I don't believe for a second they keep the data at a company of Google's size. But they don't need to anymore if they just feed your data through some neural network training that outputs a black box filled with weights. They can legally show your data is gone but now their model knows that people with similar browsing history, location, etc have a weak signal for tea cups. You'll get the occasional ad for them if their other networks don't filter out all the history for appearing suspicious. Google can bring in a bunch of experts saying how much data they've processed and no one really knows how the ML algorithm learns, etc.
They've improved their model none the less even if you delete your data after the fact.
And that is fine by me. Using my behavioral data to train an underfitting model is very different from actually storing my behavioral data. Sure, the word 'underfitting' does the heavy lifting in my previous sentence. But I don't think that overfitting is even feasible at this scale. Google does not train models that just memorize the habits of 3 billion people. Such a model would be useless.
They've improved their model none the less even if you delete your data after the fact.