The slogan

Few ideas are repeated as often in startup culture as the virtue of failure. “Fail fast, fail often.” “Failure is just learning.” Conferences have entire tracks of founders telling the story of the company that died before the one that worked. The implication is that failing is an investment: painful now, paid back later in wisdom.

Part of that is right. Fear of failure keeps people from trying, and most attempts will fail. But “failure is part of the process” quietly slides into a stronger claim: failure makes you better. That claim can be checked, at least partly, and my reading of the evidence is that a past failure, by itself, predicts little about your next attempt, though failure can teach a lot when you learn from it on purpose.

How common failure is

First, the base rate (how often failure happens overall), because “90% of startups fail” depends entirely on what you call failure.

  • All new businesses (US): of private-sector employer establishments (individual business locations with employees) opened in the year to March 2013, 79.6% were still operating a year later, 50.6% after five years and 34.7% after ten1. They span the whole economy, from construction firms and shops to software companies.
  • Venture-backed startups: research by Shikhar Ghosh at Harvard Business School, covering more than 2,000 companies that received venture funding between 2004 and 2010, generally at least $1 million, estimated that about three in four did not return their investors’ capital in full. A smaller share, 30–40%, were liquidated with investors losing everything2.

So whether failure is “the normal outcome” depends on the definition: closing within ten years, yes; returning less than investors put in, yes; total loss, a large minority. Either way, failure is common enough that it matters whether it teaches anything.

Does a failed startup improve your odds?

The most-cited direct evidence comes from Gompers, Kovner, Lerner and Scharfstein, who studied founders who joined US venture-backed companies between 1986 and 2000, with outcomes tracked to 20073. They counted a company as successful if it went public or filed to go public, and estimated each group’s chance of success with other factors (industry, year, location, stage) held equal. “Failed” here means the earlier venture-backed company did not reach that outcome, which can include acquisitions; “first-time” means no earlier venture-backed company in the data.

Axes: each bar is the study’s estimated chance that a founder’s venture-backed company goes public or files to, grouped by what happened to the founder’s previous company. Look for: how close the red bar (failed before) is to the grey one (first-timers). Takeaway: in this analysis, the one-point difference for having failed before was too small to tell apart from chance variation in the sample, which is not the same as proving it is zero; having succeeded before was associated with about nine points more.

The paper also ran the analysis at the level of companies rather than individual founders. There, previously failed founders did show a modest, statistically significant advantage over first-time teams (21.3% vs 17.1%), still far below previously successful ones (30.7%). So the honest summary is: some experience may help a little; a past failure is nowhere near a past success.

Two cautions. These are estimates for one population, venture-backed US companies of the 1990s, and one outcome, going public or filing to. They don’t measure what founders learned; founders who fail and try again may differ from first-timers in ways the study can’t see. The authors also found that previously successful founders were better at picking the right industry and the right moment to start, which looks like a skill rather than luck.

Who learns, and who doesn’t

If a past failure predicts little on average, maybe it still teaches some people a lot. A 2019 study in Nature looked at repeated attempts in three very different domains: scientists reapplying for grants, executives and board members involved in one venture-backed startup after another, and (grimly) organisations repeating attacks4.

Among people with long runs of failures, those who eventually succeeded and those who didn’t (and then stayed inactive for at least five years) looked similar at the start, but on average their sequences of failures behaved differently. In the group that eventually succeeded, attempts came faster and got better. In the other group, they didn’t speed up or improve. The difference was visible within the first few attempts.

The authors explain this with a model in which successful people keep and refine the useful parts of earlier attempts instead of starting over. They could check that directly only for grant applications. So the safe reading is modest: improvement across attempts is not automatic, and the people who improve show it early.

Related evidence points the same way. US Census data on founders shows that the founders of the fastest-growing 0.1% of new companies were on average 45 years old when they started them, not 25, and that prior experience in the same industry strongly predicts success5. What helps seems to be specific, transferable knowledge, about customers, an industry and how a business runs, rather than the general experience of having been through something hard.

What “fail fast” should mean

None of this makes “fail fast” bad advice. It makes it often misread. In the lean-startup tradition it never meant “start companies quickly and let them die”. It means run cheap experiments that test your riskiest assumptions before they cost you a company6.

Demand is the obvious candidate. In CB Insights’ analysis of 111 startup post-mortems, the most common reasons were “ran out of cash or failed to raise new capital” (38%) and “no market need” (35%); most companies gave several reasons7. “No market need” covers both products nobody wanted enough to pay for and markets that disappeared after launch. The first kind is the one an early test can catch.

Here is what such a test looks like in practice:

  • Assumption: small accounting firms have a painful invoice-matching problem and will pay €30 a month to solve it.
  • First test (interest): a one-page site describing the product and price, shown through €300 of targeted ads to accounting-firm owners, with a “join the waiting list” button. Two weeks.
  • Decision rule, set in advance (the numbers are illustrative; pick yours before you look): at least 30 sign-ups and five follow-up calls where owners describe the problem unprompted → move to a second test. Fewer → check whether the ads reached the right people and the message was clear, then change one thing and rerun.
  • Second test (payment): offer a paid pilot or a refundable deposit to the most interested sign-ups. Sign-ups show interest; only money shows willingness to pay.
  • Afterwards: write down what to keep (the audience? the price?), what to change, and what to test next.

A result like “few sign-ups” is a signal to investigate, not a verdict. But it costs weeks and a few hundred euros rather than years and your savings.

So, does failing make you a better founder?

Not automatically. In Gompers and colleagues’ data, a past failure predicted little or no advantage next time, depending on how the analysis was run, and far less than a past success. Whether failing teaches depends on what you do with it; people who improve across attempts show it early, and specific industry knowledge is associated with success in a way that survived hardship alone is not shown to be.

If you are about to start something, my advice is to spend your failures cheaply: one assumption, one test with a rule set in advance, one written note on what to keep and what to change.

  1. U.S. Bureau of Labor Statistics, Business Employment Dynamics, establishment survival table (all private sector). bls.gov; see also 34.7 percent of business establishments born in 2013 were still operating in 2023, The Economics Daily, 2024. bls.gov ↩

  2. Gage, D. (2012). The venture capital secret: 3 out of 4 start-ups fail. The Wall Street Journal, 20 September 2012, reporting research by Shikhar Ghosh (Harvard Business School). ↩

  3. Gompers, P., Kovner, A., Lerner, J. & Scharfstein, D. (2010). Performance persistence in entrepreneurship. Journal of Financial Economics, 96(1), 18–32. doi:10.1016/j.jfineco.2009.11.001. Figures are the paper’s predicted success rates at the entrepreneur level (Table 5). ↩

  4. Yin, Y., Wang, Y., Evans, J. A. & Wang, D. (2019). Quantifying the dynamics of failure across science, startups and security. Nature, 575, 190–194. arXiv:1903.07562 ↩

  5. Azoulay, P., Jones, B. F., Kim, J. D. & Miranda, J. (2020). Age and high-growth entrepreneurship. American Economic Review: Insights, 2(1), 65–82. aeaweb.org ↩

  6. Ries, E. (2011). The Lean Startup. Crown Business. ↩

  7. CB Insights (2021). The top 12 reasons startups fail. Based on 111 startup post-mortems; companies could give several reasons. ↩

  8. Armour, J. & Cumming, D. (2008). Bankruptcy law and entrepreneurship. American Law and Economics Review, 10(2), 303–350. ↩