The number that led the news in June is not the number now
On 7 August the Bureau of Labor Statistics published the July employment situation. The headline was quiet, and the agency said so itself: both nonfarm payroll employment, at -23,000, and the unemployment rate, at 4.1%, "changed little" in July. Government employment fell 53,000, with local government education down 50,000 after little net change over the prior twelve months. Health care kept rising, by 22,000, though the release is careful to add that this is slower than its 36,000 average monthly gain over the same twelve months.
Because government fell 53,000 while the total fell 23,000, private payrolls rose by about 30,000. That subtraction is ours rather than the agency’s, and it says the headline decline is a government-sector line, most of it a summer break in school payrolls.
Further down the same release sits the paragraph this edition is about. The change for May was revised down by 66,000, from +129,000 to +63,000. June was revised down by 37,000, from +57,000 to +20,000. In the agency’s words: "With these revisions, employment in May and June combined is 103,000 lower than previously reported."
Follow May backwards and it has been three different numbers. It was first published on 5 June as +172,000. On 2 July it became +129,000. On 7 August it became +63,000. Each correction arrived inside a release about a later month, under a headline about that later month, and the version of May that most people carry in their heads is the one from 5 June.
| Release | What it said about May | What it said about June |
|---|---|---|
| 5 June 2026 | +172,000 | — |
| 2 July 2026 | +129,000 | +57,000 |
| 7 August 2026 | +63,000 | +20,000 |
103,000 is an ordinary revision, and the alarming version would be the easier piece to write
The agency publishes the history of its own revisions, which makes this checkable rather than a matter of tone. Since 2003 the mean absolute revision is 33,000 from the first estimate to the second, 34,000 from the second to the third, and 51,000 from the first to the third. May’s step this month was 66,000, about twice the average for that step. June’s was 37,000, essentially the average. Two typical steps sum to 67,000, so 103,000 is around one and a half times a normal two-month revision. Larger than usual. Nowhere near extraordinary.
The second yardstick is blunter. The technical note to the same release puts the 90% confidence range on a single month’s payroll change at plus or minus 122,000. The entire two-month revision is smaller than the sampling noise around any one month’s number. If that sentence deflates the story, it is doing its job.
Revisions also run in both directions, and this year has one of each. The 5 June release raised March and April by 93,000 combined. The 2 July release lowered April and May by 74,000. The 7 August release lowered May and June by 103,000. April alone went from +115,000 to +179,000 and then back to +148,000. The upward revision of 93,000 got no headline either, which is the cleanest evidence that what is asymmetric here is attention rather than the statistics.
For scale over a longer window: twelve months earlier, the release of 1 August 2025 cut May and June 2025 from +144,000 and +147,000 to +19,000 and +14,000, leaving those two months 258,000 lower, two and a half times this month’s figure. This desk is not going to call an ordinary revision a scandal in order to have a stronger opening paragraph.
One number is a sample. The other is a tax record.
The monthly figure is a survey. It draws on about 119,000 businesses and government agencies covering roughly 622,000 worksites, which is around 26% of all nonfarm payroll jobs, and it adds a model for firms too new to have been sampled at all. The second and third estimates exist because late responses keep arriving and seasonal factors are recalculated. The revision is not an error being caught. It is a series doing what a series does, on a one-month clock instead of an annual one.
Once a year that sample is re-anchored to something built differently: the Quarterly Census of Employment and Wages, assembled from state unemployment insurance tax records that employers are required to file. It covers about 97% of the employment within the survey’s scope. Not all of it. The remaining 3% is constructed from other sources, primarily Railroad Retirement Board records and Census Bureau data, and one of those, the Annual Survey of Public Employment and Payroll, is itself a survey. "Tax records rather than a survey" is right in shape and wrong by three points, and the three points are worth printing.
