Registered-voter polls estimate views among people identified as registered to vote. Likely-voter polls estimate views among a group the pollster expects to participate.
The two groups overlap, but they are not identical. Being registered does not guarantee voting, and saying you intend to vote does not establish that you eventually will.
A model of participation
Pollsters can use questions about voting intention, interest, and past participation, or other information, to classify or weight respondents.
Different models can produce different estimates even when applied to the same interviews. A stricter screen may remove respondents who support the candidates in different proportions.
AAPOR's election polling resources explain why predicting participation is a challenge for election surveys. Source: AAPOR.
An illustrative comparison
Imagine 1,000 registered respondents with Party A ahead by two points. A likely-voter screen identifies 700 respondents, among whom Party B leads by one point.
This does not mean 300 ballots were discarded. No ballots have been counted in this example. The pollster has changed which respondents define the estimated electorate.
The crucial question is whether the participation assumptions are well documented and appropriate, not which topline better suits a reader's expectations.
Likely voters are not automatically “the truth”
A carefully constructed likely-voter model can improve relevance to an election. It can also misjudge an unusual electorate.
Unexpected participation among less frequent voters or reduced participation among previously reliable voters can challenge a model. Historical voting helps, but it cannot establish future behavior perfectly.
What to check in a release
Look for how the likely-voter group was selected, whether the screen changes over time, and whether both populations are reported.
When comparing polls, keep the population label visible. A shift from registered to likely voters can look like political movement even if the underlying responses have barely changed.
For trend analysis, prefer comparable measures. If methods differ, describe that difference before declaring that a candidate has gained or lost support.
Related reading: comparing polls over time and poll weighting.