People who are hiring for roles using smaller languages
know that there are few people local to them who are practitioners in the language. They tend to do one of two things about that:
1. They don't even bother advertising a position in local geographical "general" job boards. (Job postings on LinkedIn cost money for every day you have them up; they're a waste of resources if you can predict with high confidence that nobody will find the role through there.) Instead, they'll advertise globally but targeted to the language's community (i.e. language-specific job boards, forums and chat groups, newsletters, etc.) This is where the people using the language are looking, too, anyway, because they also know that there are too few local opportunities for it to make sense to invest the time in checking local job-boards for a job matching their skillset.
2. They don't bother hiring for the language. Instead, they hire for "experience with [relevant language paradigms]" and "experience with any of [similar, more-popular languages]" and then expect the new hire to learn the language on the job.
My personal job-criterion for hiring Elixir devs is "a polyglot in several different language paradigms, fluent in at least one functional language." I find that that filter actually predicts better whether they'll be a good Elixir dev, than actual experience using Elixir does.
-----
As an aside, there's also the fact that languages like Erlang/Elixir (or the MLs, or the Lisps, or Prolog, or...) tend not to be languages used for everything in a company, but instead tend to be languages used for the secret sauce core component of a company. A lot of the time, companies don't talk about using these languages, even though they do, because they consider them a competitive advantage over their rivals in whatever niche they occupy.
Heroku is an obvious example: much of their architecture was written in Erlang [nearly everything at first], but they never advertised that fact once in any official capacity. Likewise, HFT firms never mention they're using ML or Prolog, but many do, because trading bots are often just souped-up expert systems. The only time you find these things out, is when having a beer with ex-engineers from those companies.