We do. First off we have a public parquet-format index of all of the urls we crawl every month. And then that also lives in a HDFS table that determines when we want to recrawl a page we've crawled before.
Sorry, I wasn't clear. In order to know you have an up to date copy of the page you need two pieces of information:
1. When the page itself was last updated
2. When the crawled copy was last updated
In order to get an accurate date for 1, you have to crawl it, and if you are crawling it you might as well use that copy you just crawled.
The missing here is that for pretraining AI models should accept a cut off date and not worry about being perfectly up to date. Keeping things up to date is more useful developing internet search engines for grounding.