> dating apps that don't successfully match won't be used.

This isn't entirely true. A portion of the user base uses it almost entirely for entertainment and validation. I'd suspect it's a sizeable portion at that.

> it's not really feasible for the dating site distinguish between "will successfully date for a few months" and "will successfully date for a lifetime".

I think it is feasible for a dating site to figure that out, given enough data.

And I think the 'few months' figure is probably too high - they want to match people who will go on a few dates or a one night stand, then return to the app.

It's a pretty damn strong signal when a user matches with someone, exchanges phone numbers, then deactivates their account. The ML algorithms will be trained to do everything possible to avoid that outcome.

> I think it is feasible for a dating site to figure that out, given enough data.

I've worked on the developer-side of this; I assure you, it's not. It's hard enough to predict if people will exchange messages with any reasonable precision + recall.

> It's a pretty damn strong signal when a user matches with someone, exchanges phone numbers, then deactivates their account. The ML algorithms will be trained to do everything possible to avoid that outcome.

Yeah, any dating site that does this is doomed to irrelevance.

If it were as easy to make a successful dating site by just optimizing for that signal, you should really do it. Match will acquire you for millions - billions, and the core code is relatively simple. You could get it running in under a week. Get some VC funding, do a gradual roll out onto a few colleges, you'll be a multi-millionaire by Spring.

Have you considered the network effect? Users can't switch if there is no competitor with possible matches.

There is also a network effect for data. A dating site with few users doesn't collect much data, so can't train an ML algorithm to do a good job of showing you interesting matches.