How the Jobs Report Is Built, and Why Its First Number Moves

Each month's jobs report lands with a single headline number for how many jobs the U.S. economy added or lost. That number is an estimate built from a survey that is still collecting responses when the report is published, which is why it almost never stays the same.

The Bureau of Labor Statistics revises the two prior months' payroll figures every time it releases a new report, then publishes a preliminary estimate of that reset each summer before applying it to a full year of data the following February against a more complete count drawn from state tax records. Understanding that process is the difference between reading one month's number as news and reading it as a first draft.

What the Survey Actually Counts

The headline payroll figure comes from the Current Employment Statistics survey, which the BLS describes as covering about 119,000 businesses and government agencies representing roughly 622,000 individual worksites. That sample represents close to a quarter of all nonfarm payroll jobs in the country; the rest of the total is estimated statistically rather than counted directly, according to the BLS.

Employers are asked to report how many people were on their payroll during the pay period that includes the 12th of the month, a reference date the BLS's technical note notes may or may not line up with a calendar week. Data comes in through web reporting, electronic data interchange, phone interviews and other collection methods, per the same technical note.

Filling the Gap Businesses Don't Report

New businesses that haven't yet been added to the BLS's sampling list don't show up in the survey right away, and closed businesses stop responding. According to the BLS, a new establishment typically cannot appear in state unemployment insurance records for seven to nine months after it opens, and the CES survey generally cannot sample it for about a year. To bridge that gap, the BLS uses what it calls a net birth-death model.

The model has two parts. The first assumes that in most months, job losses from business deaths are roughly offset by job gains from births the survey hasn't caught up with yet, using an indirect imputation method. The second is a statistical forecasting model, built on an auto-regressive integrated moving average approach, that estimates whatever birth-and-death employment isn't captured by that offsetting assumption, drawing on five years of historical data from unemployment insurance records. The BLS has measured how far off that forecast can run: for the 12 months from April 2024 through March 2025, the agency found the actual net birth-death figure came in about 117,000 below what the model had forecast. Starting with the preliminary January 2026 estimates, the BLS said it modified the forecasting component to also incorporate current sample information, a change it applied retroactively to data back through April 2025.

Why the First Reading Almost Never Holds

The BLS labels the two most recent months in every report preliminary because those estimates are based on incomplete returns. A figure becomes final only after two more rounds of revisions, once nearly all sample reports for that month have come in.

That is not a small caveat, and it has grown larger as fewer employers respond on time. The research and journalism outlet Journalist's Resource, in a review of BLS data, reported that the response rate for the establishment survey fell from about 60 percent in January 2020 to about 43 percent in March 2025. The consulting firm Edgeworth Economics separately found that the rate at which newly identified establishments are added to the sample fell from more than 70 percent in 2015 to below 40 percent over the past year, a trend it linked to the likelihood of larger future revisions.

The effects show up repeatedly in the data. Edgeworth Economics noted that the BLS's August 2025 jobs report carried a revision of roughly 258,000 fewer jobs than first reported for June 2025, with state and local government education employment accounting for more than 109,000 of that change. A year later, the August 2026 report showed the same dynamic on a smaller scale: July 2026 payroll employment was first reported as a decline of 23,000, then revised up by 44,000 to a gain of 21,000, while June 2026 was revised up by 11,000, to a gain of 31,000, according to the BLS's own release.

The Annual Reset Against Tax Records

Beyond the two-month revision cycle, the BLS recalibrates a full year of estimates annually through a process it calls benchmarking. Each March, the BLS replaces its sample-based employment level with a count drawn from the Quarterly Census of Employment and Wages, built from unemployment insurance tax records that cover about 97 percent of nonfarm employment, with the remainder filled in from Railroad Retirement Board, Census Bureau and other administrative records.

Rather than rebuilding every prior month from scratch, the BLS uses what it calls a wedge-back procedure: the gap between the old sample-based March level and the new administrative count is divided by 12, and that fraction is added incrementally to each of the preceding months, with the full adjustment applied to March itself. Months after the new March benchmark are instead built forward from the new level using the survey's sample-based rate of change, alongside freshly recalculated birth-death forecasts for the coming year.

How Big Recent Benchmark Revisions Have Been

The size of that annual reset has varied sharply in recent years. For the 12 months through March 2024, the BLS's preliminary benchmark showed 818,000 fewer jobs than previously reported; when finalized in February 2025, that figure was revised to 598,000 fewer jobs.

The following cycle was larger still. The preliminary benchmark for the 12 months through March 2025, released September 9, 2025, put total nonfarm employment 911,000 lower than previously estimated, the largest downward revision since 2002, according to HR Brew's review of the release. When the BLS finalized that number in February 2026, the figure settled at a downward revision of 898,000, per the BLS's own account of the benchmarking process. By contrast, the preliminary benchmark released August 28, 2026, for the 12 months through March 2026, showed total nonfarm employment down by just 79,000, the smallest preliminary benchmark revision since 2021, according to the American Action Forum. The BLS notes that benchmark revisions have averaged about 0.2 percent of total nonfarm employment over the past decade, a scale the 2024 and 2025 cycles each exceeded by several times.

Reading a Single Month's Number

Taken together, these layers mean the headline number in any given jobs report is one estimate in a sequence that will be revised at least three more times before it's considered close to final: twice in the following two months' reports, and again the following year when the benchmark comes out. The payroll survey is only one half of the monthly release, alongside a separate household survey used to calculate the unemployment rate, but the payroll figure is the one that draws the initial revisions described here. A single month's payroll number carries real information, but the BLS's own revision history points in one direction: the figure that moves markets on a Friday morning is not the figure that will still be on the books a year later.

Related coverage: What the Jobs Report Measures, and Why Its Two Surveys Disagree.