and the middle bands shuffle.

What this series cannot tell you

Four limits, each of which rules out a sentence we would otherwise have liked to write.

The question is open-ended

This is the most important thing to understand about the dataset, and it is almost never mentioned. The Current Population Survey does not hand people a list. Item LKM reads:

“What are all the things you have done to find work during the last 4 weeks?”
Do not read answer categories.

The interviewer records a free-text answer and codes it into thirteen categories. So the series measures how people describe their job search, not what they tick on a form. A line can move because behaviour changed, or because the language people use to narrate that behaviour changed. Both are real; they are not the same. Every claim above is written to survive that reading, and it is why we lean on the cycle controls rather than on the raw slopes.

It covers unemployed jobseekers only

The universe is unemployed people who actively looked for work in the last four weeks, and it excludes people on temporary layoff — the tables say so in their own footnote. Everyone searching while still employed is absent. That is a large share of all job search and none of it is here.

Two methods are fused, and one common behaviour is invisible

“Sent out resumes or filled out applications” is a single code and cannot be split. This also means online applications sit in that line rather than in “placed or answered ads” — BLS’s own definition of an active method explicitly includes submitting a resume or application “to an employer or to a job website.” So the falling ads line is about advertisements, not about job boards, and should never be read as one.

Separately, browsing postings without applying is coded as a passive method and does not appear in this table at all. A person who only scrolls listings is not counted as an active jobseeker. Given how much modern job search consists of exactly that, this is a real hole.

The percentages do not add up, on purpose

People report more than one method, so the columns sum to well over 100. These are shares of jobseekers using each method, not slices of a pie, and nothing here supports a chart that treats them as one.

One discontinuity is worth flagging, because it is visible in the chart above: the “average number of methods used” column is published to two decimals through 2010 and to one decimal from 2011. The flat runs in the effort series after that date are rounding, not behaviour, and the chart says so on its face.

We checked ourselves, and our first correction pass had missed things

Our published standard requires that any claim on this site touching a study’s subject is audited by the same protocol before the study goes out. We scanned all 159 published pages for claims about how jobs are found. Most hits were figures inside example resumes, which are presented as illustrations. Four were unsourced factual claims, and the scan turned up two more problems we were not looking for.

“At C-suite level, 80% of positions are never posted publicly”

This was on our CEO resume page, unsourced. It is a variant of the “hidden job market” folklore. The usual reconstruction traces that to a 1966 book drawing on a pilot survey of 27 companies in Rochester, New York, in a market where the local industrial council asked members not to advertise in the city’s newspapers — we could not verify that chain against a primary source and do not rely on it here. What we can say is the part that matters: we found no study, anywhere, measuring the share of executive positions that are never posted publicly.

The sentence that followed it was worse, because it was checkable and wrong: we said those roles are filled through search firms. At large-cap level they are mostly filled from inside. Externally recruited CEOs have averaged 24% of appointments since 2016 across the Fortune 500 and S&P 500. The page now says that, with the source attached.

“The majority of senior COO roles go through executive search firms”

Same family, on our COO resume page, and an instructive one: the conclusion was right and the mechanism was wrong. Senior COO roles genuinely are not public postings — because 84.6% of COO hires in 2026 were internal promotions, not because search firms own them. That is a more useful thing to tell a COO candidate, and it is sourced.

“80% of the hiring decision weight”

Our marketing manager page said research on recruiter behaviour is “consistent” that a summary and the first two bullets carry 80% of the hiring decision weight. There is an ancestor, and the drift is precise. TheLadders’ 2012 eye-tracking document says recruiters spent “almost 80% of their resume review time” on six data points — name, current title and company, previous title and company, both sets of dates, and education. That is share of gaze time, not decision weight, and it describes identity and chronology fields, explicitly not the summary or the bullets. The same document says that beyond those six points recruiters “did little more than scan for keywords.”

We could not repair the sentence by citing that study, because our first study publicly describes both TheLadders reports as promotional documents with inadequate methodology, and repairing a claim with a source you have published a negative assessment of is incoherent. The page now uses the ResumeGo 2024 survey our own blog already relies on.

“Over 80% of consulting postings”

Unsourced, on our consultant resume page. We removed the number rather than softening it. Changing “over 80%” to “most” would have kept the false precision and dropped the accountability.

The two we were not looking for

Three pages still carried “a 6-second scan” — the exact claim our first study traced to a five-page vendor document whose entire methodology is one sentence. Our own remediation pass after that study missed them, because it matched the phrasings it had found and not the ones it hadn’t.

A correction pass that looks for known bad sentences clears the sentences it knows about and leaves the rest. All three now use the ResumeGo 2024 figure instead.

How to read a job-search statistic

Four questions, in the order that kills claims fastest.

Method

Source: Unemployed jobseekers by sex, age, race, Hispanic or Latino ethnicity, and active jobsearch methods used, US Bureau of Labor Statistics, Current Population Survey annual averages. Two tables carry it: table 34, which splits jobseekers by sex and reason for unemployment and has run throughout, and table 33, added in 2011, which splits them by sex, age, race and ethnicity. The seven method columns and the average-methods column are identical in both. We used table 34 for 1995–2010 and table 33 from 2011, where the age detail lives.

