Why 95% percentiles?
A naive approach to measuring the response time over many requests is computing the average. Computing the running average of a value can be done efficiently with many millions of requests. But from a statistical perspective, using the average to reason about response times can mislead you.
Instead of measuring averages, the 95% percentile gives a much better picture. This is because response times are not distributed normally around the average, instead they are usually log-normal distributed. In practice this means the average is extremely optimistic. Outliers in response times are usually much slower requests and you will not spot them in an average. The 95% percentile however denotes the maximum response time that 95% of all users experience.