US ELECTIONS SPOTLIGHT 2026

How to read this index

Each brief defines its candidates with an ad-hoc boolean query using primary: sentiment targeting, so sentiment is scored toward the named figure rather than toward the post as a whole. Volumes count posts, articles and comments across seven channels; Twitter/X figures are projected from a 10% sample, so they exceed the number of individually stored documents.

These numbers are not polls. They measure what was said about each candidate, how much of it there was, and who said it. In seven of the ten races here with published polling, the candidate who takes the better of the conversation is the one behind in the polling average. That is the finding worth sitting with, and it is a fact about the conversation rather than about the electorate: a candidate can lead the vote and still take the worse of the argument online. Each brief carries the current polling for its race, so the two readings can be compared on the same page.

Where sentiment runs heavily against a front-runner, the evidence here can show concentrated attack and heavy amplification; it cannot show who paid for either. Across these races between 1 and 12 percent of the conversation is original writing, so much of what separates two figures is how far a post travels rather than how many people wrote one. Net sentiment also excludes neutral posts and is computed only on posts carrying a sentiment label, a minority of the total in every race. A figure of −78 does not mean 78 percent of a state dislikes a candidate; it describes the balance of positive and negative labeled posts in that corpus. Cell sizes differ by orders of magnitude between states, and every brief states where its own base is too thin to carry a precise figure.