"Across the country, pollsters seemed to systematically undercount GOP support [in 2020], despite the fact that they were trying very hard, after some issues in 2016, not to do that."

"But with the political world entirely focused on Virginia, looking to glean lessons about the state of the national parties, the fate of Biden’s agenda, and Americans’ hope for the future, the pollsters — the people who give us a sense of what to expect in elections — are offering a new round of projections without ever having quite figured out what went wrong last year...  In talking to a range of nonpartisan and party-affiliated pollsters in recent weeks, I found that many dismissed, but laughed nervously about, the least scientifically sound idea of all, which unfortunately would have looked on the surface like a fix in 2020: just artificially slapping four extra points of support on Trump’s side. When Patrick Murray, who runs the Monmouth University poll, explored his data earlier this year, he told me, he explored whether there was 'any way I could have modeled my electorate on anything other than arbitrary guesses that would have gotten me closer to the endpoint. And the answer was basically no. If I had the ability to jump ahead in time and know exactly which voters that spoke to me were going to show up to vote, we still would have been off with the Trump vote share.'...  The likely problem, in short, is that [pollsters] simply aren’t reaching a significant number of voters activated by Trump.... And if you aren’t in any kind of contact with a voter, it’s essentially impossible to figure out what’s motivating him or her, or what distinguishes them from the people who are answering you...."

From "Polling in America Is Still Broken. So Who Is Really Winning in Virginia?" (NY Magazine).