Where does the claim that 90 percent of startups fail come from?
Nine out of ten startups fail. You have read it in a pitch deck, a newspaper and probably a condolence. I went looking for the study. There is not one.
Not a weak study, not an old study, not a study measuring something narrower than the claim. No study. Two sources carry the figure and both of them assert it.
The report that explains a number it never established
Evidence The first source is a 2011 report on startup failure. On page five it says more than ninety percent of startups fail, due primarily to self-destruction rather than competition. There is no citation on that sentence and no entry for it in the report’s own sources section.
Twelve pages later the report offers to explain “the 90% failure rate of startups,” which is a strange thing to do about a number you have not produced.
Its own caveats section says the authors did not yet have longitudinal data. Their sample is around 3,200 companies who signed themselves up to a benchmarking web app. A snapshot of people filling in a form cannot observe whether any of them later failed. Failure is never defined anywhere in the document.
The research that was never published
Evidence The second source is a Harvard Business School senior lecturer, usually cited through a 2012 Wall Street Journal article. The earliest primary appearance is actually eighteen months earlier, in a March 2011 interview, and it gives different numbers.
In 2011 he is quoted saying the failure rate is 30 to 40 percent if failure means liquidation with investors losing most of their money, 70 to 80 percent if it means missing the projected return, and 90 to 95 percent if it means declaring a projection and falling short. By 2012 the middle tier has become about three quarters and the top tier has become more than 95 percent.
Numbers drift like that when they are recalled. They do not drift when they come out of a dataset. There does not appear to be a dataset. Archived listings of his publications from 2015 and 2020 show teaching cases and technical notes. No journal article, no working paper, nothing anyone could check.
So the figure that became the internet’s fact is the loosest of three definitions in an estimate that was never written down. Under that definition a company that grows every year, pays everyone, and misses its own forecast is a failure.
What the measured literature says instead
Evidence Research using US Census records followed 12,865 venture-financed firms against more than 12 million others. Five-year cumulative failure was 51 percent for the non-venture-financed firms and 19 percent for the venture-financed ones. Their definition of failure is administrative, every location disappearing from the register, which is the most inclusive measure available. It still does not reach ninety.
Evidence A second study tracked venture-backed entrants in Census data against matched firms that looked the same at entry. By 2007, 75 percent of the venture-backed firms had shut down. So had 66 percent of the ordinary ones. The survivors of the venture-backed group employed 364 percent of what the original cohort started with. The matched controls managed 67 percent.
That pair should be the famous number and is not. Venture-backed companies close at close to the rate of comparable ordinary companies. What differs is the size of what lives.
The study that gets misread
Evidence One real finding gets quoted as a failure rate and is not one. A study of 20,961 venture-backed companies found that almost three quarters of founders receive nothing when their company exits.
That is a measure of cash reaching the founder’s own shares after everyone with a preference has been paid. A company can be acquired, employ hundreds of people, and still pay its founder zero. It describes how the money is divided, not whether the business worked.
The honest summary
Somebody wrote a sentence. It was repeated until the repetition became the evidence. The people with actual records were measuring something between a fifth and three quarters depending on what they counted, and their numbers travel less, because each one needs a definition attached and the famous one does not.
Where this stops. I could open the two sources behind the ninety percent claim, and I did. The peer-reviewed studies are paywalled, so I read the working-paper versions, which can differ from what was finally published. Every number above comes from those.
What I am not saying. That starting a business is safe, or that most of them work. Half of ordinary new businesses are gone in five years and that is a real number from a government register. The claim here is only that the famous figure is not a measurement and never was.