I agree with this take.

Our current user count calculation is very permissive: we count all public keys that have user metadata set. This can be gamed (anyone can create bot key pairs and set metadata for them), so our nostr user numbers are overstated.

The alternative is to use a web of trust algorithm to identify real users based on follow connections. The downside of this technique is that we would be underestimating the number of real users because most new users won’t have any real followers initially.

Given that other networks generously include bots when stating their user numbers, our approach is probably a decent apples-to-apples comparison. In fact, as you point out, nostr seems to have higher engagement per number of followers.

Having said all that, I am not happy with the fact that our numbers don’t reflect real users. We will work on improving these techniques, but I definitely don’t want to err on the side of excluding real users from the count.