The requirement for normality is greatly misunderstood. First, distributional assumptions are made about the conditional distribution, not the marginal (in your words, "the data itself"). For OLS, this is equivalent to the residuals being roughly normal. However, OLS actually makes no assumption about the likelihood (see the Gauss Markov theorem) and the estimates therefrom remain consistent and unbiased when the assumptions of the Gauss Markov theorem are satisfied (assuming the conditional mean is correctly specified). A good resource on this would be Introductory Econometrics by Wooldridge. It's an accessible book written for undergraduate level students with a minimal background in stats.
In all honesty, the normality assumption is perhaps the least important and one should pay more attention to endogeneity and potential omitted variables in my (humble) opinion. Whereas other violations can be rectified (non-linearity of the mean can be addressed with splines, heterogeneity of variance with robust covariance estimates) you can't fix something you didn't measure.
In the context of AB tests, you have to be a little more careful. Often times, the marginal distribution of, say revenue, may not have finite variance and so OLS shouldn't be applied. Even in the case where the variance is finite, the distribution may be so long tailed that the sampling distribution of the coefficients may not resemble a normal distribution with any sample collected in a reasonable amount of time.
