A z test relies on the normal approximation, no? I don't think that is appropriate with proportions so close to 0.

I think people are probably more concerned that it doesn't have economic significance. It might have statistical significance, but the effect is still very small, so does it really matter? That's a common trap people forget to consider. You see this a lot in finance research where some variable is statistically significant in a model, but the difference in the economic outcome is so small that it doesn't matter.

In this case, these are indeed tiny percentages. But you're going from something that a lot of people dislike and that is more complicated (from a technical standpoint), so we can simplify things and user interaction with live chat is not impacted.