Every published year from 1995 to 2025 was retrieved from bls.gov in its original format — plain text for 1995–1999 and 2002–2010, PDF for 2000, 2001 and 2011–2022, HTML for 2023–2025 — and parsed by mapping each file’s own printed header rather than assuming a column order. The parser reports any file it cannot map instead of guessing, and is published with the dataset.

The extraction was validated against an independent BLS delivery path: for the 23 overlapping years, the published annual average and the mean of the four quarterly values from the BLS Public Data API agree to a median absolute difference of 0.03 points, with a maximum of 0.20 across 154 comparisons.

The series begins in 1995 because the job-search question battery was redesigned in 1994 and BLS states directly that earlier surveys “used a different list of search methods.” Nothing in this study crosses that break.

One collection gap: October 2025 was not collected during the federal shutdown, so the 2025 annual average rests on eleven months. It is the only missing month in the series.

Correlations are Pearson coefficients between each method’s annual value and the annual jobseeker count. Trends are ordinary least squares slopes on the year. The “tight labour market” subset is the sixteen years in which the jobseeker count was at or below the series median of 6.405 million.

FAQ

What is the most common way Americans look for work?

Among unemployed jobseekers in 2025, sending out resumes or filling out applications, at 56.1%, followed by contacting an employer directly at 50.8% and contacting friends or relatives at 23.7%. These sum to more than 100% because people report multiple methods. Note that the universe is unemployed jobseekers only — people searching while employed are not measured by this series.

Is the resume dead?

Not in this data. The share of unemployed US jobseekers who sent out a resume or filled out an application was 45.1% in 1995 and 56.1% in 2025, and it has been the most-used method every year since 2010. The caveat that matters is that the code fuses resumes with applications and cannot be split, so this is a measurement of formal written application in general rather than of resumes specifically.

Do most jobs really come through networking?

This series cannot answer that, and we could not find a primary source that does. What it measures is that about a quarter of unemployed jobseekers use personal contacts as one of their methods — 23.7% in 2025, peaking at 28.8% in 2010. That is a count of people trying a method, not a count of jobs filled. The widely repeated claim that 70–80% of jobs are never advertised is usually traced to a 1966 book reporting a pilot survey of 27 companies in a single city; we could not verify that chain against a primary source, and we could not find any study that measures the quantity at all.

Has job hunting got harder or more intense over time?

Not by the one measure available here. The average number of methods a jobseeker used was 1.80 in 1995 and 1.8 in 2025, with no trend across the thirty-one years. It rises and falls with unemployment — 2.04 at the 2009 peak — correlating with the jobseeker count at 0.89. Bear in mind this counts distinct methods, not applications sent, and a person submitting two hundred online applications registers the same as one submitting two.

Why does the data start in 1995?

The Current Population Survey redesigned its job-search questions in 1994, and BLS states that earlier surveys used a different list of methods. Figures from before that break are not comparable with these. We start in 1995 because it is the first full year of published annual averages under the current question set.

Sources

Primary data

• US Bureau of Labor Statistics — Table 33: Unemployed jobseekers by sex, age, race, Hispanic or Latino ethnicity, and active jobsearch methods used, Current Population Survey annual averages. Table 34, the sex-and-reason version, is published alongside it and also covers the whole period. Archived years live under /cps/aa{year}/ up to 2022 and /cps/data/aa{year}/ from 2023; the 2000 and 2001 tables are inside the full annual compilations at /cps/cpsa2000.pdf and /cps/cpsa2001.pdf. Each row of the dataset names the file its figure came from.
• US Bureau of Labor Statistics — Public Data API v2, series LNU03044176 and siblings, used for the independent cross-check and for quarterly and monthly detail. The series identifiers were taken from the published catalogue at download.bls.gov/pub/time.series/ln/ln.series.
• US Census Bureau — CPS Labor Force questionnaire, items LKM and LKDK, for the exact question wording and the instruction not to read the answer categories.
• US Bureau of Labor Statistics — CPS concepts and definitions, for the classification of active and passive search methods.

Prior analysis of these series

• Bureau of Labor Statistics — Job search methods: internet versus traditional, Monthly Labor Review, October 2000. Covers 1994–99 using December files rather than annual averages, which is why its figures differ from ours for overlapping years. Its authors note that their window “was also a period of continuous economic expansion,” making it “difficult to disentangle secular from cyclical effects” — the limitation that thirty-one years and three recessions allow this study to address.
• Michelle Ports — “Trends in job search methods, 1970–92”, Monthly Labor Review, October 1993. Cited by the above. We were unable to retrieve the full text from bls.gov and have not relied on its contents.

Sources used in the corrections to our own site

• Crist|Kolder Associates — Volatility Report 2026: America’s Leading Companies. 665 Fortune 500 and S&P 500 companies. Source of the 24% CEO external-hire historical average and the 84.6% COO internal-promotion figure.
• ResumeGo, 2024 — survey of 418 US hiring professionals, already used elsewhere on this site, now also the basis of the screening-time statement on our marketing manager page.

Source we corrected a claim away from

• TheLadders — Eye Tracking Study, 2012. A five-page promotional document for the company’s resume-rewriting service; its entire stated methodology is that “thirty professional recruiters took part in the study during a 10-week period.” Its real finding — 80% of gaze time on six identity and chronology fields — was the ancestor of the “80% of hiring decision weight” claim we removed. Linked here so the drift can be checked, not as support.