A normal "pattern" of optimization in our system is slightly above average to slightly below average performance for the first period (depending on how much traffic your site gets - the greater the traffic, the shorter the time) while the system gathers statistically significant data on your old site and "weeds out" some lower performing variants. Undoubtedly, in the early days, some of our experiments are going to fail and some succeed, and that will be reflected in overall results. The results from experiments using the better performing layouts can be "diluted" by lower performing layouts until they are excluded from testing. During that early period, results can be somewhat inconsistent but tend to improve gradually as time goes on. Then, once we get enough data, we often see a sharper improvement in results, followed by a period of solid / steady improvement. After that, we see gradual improvement month-over-month for eternity (ok, only for 3.5 years so far, but with the systematic improvements in the technology we expect it to continue over a long period of time).
So, overall, you should mentally prepare yourself to give Ezoic sufficient time to demonstrate its potential. We speed up the process as much we can by using statistical clustering etc. but we are at the mercy of math in that the variation of results is higher on smaller data samples. Because of this, in an analysis of past publishers, we've determined that the number one factor of a site's success or failure on our platform is time - the lower the time, the lower the likelihood of success, the greater the time, the higher the likelihood of success.
A quick summary:
Stage 1: Data gathering. Performance will vary.
Stage 2: Optimization time - results similar to stage 1 until one larger improvement "jump", followed by steady improvement).
Stage 3: Ongoing incremental gains.