The Bureau of Labor Statistics publishes two employment numbers on the first Friday of each month. They come from two different surveys, measure two slightly different things, and routinely disagree by enough to flip the narrative of a labor report. The establishment survey produces nonfarm payrolls (the headline number). The household survey produces the unemployment rate, the labor force participation rate, and a second employment count that markets mostly ignore until it stops moving in the same direction as the first one. When the two surveys agree, the report reads clean. When they disagree by more than 200,000 in a month, or by more than a million over a year, the household survey is the one worth reading carefully.
The structural differences are mechanical. The establishment survey (formally the Current Employment Statistics program) samples roughly 119,000 business establishments covering about 631,000 worksites and asks how many jobs were on payroll during the reference week. It is a count of jobs, not workers. A person holding two payroll jobs is counted twice. Self-employed workers, unpaid family workers, household-domestic workers, and most agricultural workers are not counted at all. The series covers about 80 percent of total nonfarm employment by headcount but only the W-2 portion of it. The household survey (the Current Population Survey) samples about 60,000 households and interviews adults about their employment status during the same reference week. It is a count of workers. A person holding two jobs is counted once. The self-employed, unpaid family workers, agricultural workers, and household-domestic workers all appear in the count.
The two surveys also handle new businesses differently, and this is where most of the gap shows up over time. The establishment survey cannot sample firms that did not yet exist when the sample frame was drawn, so the BLS adds a statistical adjustment called the birth-death model to estimate jobs created at firms too new to appear in the survey. The adjustment runs through a non-seasonally-adjusted ARIMA model that lags actual business formation by 18 to 24 months. When business formation is accelerating, the birth-death model under-counts jobs and the establishment survey reads low. When business formation is decelerating off a peak, the model over-counts and the establishment survey reads high. The household survey has no equivalent adjustment because it samples people directly. When a worker takes a job at a new firm, the household survey picks it up on the next interview cycle.
This is the gap that opened through 2024. The establishment survey reported roughly 2.9 million net jobs added across the year. The household survey reported roughly 100,000. The two prints implied two completely different labor markets: a tight one still hiring, and a flat one that had already stalled. The benchmark revision released in February 2026 cut the prior-year establishment-survey total by 818,000 jobs, the largest downward revision in the modern history of the program. The benchmark revision did not adjust the household survey at all, because there was nothing to adjust. The household survey had been reading the cycle correctly the entire time. Markets had been pricing the establishment-survey print, which is part of why fed funds futures whipsawed when the benchmark landed.
The pattern repeats at every turning point. The household survey led the establishment survey by four months into the 2008 recession and by three months into the 2001 recession. It led at the 2010 and 2003 recovery points by similar windows. The reason is methodological: business formation and contraction are leading indicators, the birth-death model is a lagging indicator, and the household survey captures the underlying flow without the lag. The trade-off is that the household survey has a much smaller sample (60,000 households versus 119,000 establishments), so its month-to-month sampling error is larger. The 90 percent confidence interval on the monthly household employment change is roughly plus or minus 600,000. The same interval on the establishment-survey change is roughly plus or minus 130,000. In any single month, the establishment print is more precise. Over rolling six-month windows, the household survey is more accurate, because the precision advantage of the establishment survey gets eaten by birth-death misestimation.
The practical rule for reading any NFP morning: use the establishment-survey print for the level read on payroll demand and for short-term wage and hours data. Use the household survey for the direction read on the cycle, for self-employment and gig-economy shifts, and for any month when the two surveys disagree by more than 200,000. When they agree, the report is a clean read. When they disagree, the household survey has the better historical track record at the points where the disagreement actually matters. The unemployment rate, which comes from the household survey, is the line the Fed Chair has cited most often in post-FOMC press conferences across the last three cycles. There is a reason for that, and it is not that the Chair finds the household survey more interesting. It is that the household survey is the one that has been right at the turns.
A 100,000-job gap in any given month is noise. A persistent 200,000-job gap that runs for three or more consecutive months, in the same direction, with the household survey weaker, is the configuration that has preceded every modern downturn. The gap to watch through this summer is whether the household-survey employment count keeps printing flat or negative while the establishment survey holds positive. If it does, the establishment survey is the one that will get revised, not the other way around.