In the past year, you’ve probably heard more about revisions to Bureau of Labor Statistics (BLS) jobs data than ever before. Usually a bureaucratic afterthought, they made headlines last August when President Trump publicly scrutinized the agency for its monthly revisions before firing former commissioner Erika McEntarfer. Then earlier this month, the agency lowered its March 2025 estimate during its annual benchmarking process by nearly 900,000, the largest revision since the Great Recession.

Revisions to the government record can be a prickly concept, especially when people’s trust in their sources of government information is low. Without transparency behind changes in information, and the time it takes to understand them, it can be hard to wrap your head around. Why is data changing? Was it “wrong,” and, if so, can we trust that it’s “right” now?

Changing data can be a bit of a pain for us USAFacts analysts, too: When numbers don’t match what was previously reported, we need to take care to understand those changes and reflect them in our work, too. This isn’t always easy – sometimes data sources overwrite the previous data and there’s no accessible record of past releases. (When all else fails, we may even rely on the old internet utility the Wayback Machine for a bit of time travel.)

But understanding the jobs revisions requires no such archival digging. The BLS’s Current Employment Statistics (CES) survey, the monthly source for nonfarm employment data, is my gold standard of data revision transparency.

So let’s humor the metaphor and mine some of that gold — we’ll cover both the monthly revisions and the annual benchmarking — and I promise a few fun charts along the way.

In each monthly release, the CES — which surveys a sample of approximately 119,000 businesses and government agencies — features new data from the previous month (e.g., the January jobs numbers first show up in February).

But not all 119,000 participants get their payroll numbers in by the time the first release comes out. BLS accounts for this by revising the number the following month as more data comes in, and then the month after that, to try to get more precise results.

Historically, this has meant an average change of about 57,000 jobs after seasonal adjustment, in either direction, between the first reported month and the third. They average a change of 41,000 jobs the first time they are revised and 31,000 the second time (note: these don’t add up because working with absolute values can be funky that way).

Let’s look at the month-over-month changes as reported with each release in 2025. In August, you can see the drop that led to the commissioner’s firing, and if you stick around until the end, you’ll get a sneak peak on what happens during annual benchmarking (much more on that later):

These revisions account for an imperfect reporting system. Take a look at this scribbly chart, which tracks the percentage of responses from employers who have agreed to participate in the CES. You can see the first month’s release (the dark green line) includes quite a bit less input — an average of around 65% of respondents in 2025 — while the second and third releases tend to average close to 90%.

Notably, too, those first-month collection rates have been declining since 2015, after a few decades of improvement. The BLS has indicated that this is a government-wide issue getting responses to survey data (are you screening your calls?). They also note the way the COVID-19 pandemic in particular changed the work world.

Ok, so we’re getting fewer responses in the first month — does a lot change by the time we get our third report? Here’s a look at the magnitude of the overall change, in either direction, between the first and “final” data releases (again, seasonally adjusted). Basically, how much is the data changing?

Hmm. This chart is, if nothing else, a reminder of just how dramatic an impact COVID had on, well, everything, with economic data collection being certainly no exception. But, with a large outlier and so many months of data in there, it’s hard to tell how revisions have changed over time — let’s instead look at a 12-month rolling average.

COVID again sticks out, but aside from that, it looks like revisions got generally smaller between 1979 and the mid-2010s, with some spikes during recessions in 1990-91 and 2008-09. The Congressional Research Service attributes at least some improvement to modernizing the survey and collecting responses via the phone and internet instead of through the mail, though surveys continue to face challenges.

But then all the way to the right, you can see that rolling average creeping back up. The 10 months we have this data for in 2025 averaged an overall adjustment of 57,000. That’s right at the historical average, but the highest average in a calendar year since COVID, and since 2008 before that.

Is this the fault of declining collection rates? Hard to say; there’s a confluence of factors that contribute to these types of large-scale national surveys. But the revision process is an exercise in balancing the expediency of a data product reporting just one month out and the precision of one getting 90% survey collection rates, and I, for one, appreciate that effort of survey design.

Speaking of surveys, they are a great tool for making large-scale estimates, but the result is just that — an estimate. And the decades of refinement that have gone into a survey like the CES typically make that estimate a pretty good one.

But what’s better than a survey? A census! (That felt less exciting than I’d hoped.)

The BLS also runs a program called the Quarterly Census of Employment and Wages (QCEW), to which all employers covered by unemployment insurance are required to report on a — you guessed it — quarterly basis. The QCEW covers 97% of all nonfarm employment, which the BLS can combine with other sources to get a full count (not an estimate) of total nonfarm employment.

Since this count can be trusted as more reliable than the survey estimate, BLS goes ahead and replaces the previous data on an annual basis in what it calls the “benchmark revision.” It applies this new benchmark revision to the data from March of the previous year and then smooths out the surrounding data with a series of adjustments, including a recalibrated seasonal adjustment for the last five years.

This was the origin of this month’s news. The annual benchmark revisions were released February 11 — late a few days due to the partial government shutdown — and adjusted the March 2025 estimate down by 898,000 jobs, or 0.6% of the 158 million jobs in the economy.

For all the praise I’ve offered the BLS for access to its data, this was a trickier find: That was the largest benchmark revision since 2009, when the March data was edited down by 930,000.

This data will continue to be refined and revised — that March 2009 figure has been published with over 15 different values. But revision is very much built into the process — as the BLS learns more about the economy, it applies those learnings to what is arguably one of the most important economic datasets in the country.

In data analysis, scrutiny is not a four-letter word. Bringing a critical lens to government data is a healthy and necessary part of understanding that data, and in turn what’s going on in the world around us. Meanwhile, government agencies can do us a great service by bringing this level of transparency to their data collection. Transparency is a win-win for the people — and essential for when government data is called into question.