I spent a while coding 542 digital mental health organizations that left the market between 2000 and 2026 — shutdowns, bankruptcies, acquisitions, pivots. Up to 18 fields each plus four classification axes. The report built on it has been out for a week; the data itself had never been published, which always bothered me, so here it is under CC BY 4.0.

The strongest signal is the payer: 53% of consumer-paid companies died against 21% where an institution paid, on near-equal groups of 258 and 254. The least intuitive one: a medical co-founder does nothing on its own — 47% exits with one, 47% without.

One thing I want to flag rather than have someone find it. The four classification axes were not coded by hand: LLM agents read each company description and assigned labels against a fixed taxonomy. That is the weakest part of the method, so the written reasoning behind every single label ships as a separate file. The factual fields — funding, dates, country, outcome — come from Crunchbase, CB Insights, Tracxn, deadpool databases and trade press, not from that pass.

What it does not show: this is a graveyard, not a random sample, so the shares compare groups against each other and do not predict failure. Funding is disclosed for 59% of the companies, 67% of the sample is US and UK, and anything under n=25 is a direction rather than a number.

If you find labels you disagree with, that is genuinely the most useful thing you can do with this.

Shouldn't this be a private dataset?

Does this include the several stupid "thin wrapper on ChatGPT to 'help' people with depression/ADHD/etc." startups that I recall from about two seconds after ChatGPT came out?

Basically: "I want to be a founder, I have no expertise and a superficial understanding of the domain but I've got some tech, so here's a non-solution I'm selling as 'the best thing ever for you.'"

Starting at page 20, the report looks like someone let AI create a HTML page and then just converted it to PDF. There are misplaced "Calendar", "Money Bag" and "Globe" emojis floating above the title of each startup analysis.

I suspect that nobody, including the authors, ever read this 391 page report.