In the same chart you see a few other not-massively-popular languages that have very dedicated user-bases of people who are really into the language (Rust, Elixir, Clojure, Julia) and they understandably score very high on this metric. Why doesn't Haskell?
Compared to the JVM languages I would choose Haskell any time (if given the choice), but for most native tasks the "burden" of Haskell's type system and mediocre performance makes it hard to pass the "value return" threshold compared to Go, Nim, Rust or even C++.
I'm sure that Haskell's performance is plenty for most applications, and the low level/unpure escape hatches are there, but they also raise the complexity significantly from that I have seen.
To me Haskell still has an aura of language research.
This may be a prejudice assumption, I have not done anything more than toys with it.
Whether it means you should lower your expectations for Go or raise them for Haskell, is up to you but Haskell can be a quite performant language for many kinds of problems - even if laziness may make performance reasoning harder.
It's a pure language, too pure, there are no variable, no state etc.;
It is difficult to motivate people (outside academia) to learn it and use it for something productive.
Moreover, there are other FP languages that are more attractive and fairly easy to learn: Elixir, Closure, Elm. They are not so "pure"...
The main reason I suggest people learn Haskell:
"A language that doesn't affect the way you think about programming, is not worth knowing." – Alan Perlis
Just kidding, I have bills to pay, too. But my standard for whether it's worth learning another langage is quite a bit higher now.
I definitely felt that learning Haskell improved my non-Haskell programming by making me more conscious (and cautious) of the impacts of using state. And more aware in retrospect when that has been the cause of wasted days debugging.
I feel learning them through Haskell in a way where they have a name and a concrete concept, have proper types and thus compile time errors is so much more easier.
The lightbulb moment can be arbitrarily bright (and it is) but if the path to it exceeds readable maximums around pre-return investment you’re going to get modest adoption.