Why US jobs numbers get rewritten — the revision machinery
Payroll figures are survey-based estimates revised twice as late responses arrive — and consistent downward revisions are an economic signal
The three lines
- The first payroll print uses early survey responses only; two monthly revisions follow
- This round erased 103,000 jobs from May and June combined
- Consistently one-directional revisions signal the economy is outrunning the statistics
Key questions
- What are payroll revisions
- Retroactive corrections to the monthly jobs number. The first release — the one that makes headlines — counts only businesses that responded by the survey deadline. Late responses are folded in over the next two months, and once a year a benchmark revision resets the whole series against near-census tax records.
- Why are the revisions so large
- Low initial response rates. A large share of surveyed businesses miss the first deadline, and when late responders differ systematically from early ones the correction grows. At economic turning points the bias is predictable: closing businesses stop responding at all, making initial samples look healthier than reality — so downturns breed chains of downward revisions.
- How should revisions be read
- Direction over size. Several consecutive months of downward revisions mean the labor market is weaker than each headline suggested — that is the signal, not noise. This week's case: with 103,000 jobs revised away from May and June, July's -23,000 reads not as a shock but as confirmation of a slide already underway.
The scariest number in Friday's jobs report was not -23,000. It was the rewrite: May's gain cut from 129,000 to 63,000, June's from 57,000 to 20,000 — 103,000 jobs erased from months everyone thought they understood. A labor market described as "holding up" had in fact stalled some time ago. This reference explains why US jobs numbers change after publication, by how much, and what the changes themselves tell you.
1. Why the numbers move — responses that arrive after deadline
The payroll survey covers about 120,000 business establishments, and its problem is time. The first release — the headline number — tabulates only the responses that arrived by the initial deadline. Many businesses answer late. So the Bureau of Labor Statistics rewrites each month's figure twice, in the two following months, as stragglers report. Separately, once a year, a benchmark revision resets the entire series against unemployment-insurance tax records — near-census data that arrives with a long lag.
Revisions swell at turning points, with a predictable bias: businesses that are shutting down tend to stop answering surveys entirely, which makes the initial sample look healthier than the economy it samples. Downturns therefore breed chains of downward revisions — the pattern this week's data uncomfortably resembles.
2. The machinery at a glance
| Stage | Timing | What happens |
|---|---|---|
| First print | Month +1, first Friday | Early responses only — the headline |
| First revision | +2 months | Late responses folded in |
| Second revision | +3 months | Final monthly figure |
| Benchmark | Annual | Full reset against tax records |
| This round | May -66k, June -37k | 103,000 erased combined |
Two rules for reading. First, the headline is a draft — markets trade the first print, but trend judgments belong to revised data. Second, direction is the signal — when revisions run one way for months, the statistics are chasing an economy moving away from them, and the real economy is usually cooling faster than the numbers admit.
3. What remains
The revisions redrew America's labor-market picture from "suddenly bad in July" to "cooling since late spring." The next checkpoint: which direction July's -23,000 gets revised next month. The report's basic structure is in "The US jobs report, explained"; Friday's full damage in "America lost jobs in July"; the unemployment-rate illusion in "How the unemployment rate is measured." This document updates at the annual benchmark.