We read what 191 Democratic nominees posted in October and placed each one on a scale from the party's moderate wing to its progressive wing. In the Senate, the most progressive-sounding nominees ran behind. For governors the evidence is mixed, and in the House districts counted so far there was no clear pattern.
We looked at what candidates actually posted, day to day, in the last month of the campaign.
In Senate races, the third of nominees whose feeds sounded most progressive ran 3.8 points behind the typical Democratic shift from 2024, on average; the most moderate-sounding third ran 2.1 points ahead. And in 13 of the 17 states where we could compare a Senate race with a governor's race on the same ballot, the Democrat who sounded more progressive ran behind the other.
There are reasons for caution, though:
We also looked into populist language and its relationship with electoral performance. Reading one post at a time, the share of a nominee's posts that were progressive without populist framing correlates with running behind more clearly than the share that were populist but not progressive, though that split leans on the same four races.
Each dot below is one October post from one of the 191 Democratic House and Senate nominees, placed so that posts saying similar things sit near each other. The shade is how progressive the whole feed of the post's author sounded.
The clusters whose authors sound most moderate are about farmers and agriculture, local events and Friday-night football, how to vote by mail or drop box, stock trading and corruption in Congress, and breast cancer and domestic violence awareness. These are topics whose authors lean moderate on average, not topics that are moderate in themselves: candidates from across the party post about each of them.
The clusters whose authors sound most progressive are about Trump and abuse of power; ICE and immigration detention; Israel, Gaza and the West Bank; tributes to Dolores Huerta and labor and civil rights; and Indigenous Peoples' Day.
Hover over or tap any dot to read the post.
Comparing the most progressive-sounding third of the 191 Democrats with the most moderate third, these are the words each side used much more often. Every word on the list was used by at least eight different candidates on its side.
The progressive-sounding end talks about money, billionaires, power, government and ICE. The moderate-sounding end talks about costs, prices, families, local roads and businesses, and the word "bipartisan". In the Senate, the share of a nominee's posts using a fixed list of bipartisan words, given in the technical note, lines up with the results only weakly: a larger share correlated with running further ahead, within noise (rank correlation +0.20).
Rank correlation among the 28 scored Senate nominees. Counts as of Nov 8. AP results via Fox News.
Each dot is a Democratic Senate nominee. Further right, their October posts sounded more progressive. Higher up, they ran further ahead of the typical shift: how much better or worse they did than Kamala Harris in their state in 2024, compared with other nominees in the same position.
Split into thirds, the most moderate-sounding third ran 2.1 points ahead, the middle third 0.2 ahead, and the most progressive-sounding third 3.8 points behind. Holding constant how Democratic the state was, whether the race was rated competitive and whether the Democratic nominee already held the seat, each step more progressive on the chart's scale correlated with running 2.2 points further behind (95% interval −4.1 to −0.2); an interval that crosses zero is within noise, and this one stays below it.
Four races carry most of it: Ed Markey in Massachusetts, Peggy Flanagan in Minnesota, Abdul El-Sayed in Michigan and Angie Nixon in Florida. They are the four who ran furthest behind of the 28, and each is among the five most progressive-sounding. Take out any single race and the estimate stays between −1.8 and −2.5. Take out all four and it falls to −0.7, the same direction but within noise. Some shrinking is expected when the races that ran furthest behind are left out: in simulations where the pattern runs evenly through all 28 races, about four-fifths of it would remain, on average. Here about a third remains, and an even pattern leaves that little only about one time in ten. So the result leans heavily on these four races, which is a reason to read it with caution.
Hover over or tap any dot to read a post that two AI readers, working separately and without the results, picked as typical of that feed. Two more checks on the Senate result come near the end of this page: a re-read with names masked, and one using only September posts.
In the House the picture is different. The most moderate-sounding third ran 2.5 points ahead of the typical shift. The middle third and the most progressive-sounding third both ran slightly behind it (−0.4 and −0.7).