The agency goes further than most of its readers do. It states that the benchmark represents the difference between two employment estimates derived from separate statistical processes and therefore reflects the sum of the errors present in each. The tax records have their own error. Anyone treating the benchmark as the true number against which the survey is graded is making the same mistake this piece objects to, one layer up.
| The question | The monthly estimate | The annual benchmark |
|---|---|---|
| What is actually counted | A sample of about 119,000 businesses and government agencies, roughly 26% of all nonfarm payroll jobs, plus a model for firms too new to be in the sample. | Unemployment insurance tax records covering about 97% of the employment in scope. The remaining 3% is built from other sources, one of which is itself a survey. |
| How often it changes | Three times. A first estimate, then a second and a third as late responses arrive and seasonal factors are recalculated. | Once a year, referenced to a single month of March, and at the preliminary stage only for major industry sectors. |
| How wrong it can be | The published 90% range on one month’s change is plus or minus 122,000, wider than most monthly moves that get reported as news. | No equivalent range is published. The agency describes it as the difference between two statistical processes, carrying the errors of both. |
| What it changes when it lands | The published number, in the next release and the one after that. | Nothing, on 28 August. Official estimates are not updated on the preliminary revision. The final one enters the series in February 2027. |
The large correction arrives on 28 August, and it will not change the number
At 10:00 ET on 28 August the agency publishes the preliminary estimate of the annual benchmark revision, referenced to March, on the same day it issues first-quarter 2026 data from the tax records. Then comes the sentence almost nobody quotes: "Official establishment survey estimates are not updated based on this preliminary benchmark revision." The final revision enters the published record with the January 2027 employment situation, in February 2027. So for roughly six months the better count will say one thing and the official series will keep saying another, and the official series is the one that gets quoted.
The preliminary is itself a draft. The March 2025 benchmark was published on 9 September 2025 at -911,000 and finalised on 11 February 2026 at -898,000 seasonally adjusted, or -862,000 on the not-seasonally-adjusted basis. The year before, the March 2024 preliminary was -818,000 and the final was -589,000, a pair we have from secondary sources rather than from agency text we could retrieve: 28% of that correction evaporated between the two versions. And 2025 should not be called the largest downward revision on record. On the agency’s own consistent basis its 2009 benchmark article carries a larger cut; only the preliminary beat 2009, and comparing a preliminary with a final is precisely the error this edition exists to name.
Why does a correction approaching a million jobs never feel like one? Because it is spread. The agency applies a wedge-back procedure that assumes the estimation error accumulated at a steady rate since the last benchmark, so a level cut of 898,000 surfaced as twelve ordinary-looking monthly changes: 159,000 off January, 6,000 off May, 8,000 off July, and 33,000 added back in October. Not one of them was large enough to be a story on its own. The correction was real, it was published, and it had no day.
Nor is the benchmark always a cut. March 2022 was +506,000 and March 2017 was +135,000; two of the last ten were upward, and the agency puts the ten-year average absolute revision at 0.2% of total nonfarm employment. This desk is making no forecast of the size of the 28 August figure. It does not exist yet.
How to read a number that is still moving
If you are hiring or looking for work, the operational rule is that a single month’s payroll change is not evidence of anything on its own. The uncertainty band around it is wider than most of the moves reported as turning points. The flows underneath it barely moved: the June JOLTS release, published 4 August, put job openings at 7.4 million, hires at 5.3 million and total separations at 5.4 million, described as little changed, unchanged and little changed. A market in which 5.3 million people are hired in a month is poorly summarised by a net figure of -23,000, which is the small difference between two very large numbers.
Three dates do the work from here. 28 August: the preliminary benchmark, which changes nothing official. 4 September: June gets its third and final sample-based estimate, which means the 103,000 in today’s piece is not final either. February 2027: the benchmark is written into the series. The number that corrects the number is also a draft, and the desk that tells you otherwise is selling certainty it does not have.
This edition contains no figure from our own data. We hold nothing that measures American payroll employment and would not publish it if we did. What we do have is a small version of the same problem: our own advertised remote-work share moves with how many sources we read, which is why every city page prints its source count next to its number. The thing that makes a public statistical agency worth reading is that it publishes the history of its own revisions and the width of its own error band. The part of that discipline anyone can copy is printing the denominator.
Our own numbers carry the same warning, so we print what they were built from. Every city page at babzituna.com/jobs shows the number of sources behind it, which is the one column that tells you when a figure is about us rather than about the market.