GP has argued otherwise. PyPI sampling is not random. It is heavily influenced by automatic downloads and such and completely ignores that most simple 2.x libs / non-3.3-tagged libs will work just fine on python 3.
In other words, it is neither random nor representative of python 3 usage, failing the premise you yourself mentioned.
There's no easy way to determine Python 3 usage overall. Your best bet is to go around polling companies of all sizes as well as indie devs. What you will find is that:
1. Some amount of people still use Python 2.x because they don't know any better
2. Some amount of people want to use 3.x, but still use Python 2.x because of incompatible dependencies.
3. Some amount of people (companies mostly) still use Python 2.x because a migration to 3.x is currently too costly.
1 -> This is partly caused by FUD. Hating on Python 3 has become a popular past time of people who don't know how to be productive otherwise. It's also caused by the very high amount of tutorials etc around the web that "require" 2.x because, for example, they have an unparenthesized print call (relevant XKCD: 353). Arguably this is one of the big failures of the 3.x push.
2 -> That number is constantly diminishing as more dependencies become available on 3.x. The most recent big win is Scrapy, a massively popular scraping framework, which just finished its 3.x port a couple weeks ago. It's an ongoing process - we are in a much better state than we used to be, the Py3WOS is looking almost entirely green.
3 -> This number is mostly irrelevant as it tends to be in closed ecosystems. So maybe CloudShareBuzzr Inc. is using Python 2.7 internally, but they don't produce anything open source so it doesn't affect the rest of the ecosystem. As new, major features are constantly added to Python 3.x (venv in 3.3, pip in 3.4, async def in 3.5), eventually the cost of switching will be worth it.