Part of the rise came from what users asked for. In June and July, more requests were for full apps, which draw more complaints than single web pages.
But with the task, language, point in the chat, how long each user had been using the chat, and request length all held fixed, the same models still drew about 30% more complaints, from 4.2% of replies to 5.5%.
To measure this, we regressed against an indicator variable for each model and for each category of those factors, which estimates how much each one raises or lowers the complaint rate. The 30% is the rise left after taking those effects out.