The monthly payroll number that moves headlines is an estimate built from a partial sample, and it isn't final until the following winter. Each February, the Bureau of Labor Statistics realigns its payroll series to near-universal tax records — and in the release published February 11, 2026, that step cut March 2025 employment by 898,000 on a seasonally adjusted basis.
That is not a scandal. It's the design.
The establishment survey behind the monthly jobs report is fast on purpose, and speed costs precision. Understanding where the imprecision lives — and when it gets corrected — changes how much weight a single month deserves.
Why does the first jobs number change at all?
Because the first estimate comes from a sample, not a census. The Current Employment Statistics survey covers about 119,000 businesses and government agencies across roughly 622,000 worksites, according to the program's technical notes — about 26 percent of total nonfarm payroll employment, drawn from a frame of some 12.1 million unemployment insurance tax accounts.
Twenty-six percent is a large sample by survey standards. It is still not everyone.
Two things then happen to that first print. Reports keep arriving after the initial deadline, which is why each month gets revised twice in the two releases that follow. And once a year, the sample-based estimate is replaced at a single reference month by counts from the Quarterly Census of Employment and Wages, the near-complete tally of jobs covered by unemployment insurance.
That annual replacement is the benchmark. Its reference month is March, and BLS publishes it the following February alongside the January estimates, then carries the correction back to the prior April and forward to the current month.
What is the birth-death model actually doing?
It is filling a hole the sample can't see. New businesses open continuously but reach the survey frame only after a lag, and firms that close often stop reporting rather than reporting a zero. Without an adjustment, the estimate would systematically miss both.
BLS handles this in two parts, per its published description of the model. Employment losses from business deaths are excluded from sample-based estimation, so the trend from surviving firms stands in for the closures — on the reasoning that births and deaths partly offset each other. Whatever net employment change from business formation and closure remains is then forecast with an ARIMA time-series model fitted to the past five years of actual birth-death residuals measured in the QCEW microdata.
So the adjustment is a forecast built on recent history. That is exactly the design that struggles when history stops being a guide — when formation slows or closures accelerate faster than the last five years imply.
The method has been tightened. Model inputs have updated quarterly since the January 2011 benchmark, and as of the February 2026 release the birth-death model incorporates current sample information each month rather than running on fixed annual forecasts.
How big is a normal benchmark revision?
Small. Over the 10 years before the latest benchmark, the absolute percentage benchmark error at the total nonfarm level averaged 0.2 percent, with a range from less than 0.05 percent to 0.4 percent, BLS reports in its technical notes for the national benchmark. On a payroll base above 150 million, 0.2 percent is a few hundred thousand jobs spread across a year.
That average is the yardstick worth carrying. It tells you a benchmark revision of a couple hundred thousand is routine housekeeping, and one several times that size is a signal about the sample and the model, not just noise.
What did the March 2025 benchmark change?
It was one of the large ones. The preliminary estimate published in September 2025 pointed to 911,000 fewer jobs added over April 2024 through March 2025 than previously reported, with leisure and hospitality down 176,000, professional and business services down 158,000, and retailers down 126,000, PBS NewsHour reported on September 9, 2025.
The final figure landed close to it. The Employment Situation release for January 2026, published February 11, 2026, put the March 2025 benchmark revision at 898,000 seasonally adjusted, or 862,000 on an unadjusted basis — a decline of 0.5 percent.
Same release, ordinary business elsewhere: payrolls up 130,000 in January 2026, unemployment at 4.3 percent, health care and social assistance adding jobs while federal government shed 34,000.
How do the people who read this data for a living handle it?
They discount the first print and wait. Asked about the preliminary benchmark at the September 17, 2025 FOMC press conference, Federal Reserve Chair Jerome Powell said the revision "was almost exactly what we expected," adding: "It was amazing how close the expectation was."
His account of why early months are soft is about collection, not conspiracy. "The response rate is quite low for the first month, or lower for the first month," Powell said in the published transcript. "By the time you get to the second or third month … you get to the place where the data are much more reliable by the second, and certainly the third, month." His verdict on the series overall: "The data we get is still well good enough for us to do our work."
Response rates are a live variable, not a constant. In the January 2026 report, BLS noted household survey collection fell to 64.3 percent, affected by severe winter weather — a reminder that measurement conditions vary month to month.
What does this mean for reading a jobs report at work?
Mostly, it means matching your confidence to the vintage of the number. A first print is a fast estimate with known gaps; a third print is firmer; a post-benchmark figure is anchored to tax records. Three practical habits follow.
- Check which vintage you're quoting. A month that has been through two revisions carries more information than one released last Friday.
- Read the trend, not the month. Sampling error and birth-death forecast error both wash out faster across quarters than across single months.
- Note the benchmark date. Anything citing payroll growth for a period before the last March reference month should reflect the benchmark, or it's quoting a superseded series.
None of this is a forecast, and none of it is investment or career advice. It's a description of how one widely quoted number is produced — which is usually the more durable thing to know.
For a related careers perspective, read Employers dropped the degree requirement — hiring barely moved.