July's jobs number went from −23,000 to +21,000. That is not a rounding error.
The Bureau of Labor Statistics revised two months of hiring up by a combined 55,000. One of them changed sign, and it is the figure this publication built an article around.
- What happened
- In the August employment report, published September 4, the Bureau of Labor Statistics revised June's job gain up by 11,000 and July's up by 44,000. July moved from a reported loss of 23,000 jobs to a gain of 21,000.
- Why
- The monthly payroll figure is a survey estimate. Employers report late, and the Bureau revises the number twice as more of them do. A first print is the earliest estimate, not the measurement.
- The numbers that matter
- −23,000July, as first reported+21,000July, as revised+55,000Combined June and July revision+162,000August payrolls
- Why it matters
- Decisions get made on first prints — by markets, by commentators, and by this publication. Knowing how much a first print can move is the difference between reading the data and being led by it.
- What happens nextas we saw it on 21 Sep 2026
- The September employment report is scheduled by the Bureau of Labor Statistics and will carry its own revisions to July and August.
July was reported as a contraction and revised into a gain. The direction changed, not just the size.
| August 2026 | 162 |
|---|---|
| June, revised | 31 |
| July, revised | 21 |
| July, first print | -23 |
On August 7 the Bureau of Labor Statistics reported that the American economy had lost 23,000 jobs in July. That number was read as evidence of a labor market rolling over, and it was quoted widely — including here.
On September 4 the same agency revised it. July did not lose 23,000 jobs. It gained 21,000.
June was revised from +20,000 to +31,000, an increase of 11,000. July was revised from −23,000 to +21,000, an increase of 44,000. The Bureau reports the combined effect as 55,000 higher than previously reported.
Why the number moves
The payroll figure comes from a survey of employers. Not all of them respond by the time the first estimate is published, so the Bureau publishes what it has and updates the number in each of the following two months as more responses arrive. Revisions are the system working, not failing.
What is unusual here is the sign change. A revision of 44,000 on a figure near zero is enough to turn a contraction into an expansion, and a month that was described as job losses becomes a month of modest hiring.
The size of a typical revision does not shrink because the headline happens to be small. When the reported number is close to zero, an ordinary revision is large enough to flip it, which means the least informative first prints are exactly the ones that generate the strongest headlines.
What we got wrong
Our August 17 article on the economy was built around the July figure as first reported. Its central claim was that payrolls had fallen while inflation stayed elevated. Half of that pairing no longer holds: payrolls did not fall in July.
We have not rewritten that article. It described the evidence available when it was published, and quietly editing it would destroy the record of what we actually argued. It now carries a note at the top pointing here.
Sources
- The Employment Situation — August 2026U.S. Bureau of Labor Statistics · Sep 4, 2026 · Primary source
- The Employment Situation — July 2026U.S. Bureau of Labor Statistics · Aug 7, 2026 · Primary source
Related analysis
The economy is sending two signals at once. The Fed is listening to the louder one.
Payrolls fell. Retail sales fell. Inflation held at 3.4%. And three Federal Reserve officials voted to raise interest rates. Here is how those facts fit together.
August added 162,000 jobs. The average of the prior twelve months was 31,000.
The strongest hiring month in a year sits on top of a twelve-month average five times smaller. Both numbers are true, and only one of them is a trend.
The Fed stopped arguing with itself, and raised rates
In July three officials wanted higher rates and lost. On September 16 the vote was 12–0. Nothing about the committee changed — the data did.