Viral Polls Are Not Representative Polls

The central argument is that the polls people see most often online are not necessarily the polls that best represent the wider polling picture. To test this, Kevin Reuning collected 90 days of posts from @usapolling.bsky.social, a highly active polling-distribution account on Bluesky, then used an LLM to extract poll details. The analysis focuses on posts containing a single general-election poll and compares each poll’s Democratic-Republican margin with the average margin for that race. Bluesky’s left-leaning user base and the choice of one account make the sample limited, but the pattern echoes an earlier analysis of Twitter.

Poll posts with more Democratic-leaning results than the average for their race tended to receive more reposts. This relationship appeared across nearly all races with enough polls to examine, with North Carolina’s Senate race as an exception. A model accounting for differences among races likewise found that a poll’s distance from its race’s average predicted reposts better than its margin alone. Race-specific effects also varied: Texas’s Senate race drew relatively high sharing, especially when a poll showed the Democrat ahead, while Pennsylvania’s governor’s race attracted little attention. The findings suggest that viral polling can distort passive news consumers’ impressions of the electoral landscape: attention favors unusually favorable results, rather than simply reflecting the overall state of a race.