The economy got quieter than the jobs survey can hear

Nonfarm payrolls changed little in July, down 23,000, against an average monthly gain of 34,000 over the prior twelve months. In the July release, BLS cut May’s number by 66,000, from a previously published +129,000 down to +63,000, and cut June’s by 37,000, from +57,000 to +20,000. Put the two together and BLS’s own figure is stark: employment in May and June combined stands 103,000 lower than previously reported.

The monthly jobs number has not got worse. The economy got quieter than the instrument can hear. BLS’s own long-run mean absolute revision, first print to final, is 57,000 jobs; its own stated average monthly job gain, a signed figure, over the year to July is 34,000. One of those is a revision record BLS has kept since 1979, the other a twelve-month hiring average, and putting them next to each other only works if both windows stay named, which they now are. The error bar did not grow. The signal fell inside it.

That is why December could be published as a gain of 50,000 jobs and, two revisions later, settle at a loss of 17,000 — a swing of 67,000 that crossed zero. Nobody who reacted to Friday’s headline could have known, in real time, whether the month underneath it was good or bad. Neither could BLS.

Seven months of revisions

BLS publishes this itself. Its revisions table, updated with every release, carries each month’s first, second and third estimates and the difference between the first and the third. What follows is that record, with each first print linked to the release it appeared in. What BLS has not published is a 2026 average: the row sits on the same page, and it is blank until the year closes.

Month	First print	Stands now	Net
December 2025	+50,000	-17,000	-67,000
January 2026	+130,000	+160,000	+30,000
February 2026	-92,000	-156,000	-64,000
March 2026	+178,000	+214,000	+36,000
April 2026	+115,000	+148,000	+33,000
May 2026	+172,000	+63,000	-109,000
June 2026	+57,000	+20,000	-37,000, one revision still owed

Of the six months whose revision cycles are closed, three finished net higher than their first print and three finished lower; June, still open, is not counted either way. December flipped sign outright.

Average the five 2026 months among those six and the mean absolute revision is 54,400 jobs, against a mean signed monthly change of 85,800. BLS’s own 2026 average stays blank on that page until the year closes; this is the part of it that has already settled. It sits inside BLS’s own long-run band of 51,000 to 57,000. The noise did not get bigger. There is no slope here to point at, only a series bouncing around a number BLS already tells you to expect.

Where the missing 21 points go

Take one month apart to see why revisions happen at all. When BLS published December’s first estimate, it had collected 66.8 percent of what it calls the active sample: the employers it still expects to hear from, refusals and dead firms already removed. Two releases later, when the book closed on December, that collection rate had reached 87.9 percent. So 21.1 percentage points of the active sample arrived after the number had made the news.

The Government Accountability Office measured the same survey on the same basis: in 2024, responses rose on average from about 60 percent at the first release to roughly 90 percent by the third.

That gap is not new, and it is not shrinking on its own. Widen the denominator and it gets worse. Response rates, which keep refusals in and which BLS reads at the third release, the same moment that 90 percent describes, fell from 62 percent to 42 percent between 2015 and 2025, on GAO’s count. The survey hears from nine in ten of the firms still willing to talk to it, and from fewer than half of the firms it asked. Businesses are answering a voluntary government survey less often than they used to. BLS did not choose that.

The case that this is fine

None of this is hidden. The table above is BLS’s own, and the summary of mean revisions on the same page cuts against a one-directional story: the mean signed revision since 1979 is +11,000, slightly upward, not down. A professional who trades on this data does not read the first print. She reads the three-month average, which smooths exactly the noise described above. That is a real defense, and by the numbers a fair one.

It has one hole. The three-month average is built from the same first prints that later get revised — the same instrument, taken three times instead of once, not a different one. The people who reprice a mortgage curve or vote on a rate move act on the Friday print. They do not get to wait for October.

Nine days

Nine days before the July release, the Federal Open Market Committee held its target rate at 3-1/2 to 3-3/4 percent, on a 9-3 vote. All three dissents came from the hawkish side: Beth Hammack, Neel Kashkari and Lorie Logan each preferred a quarter-point hike. The FOMC statement’s own description of the labor market read: “Job gains have kept pace with the workforce, and the unemployment rate has changed little.”

That was an accurate sentence about the data the Committee could see on July 29. It could not see nine days ahead, when 103,000 jobs disappeared from the two months behind it. The next meeting opens September 15, reading a data set that, by BLS’s own history, is still not finished revising itself.

What comes next

BLS will publish its preliminary annual benchmark on August 28, tied to state unemployment-insurance tax records. Those records cover about 97 percent of total employment. The monthly survey covers about 26 percent. The last three annual benchmarks have each finished negative, and each larger than the one before: -187,000 in 2023, -598,000 in 2024, -861,000 in 2025. The ten-year average revision is 0.2 percent of total payrolls in absolute terms. That pattern makes a fourth negative reading plausible, at moderate confidence — three years is a base rate, not a projection. The finding this piece rests on carries no such hedge. It is arithmetic on numbers BLS has already published.

June’s own file does not close until September 4, when its second and final revision arrives alongside August’s first print. Until then it is a number with one revision still owed it, sitting inside an error bar that never grew and that the average month has now fallen inside.