Nonfarm Payrolls Explained: How the US Jobs Report Is Counted — and Why It Keeps Getting Revised

Why it’s suddenly everywhere

1. 01

   In August 2025, President Trump fired BLS Commissioner Erika McEntarfer hours after a jobs report revised May and June payrolls down by a combined 258,000 — the largest two-month downgrade outside the early COVID months, per Reuters.

2. 02

   Weeks later, BLS published a preliminary benchmark revision of −911,000 jobs for the year through March 2025, implying job growth had been overstated by roughly 76,000 per month.

3. 03

   The October 2025 government shutdown then delayed the September jobs report for weeks — the first interruption of the monthly series since 1948 — turning the data’s reliability itself into the story.

4. 04

   Survey response rates keep sliding: BLS’s own tracking shows the CES first-month collection rate fell from above 60% pre-pandemic to the low 40s, making first estimates noisier and revisions larger.

The timeline

1. 1915-10Four manufacturing industries

   BLS begins collecting employment and payroll data for boots and shoes, cotton, hosiery and lumber — the seed of today’s establishment survey, prompted by the 1913–14 recession.

2. 1940The household survey arrives

   The Works Progress Administration launches the sample survey that becomes the Current Population Survey — the future source of the unemployment rate, a complement to employer tallies.

3. 1943Birth of the modern report

   Employer and household data are combined into the monthly Employment Situation release — the format markets still parse at 8:30 a.m. Eastern on the first Friday.

4. 1979Two surveys, one report

   The payroll count from the establishment survey (CES) and the unemployment rate from the household survey (CPS) settle into their modern division of labor.

5. 2003Birth-death model added

   Because brand-new and dead firms can’t be sampled in real time, BLS adds a model-based estimate of jobs from business births and deaths — later a favorite target of critics.

6. 2024-08The −818,000 warning shot

   The preliminary benchmark for the year to March 2024 cuts 818,000 jobs from published estimates — the largest since 2009 — hinting the post-pandemic numbers ran hot.

7. 2025-08-01A statistician is fired

   A +73,000 July report lands with −258,000 in revisions to May and June; Trump calls the numbers "rigged" and fires McEntarfer the same day. Former BLS chief William Beach calls the move "totally groundless."

8. 2025-09-09The −911,000 benchmark

   BLS’s preliminary benchmark implies the year to March 2025 created nearly a million fewer jobs than reported — fuel for both critics of the data and defenders who note this is exactly how the process is designed to work.

9. 2025-10The data blackout

   A government shutdown delays the September jobs report for the first time since 1948; the October report is withheld entirely, leaving the Fed to set rates without its key input.

10. 2026-08-28Quiet normalcy returns

    The March 2026 preliminary benchmark comes in at just −79,000 (−0.1%), published with little fanfare — a reminder that outsized revisions cluster at economic turning points, not every year.

The backstory

What the number actually is

The “nonfarm payrolls” headline is not a headcount. It is a statistical estimate built from the Current Employment Statistics survey, in which roughly 121,000 businesses and government agencies — covering about 631,000 worksites — report how many people were on their payrolls for the pay period including the 12th of the month. BLS then blows that sample up to the whole economy. The coverage is big but bounded: farm workers, the self-employed without employees, unpaid family workers, private household staff and the military are all excluded, which is where the awkward word “nonfarm” comes from.

Alongside it runs a completely separate machine: the Current Population Survey, a monthly knock on about 60,000 households’ doors that produces the unemployment rate. The two surveys measure different things — jobs versus people — so they regularly disagree. A person holding two payroll jobs counts twice in the payroll survey but once as employed in the household survey; a gig driver with no employer shows up only in the latter. When headlines say “the two surveys diverged,” this is why.

Why the first print is always a draft

The first estimate published on the first Friday is built on the responses that arrived in time — and late responses are disproportionately from small businesses, whose hiring is more volatile. As stragglers file in, BLS revises each month twice: once in the next report, once in the report after that. The 90% confidence interval around a monthly change is on the order of ±100,000 or more, which means a “+73,000” month is statistically indistinguishable from a slightly negative one. Traders move billions on the first print anyway; the revisions that rewrite it barely ripple through markets — but they do ripple through politics.

Then there is the annual reckoning. Once a year BLS re-anchors the survey to the Quarterly Census of Employment and Wages — near-complete counts derived from the unemployment insurance tax records that virtually every employer must file. This benchmark revision is where the big numbers come from: −818,000 for the year to March 2024, −911,000 for the year to March 2025. They are not admissions of error so much as the designed reconciliation of a fast, noisy survey with a slow, near-complete census. The errors cluster for a reason: at turning points, new firms die and are born faster than the birth-death model can track, so the model overshoots in slowdowns and undershoots in recoveries. That is also why 2026’s preliminary benchmark of just −79,000 attracted so little attention — a stable economy is easier to count.

The political firestorm

Revisions have happened every month for decades, in both directions and under every administration. What changed in 2025 was the reaction. Within hours of the August 1 report, Trump accused McEntarfer — a Senate-confirmed labor economist — of manipulating the data and fired her, a step without modern precedent. The nominee floated to replace her, Heritage Foundation economist E.J. Antoni, was withdrawn on September 30 after bipartisan unease in the Senate, leaving veteran deputy William Wiatrowski as acting commissioner.

The statistical community’s response was unusually blunt. William Beach, who ran BLS during Trump’s own first term, called the firing groundless and warned it set a dangerous precedent for every statistical agency. The deeper worry is structural: the system works because roughly a thousand career staff process surveys under strict confidentiality, insulated from political appointees — and trust in that insulation, once lost, cannot be revised back into existence the way a jobs number can. For anyone trying to read the US economy from abroad, that is the real story of 2025: not that the numbers move, but that the machinery moving them suddenly became the news. The same pattern — a dry policy instrument turning into a headline — repeated a year later when record diesel prices pushed the G7 into another emergency oil reserve release, and again when Beijing recast a decade-old gender-test smuggling trade as a biosecurity story: over 100,000 vials of pregnant women’s blood became a national security headline.

Key facts

Headline source

CES establishment survey

Sample size

~121k firms, 631k worksites

Reference point

Pay period incl. the 12th

Unemployment rate

CPS, ~60k households

Monthly revisions

2 per estimate

Benchmark anchor

QCEW tax records, yearly

What people are saying

Common questions

Why do revisions go down more often than up lately?

Revisions are symmetric over full cycles, but they bunch at turning points. In slowdowns, struggling small firms respond late and the birth-death model keeps crediting the economy with jobs from hypothetical new firms that no longer exist — so corrections cluster downward. In recoveries the same machinery undershoots and revisions turn positive.

Does a big revision mean the data was rigged?

No. Revisions are published on a fixed schedule, in both directions, by career staff using documented methods — the benchmark process exists precisely to catch and correct drift. The −911,000 preliminary benchmark is the system working as designed: reconciling a fast survey against near-complete tax records.

If the number is so uncertain, why do markets care?

Because it is the fastest comprehensive read on the US economy, and the Fed reacts to it. The first print moves bond yields and rate expectations within seconds; the smarter money then watches the revision trend, which is often the better signal of where the cycle is turning.

Could the US just wait for the tax data instead?

It would be accurate but useless in real time: QCEW records arrive with a lag of five months or more. The jobs report is the classic trade-off — a quick, noisy estimate now, reconciled against slow, solid counts later. Every advanced economy makes some version of the same bargain.