If you take a look at the data, I tested the database in cross-az functionality as well. Still better reports than what Google has shown in their benchmark. The only thing I can gather is they were testing from an instance on Google's platform.
Everything was stood up using Terraform, the benchmark tests were conducted using Sysbench, and all data was plotted using R.
Since we didn't have access to Google's original data, we provided some overlays in the post as an easy visual comparison. However the data used to produce all of those graphs as well as additional data is in the repo.