No, not in the sense most posts mean. Official US data show that about a third of new business locations are still operating ten years after they open. So roughly two in three closed, not nine in ten. And the best-known 90% figure measures something else: start-ups that fell short of their own projections, not start-ups that shut down.
The real question is what “fail” means, which companies are counted, and over how many years. Change any one of the three and the answer moves from about 20% to about 95%. This page shows where the numbers come from, then gives you a ten-minute method for the next viral statistic.
Where does “90% of startups fail” come from?
We could not find a dataset behind the round number. The oldest dated document we could open that states it is a 2011 Startup Genome report on “premature scaling” (growing faster than the business can support), edited in March 2012. Its introduction says: “More than 90% of startups fail, due primarily to self-destruction rather than competition.”
That sentence has no source attached, no time period and no definition of “fail”. The report’s own data came from more than 3,200 high-growth technology startups, and it studied why they scale too early, not how many close.
Since then the line has been repeated in thousands of posts. Repetition is not evidence. The Evidence-Aware Life has a phrase for it: one source in twenty coats.
What does the official data show?
The US Bureau of Labor Statistics (BLS) follows private-sector business locations that report employees, through its Business Employment Dynamics program. Its Table 7 tracks each year’s new openings over time; “still operating” here means the location still reports employees. From the table as published, with data through March 2025:
| Opened in the year to | Still operating in March 2025 | Years later |
|---|---|---|
| March 2024 | 77.9% | 1 |
| March 2020 | 51.4% | 5 |
| March 2015 | 34.7% | 10 |
| March 1994 | 12.6% | 31 |
So roughly one in five new locations closes within a year, about half within five years, and about two in three within ten. That is a real, checkable pattern. It is not 90%.
The Netherlands tells a similar story. Statistics Netherlands (CBS) reported in November 2022 that of the roughly 153,000 businesses founded in 2015, almost 96% still existed a year later and 64.7% in 2020. Among new employers (businesses hiring staff for the first time), 48.6% were still employers five years later.
Why does the definition change the answer?
Three things carry the whole statistic: which companies, what counts as failing, and by when. The table below puts six versions side by side. For the BLS and CBS rows, the failure share is 100 minus the published survival rate.
| If “fail” means | Who is counted | Share that “failed” | Source |
|---|---|---|---|
| Location stopped operating within 1 year | US private-sector locations opened in the year to March 2024 | 22.1% | BLS Table 7 |
| Same, within 10 years | Opened in the year to March 2015 | 65.3% | BLS Table 7 |
| Business no longer exists after 5 years | Dutch businesses founded in 2015 | 35.3% | CBS |
| Assets liquidated, investors lose most or all their money | Start-ups, in Ghosh’s estimate | 30-40% | HBS Working Knowledge, 2011 |
| Did not deliver the projected return | Same | 70-80% | Same |
| Declared a projection, then fell short of it | Same | 90-95% | Same |
The last three rows come from Shikhar Ghosh, a senior lecturer at Harvard Business School, as reported by the school’s Working Knowledge site in 2011. The article does not describe his underlying sample, so treat them as an expert’s estimates, not a published table.
Look at what the 90-95% row actually measures. Ghosh put it plainly: “Very few companies achieve their initial projections. Failure is the norm.” Missing your own forecast is common. It is not the same event as closing down.
Two more traps sit inside the definitions:
- A business is not a startup. The BLS and CBS figures cover every kind of new business: restaurants, plumbers, consultancies, shops. They are not startup-only figures, and a venture-funded tech company is a small, different group.
- Closed is not the same as failed. The BLS counts locations, not companies. Its own FAQ gives the example: when a chain opens a new branch, that counts as a new opening. A chain that shuts one branch adds a closure even if the company is doing well. The table records that a location stopped reporting employees; it does not record why.
How LEVEL grades the claim
LEVEL is a five-step check for deciding how much weight a claim can bear before you act on it. It comes from The Evidence-Aware Life, which is available now. You do not need the book to use it. Here it is, applied to “90% of startups fail”:
- Locate: say the claim in one sentence that could be checked. “Nine out of ten new businesses close within X years.” If you cannot fill in the X and the kind of business, you are judging the packaging, not a claim.
- Estimate: decide what being wrong would cost. Quoting it in a pitch deck to an investor who knows the BLS table costs credibility. Using it to talk yourself out of, or into, a big decision can cost far more. Higher stakes need better evidence.
- Value: place the evidence on the book’s Evidence Ladder, seven kinds of support: assertion, anecdote, authority, mechanism, study, convergence and your own test. The book treats these as kinds, not a ranking, so you still check quality within each. The Startup Genome sentence is an assertion. Ghosh’s estimates are authority: an expert’s judgment without a published table. The BLS and CBS tables are systematic measurement.
- Expose: state the strongest case against the claim before you use it. Against “90% close”: the BLS table shows about 65% after ten years. For the claim: if “fail” means missing your own projections, Ghosh’s 90-95% supports it.
- Lean: act only as firmly as the evidence and the stakes allow. Use the version you can defend: “About a third of new US business locations are still operating ten years after opening (BLS).”
The LEVEL framework page has the full method.
The ten-minute method
For the next viral statistic, set a timer.
| Minutes | Do this |
|---|---|
| 0-2 | Rewrite the claim with three blanks filled in: who is counted, what event, over what time. |
| 2-5 | Trace it. Search the exact number plus “report” or “data”. Open the oldest document you can find. Does it show a table, or only state the number? |
| 5-8 | Find the primary source for the same definition: your national statistics office, a regulator, or the original paper. |
| 8-9 | Compare definitions. Same population, same event, same time window? |
| 9-10 | Keep, rewrite or drop. Write the version you can defend, with its source and date. |
If step 2 ends in a blog that cites a blog, stop: you have found an assertion. Record the result in an evidence ledger (one row per claim, with its source and what would retire it), so you never check the same number twice. The same habit of starting from real frequencies runs through Calibration techniques for founders.
Try this today (10 minutes)
Open the last deck, post or report you wrote that contains a statistic. Run the ten-minute method on the one number your argument depends on. If you cannot fill in who, what and when, rewrite the sentence or delete it.
Cite this:Do 90% of startups fail? How to fact-check a viral statistic in ten minutes.Len P. van der Hof. https://lenvanderhof.com/en/blog/do-90-percent-of-startups-fail/ ·