I thought I knew exactly how the ab test start callback behave but now that we have a different experiment here I am not very sure anymore. We used one event that is the one that happens the least compared to others as metric and the others as guardrails. So these numbers for control and variation 1 are from the number of people that reached that metric with the respective variation, right? So even though we have more people in the control or the variation group they only count as users once they trigger this metric?

The user number should be the total unique users who saw the experiment within the time window.

So it has nothing to do with the metric, right?

usually - you can add 'activation metrics' which will reduce that number to include exposure + a metric event. And if you have multiple Experiment Exposure Tables, that can also adjust the set of users

It could also just be due to low traffic, I was just wondering. So if my first assumption is correct, the callback is called as soon as we do a "eval" for the user to see in which group he is, correct?

yes

The queries we run require at least one conversion for a variation in order to get the number of users. We're looking to change that soon so that you can always see accurate user counts.