Why payroll numbers are revised after the first report

The scenario below is fictional. Its figures are invented to show the reasoning, not to describe a real month.

At 8:30 a.m., a fictional newsroom receives the August employment report. Payrolls are up 92,000, unemployment has moved from 4.0 to 4.2 percent, and June plus July have been cut by 118,000. The desk must explain what those numbers show now, and why some may change again.

The first intake

An initial payroll estimate can change because it is provisional, later reports add information, and annual benchmarking can re-anchor the series.

In this scenario, the desk first records the August release as published, keeping a copy of its tables so later changes can be distinguished from what was known at the time.

The assignment is to report the figures without treating them as a final verdict on the labor market. The desk notes that the latest payroll number is preliminary and that the reductions to June and July affect the recent baseline. It also flags a common reporting trap: letting a first estimate read like settled history.

Why the first number is the least complete

The payroll figure comes from the BLS Current Employment Statistics survey of employers. On release day, only part of the sample has reported, so the first estimate rests on incomplete information. The estimate is revised in each of the next two months as more employers respond.

Each year the series is also benchmarked to unemployment-insurance tax records, which cover nearly all employers. How large the changes are varies from month to month, so the desk checks the revision table that BLS publishes before it calls any revision unusual. In the invented case above, June and July fell by 118,000 together because late reports pulled both months down.

How large revisions have been in practice

Month to month, revisions are usually modest. BLS data show the third estimate of the monthly payroll change differs from the first by an average of about 51,000 jobs (seasonally adjusted) since 2003. The 2024 average was 48,000, while 2020 and 2021 reached 130,000 and 181,000 during the pandemic.

The annual benchmark can be much bigger. Over the prior 10 years, BLS says it averaged 0.2 percent of total nonfarm employment in absolute terms, ranging from under 0.05 to 0.4 percent. The March 2025 benchmark was larger: the not seasonally adjusted level fell by about 861,000, or 0.5 percent.

Testing the headline against the sources

The first check is whether the draft joins measures that describe different things. Payroll employment comes from an establishment survey counting jobs, while unemployment comes from a household survey classifying people by labor-force status. They are separate readings, not components of one calculation.

The reporter checks the household survey distinction and the BLS Employment Situation methods. A person with two covered payroll jobs can appear twice in the establishment count and once in the household survey. The desk therefore avoids presenting the employment change and the unemployment-rate movement as if they were directly interchangeable.

A hurried draft says the economy added 92,000 jobs as unemployment suddenly jumped. The desk rejects that phrasing because it hides the revisions and links two separate measures without explaining their relationship. It returns to the source tables and follows a release-reading routine before deciding what the headline can support.

Rebuilding the recent baseline

The desk finds that the earlier estimates have changed enough to alter the recent average. It compares each month's first payroll estimate with its current estimate, while marking August as unrevised. That comparison shows why the newest monthly total should not be read without the history around it.

Using the current June and July estimates with August gives an average of 80,333, rather than the 119,667 the first estimates imply. The desk labels August preliminary and calculates from the current values, rather than carrying over last month's average. Its guide to developing news stories helps keep the change in the record clear.

Identifying what changed

The desk treats monthly revisions, annual benchmarking, and seasonal adjustment as distinct processes. Monthly estimates can shift when more sampled employers respond; benchmarking re-anchors the series to more complete administrative counts; seasonal adjustment can restate earlier adjusted values.

The Current Employment Statistics revision FAQ says initial monthly estimates are revised during the next two months as late sample reports arrive. It also describes annual benchmarking, based mainly on unemployment-insurance records, and notes that seasonally adjusted estimates may keep changing for years. The desk names the process behind each change instead of using revision as a catch-all label.

Filing the story and setting the next check

After checking the tables and survey definitions, the desk uses a measured headline: payroll growth slows to 92,000 in August as earlier gains are revised lower. It reports the unemployment-rate movement separately, without implying that the two figures come from one count.

Before calling the rate change a surge, the reporter checks household-survey levels, labor-force participation, the employment-population ratio, and published precision information. The desk considers whether the labour force grew faster than employment or whether sampling variation could explain the movement.

The desk reviews employment by industry, average weekly hours, and average hourly earnings. It checks whether a strike or severe weather disrupted collection during the reference period, and avoids implying that every worker received the published average pay increase. A change in worker mix can affect that average, as the indicator comparison distinction helps explain.

Finally, the desk records the next monthly release dates, the preliminary benchmark date, and the final benchmark publication. It keeps the August table on file. If later revisions alter the picture, the live explainer gets a dated update, while the original alert remains a record of the information available when it was published.

Rules the desk carries forward

The scenario leaves the desk with a few practical rules: preserve the first version, keep payroll jobs distinct from household employment measures, recalculate averages from current estimates, and identify whether a change is a monthly revision, benchmark, or seasonal recalculation.

A downward revision by itself does not prove that the agency made an error. Later employer responses and more complete records can move a sample estimate; a correction for an identified error is a different event and should be labeled accordingly. When describing a rate change, keep its current value, prior value, and relevant context together.

For desk reference, the CES Handbook of Methods and CPS Handbook of Methods sit alongside the Employment Situation technical note and the explanation of how the government measures unemployment. Related checks can use CES benchmark information, seasonal adjustment methods, the Quarterly Census of Employment and Wages overview, and labor-force participation definitions.