How many new businesses actually survive?
Half of new businesses are gone in five years. Two thirds are gone in ten. Those figures come from a government table that follows every business that opened in a given year and counts how many are still reporting employment each March afterwards.
Reading the table directly turns up a better problem than the number. Two pages on the same government website disagree, and the thing being counted is not what almost anyone thinks it is.
The numbers
Evidence Across 27 completed cohorts, five-year survival runs between 45.4 and 51.9 percent, averaging 49.3. Across 22 cohorts that have reached ten years, survival runs between 32.4 and 35.3 percent, averaging 33.9.
Take one cohort to make it concrete. Of the 677,876 establishments that opened in the year ending March 2015, 79.6 percent were alive a year later, 50.2 percent at five years, and 34.7 percent at ten. That last figure is 235,071 places still open a decade on.
The word that does the damage
Evidence The table counts establishments. An establishment is one physical location, not a firm and not a company. A business with three sites is three establishments. A closure in this table can be a company failing, or a location shutting while the company thrives, or an ownership change, or a reclassification.
Every article that turns this table into a sentence about startups has swapped the noun. The table cannot tell you how many businesses failed, only how many places stopped reporting employment.
The thing I did not expect
Evidence The year a business opens turns out to matter much less than it should. Across 22 cohorts, ten-year survival spans 2.9 percentage points. The lowest is 32.4 percent for businesses opening into the year ending March 2001. The highest is 35.3 percent for those opening in the year ending March 2010.
One of those cohorts opened into the dot-com collapse and the other into the recovery from the financial crisis. Ten years later they are three points apart.
Year-one survival moves a little more, between 75.2 and 80.9 percent. Its worst year is the cohort born into the year ending March 2008. Its best is the cohort born into the year ending March 2020.
The pandemic cohort had the best first year in the series. There is no confident explanation here and none is going to be invented. Emergency support programmes are an obvious candidate and so is a change in who opened a business that year.
This result cuts against the argument the rest of this site makes. Thirty-two years, two recessions and a pandemic, and the ten-year survival rate of American establishments will not move more than three points. Whatever closes two thirds of them is not the economy.
The same agency, two answers
Evidence There is a second page on that website, an anniversary feature, giving five-year survival by birth cohort as 54.3, 52.3, 55.3, 49.8, 56.0 and 57.3 percent for six named years. The table gives 49.6, 48.1, 49.9, 45.4, 51.1 and 51.9 for cohorts carrying those same labels.
About five points apart, in the same direction, throughout. Both read on the same day. I do not know which is right. The likely explanation is that the two pages label cohorts differently, calendar year against year ending March, but that is a guess and I have not confirmed it.
What I can say is that every article quoting government survival rates picked one of those two pages, and almost none of them knew the other existed.
One practical note
Observation The table blocks automated downloads. Three scripted attempts to read it returned three different wrong answers for the same cohort, quietly, with no error. It took opening it in an ordinary browser to get it right.
An automated read that finishes without complaining is not evidence that it parsed the columns correctly. That is worth knowing in a year when a great many numbers are being fetched by machines.
The honest summary
Roughly half at five years and a third at ten, for establishments, from a table that another page on the same site contradicts by five points. That is a less satisfying sentence than nine out of ten fail, which is probably why the other one won.
Where this stops. United States only, private sector, establishments rather than firms. Nothing here says whether a closure was a failure, a sale or a choice, because the register cannot see the difference.
What I am not saying. That the cohort finding means conditions do not matter. It means the national economy does not move ten-year survival much. The conditions an individual business faces, its industry, its money, its location and its owner, are a different question and that one is still open.