But the progressive-sounding third sits in safe Democratic seats far more often: 13 of its 33 nominees ran where Harris won by 15 points or more, against 5 of 32 in the moderate third. So the thirds partly compare safe seats with competitive ones.
Lay the same nominees out by how Democratic their seat was. Across all 97, sounding more progressive correlated with running 1.3 points behind per step on the scale, within noise (−2.8 to +0.2). In the 76 seats under D+15 it was −1.6 (−3.3 to +0.1), the same direction as the Senate, but within noise. Two South Texas races carry much of that: Henry Cuellar (TX-28) and Bobby Pulido (TX-15), both in the moderate-sounding third, ran 29.6 and 28.8 points ahead. Without them the moderate third's average falls to +0.8, and the estimate under D+15 to −0.5 (−1.8 to +0.9), within noise.
Hover over or tap any dot for that nominee's typical post.
We scored 24 Democratic governor nominees the same way and put them on the same scale: every race at least 95% counted, leaving out four where other candidates took more than 10%. How progressive a governor nominee sounded did not line up with how they ran: +0.1 points per step with the same controls (−4.6 to +4.8). Re-read with names masked, and measured like the Senate re-reads near the end of this page, it was +1.3 (−5.5 to +8.1); measured that way, the first reading is still +0.1. With 24 races, those intervals are wide enough to include the Senate's estimate as well as zero, so this test rules out little either way.
Where a Senate race and a governor's race shared a ballot, the comparison is direct: the same voters, the same day. It also has to allow for how the governor nominee sounded, so the chart places both nominees on the scale. In the 17 states where other candidates took no more than 10% in either race, the more progressive the Senate nominee sounded than the governor nominee, the further the Senate margin fell below the governor's: 3.8 points per step of difference (−6.6 to −0.9). In five of the six states where the Senate nominee sounded most progressive next to the governor nominee, the Senate nominee got the smaller margin. In eight of the other eleven, the bigger one. Put the other way, whichever of the two Democrats sounded more progressive got the smaller margin in 13 of the 17 states, including all five where that was the governor nominee.
Back in the Senate, populist language showed up across the field, from the most progressive-sounding nominees to the most moderate, though far more often at the progressive end. Did it also correlate with electoral performance? To check, Anthropic’s Claude Sonnet, from the same family of AI models that placed the nominees on the scale, read each of the Senate nominees' 10,163 October posts with enough text to read and answered two questions: does the post lean toward the party's progressive wing, and does it frame politics as working people against the wealthy, corporations or a rigged system? Read one post at a time, posts that are progressive, populist or both fall into three kinds: progressive and populist, progressive but not populist, and populist but not progressive.
The share of posts that were progressive but not populist correlated with running behind at least as closely as the share that were progressive and populist, and posts that were populist but not progressive showed no clear link. About three-fifths of the progressive-but-not-populist posts came from the four races that carry the Senate result; without them, that link keeps most of its size (2.7 points further behind per step of about 5 percentage points of the feed, against 3.8 with all 28) but is on the edge of chance (p 0.050). In 28 races we cannot say whether populist framing without the progressive lean correlated with running behind.
Each nominee's October feed is 100 squares, each about 1% of their posts, in order from the nominee who ran furthest behind the typical shift to the one who ran furthest ahead. For most nominees, most squares stay gray: posts that did not lean progressive and used no populist framing, such as get-out-the-vote posts, rally photos, biography and thank-yous.
Progressive and populist. 1,755 posts, 17% of all the Senate posts, lean progressive and use populist framing.
Progressive, not populist. 596 posts, 6%, lean progressive without populist framing. Immigration, the economy and health care come up most, each in about a fifth of them.
Populist, not progressive. 681 posts, 7%, use populist framing without leaning progressive; 655 of them use it only lightly rather than strongly. About half are about the economy.
