They ask seven questions, known to correlate generally with turnout, about voting intentions, past voting, interest in the campaign and knowledge of voting procedures. They then combine those questions into a index that they use to designate some predetermined fraction of their adult sample as most likely to vote.
Yes, this moves deeply into the methodological weeds for most people, but why should you care? Because the switch from registered to likely voters usually comes with a Republican advantage of between 1 and 3 percentage points. GOPers simply are more likely to vote and more likely to say they're gonna vote (or did vote, or meet the other criteria listed above). Thus, the Obama-Romney results shift.
Every polling shop has it's own special sauce when it comes to deciding who gets to be called a likely voter. People lie about whether they voted int he past and their plans to vote in the future, so other questions are designed to tease out the most likely of likely voters, thus making polls more predictive when Election Day rolls around.
If you read the HuffPo column -- and I recommend you do so if you're into polling or consume polls or worry about polls -- read #5 on the list about how "internals" are being conducted. It's interesting stuff, the kind of thing rarely shared with news folks except tangentially, when the news is good. A few well-connected journalists do get to hear or see some of these "internals." Pay attention to what they say versus the typical political pundits.