Measured one kind at a time, with the same controls as above, a larger share of progressive posts without populist framing correlated with running behind: −3.8 points per step (−5.6 to −2.0), a step being about 5 percentage points of the feed. So did progressive populist posts, −2.2 (−3.8 to −0.6) per step of about 12 percentage points. Populist framing on posts that were not progressive was within noise, −1.3 (−3.1 to +0.5). Measured together, the progressive share without populist framing holds at −3.3 (−5.6 to −1.0), and the progressive populist share falls to −0.6 (−2.4 to +1.1), within noise.
Hover over or tap a coloured square to read a post of that kind from that feed; a few have no post to show.
The pile. Every post the 191 Democratic nominees published from Oct 1 to Nov 2: 57,100 posts on X, Facebook, Instagram and TikTok.
The filter. Their own posts, not shares or ads: 51,879. With at least 100 characters of text, enough to read: 42,588.
By candidate. Sorted into 191 feeds, one per nominee.
The sample. Ten posts drawn at random from each feed, each cut at 260 characters: 1,910 posts in all.
One judgment. A reader sees two feeds side by side, labelled only A and B, and answers one question: which candidate is closer to the progressive wing of the party? The reader is an AI model (Anthropic's Claude Sonnet) given a fixed list of what counts on each side. Attacking Trump counts for neither.
A thousand judgments. 1,019 such comparisons link every candidate to about ten others. Shown again with the sides swapped, 91% of 120 comparisons got the same answer. Two other Claude models, Opus and Haiku, re-judged the Senate pairs from the first masked draw (described in the next section) and agreed with Sonnet on 91% and 85% of picks.
The scale. A standard model for paired comparisons (Bradley-Terry) turns the picks into one position per candidate, with an uncertainty of about 0.3 on a scale where the typical candidate sits within ±1. That covers the judging only. Drawing a different set of posts moves a nominee too: across six separate draws for the 28 Senate nominees, a nominee's place had a standard deviation of about 0.3 on this scale. G. Elliott Morris built a similar scale from campaign websites at Strength In Numbers; ours ranks the 29 Senate Democrats on both lists in close to the same order (rank correlation 0.88). Among sitting House members, those in the Progressive Caucus alone averaged +0.9 on our scale, New Democrats −0.5 and Blue Dogs −1.2.
The result. For each race: the 2026 margin, minus Harris's 2024 margin in the same state or district, compared with the typical shift for nominees of the same party, office and incumbency. Above zero ran ahead of it, below zero behind.
The checks. The Senate result, re-estimated ten ways. Scroll the chart to see every row.
Three questions about the Senate result. Was the AI model that read the feeds reacting to the candidates' names rather than their posts? Did the progressive voice show up only in the final month? And which lines up more closely with running behind: how much of a feed leans progressive, or how far left its progressive posts go?
Could the AI model that read the feeds have been reacting to ideological associations with names rather than posts? We re-read all 28 feeds from scratch: a fresh draw of 12 posts each, with names, opponents, states and links masked. The order held (rank correlation 0.87 with the first reading). Once the names were gone, Chris Coons moved most, from twenty-fourth of 28 to fourteenth. Near the top, Sherrod Brown went from second to seventh. Five more fresh draws of posts mostly agreed at the top: Markey, Flanagan, El-Sayed and Nixon held the top four places in four of the six masked readings, and at least three of them did in every one. The result held too: masked, the estimate is about the same as the first reading’s.
Could the voice have come only in the final month, from nominees who were already falling behind and trying to mobilize their progressive base? Using only September posts gives the same answer: −1.8, against −1.9 for October posts measured the same way (the note under the chart gives the interval). The voice was there before the final month.
How much of a feed leans progressive and how far left its progressive posts go need not move together: one nominee might voice positions far to the left only now and then, while another fills the feed with milder progressive content. Read one post at a time, as in “Progressive, or populist?”, the two can be counted separately: the share of the feed that leans progressive (below, simply the share), and how often the reader judged those posts clearly progressive rather than only leaning that way (how far left).
The share is what lines up with running behind. How far left the posts go is weaker and within noise. Held side by side with the share, it adds no clear link of its own. The test is thin, though: few nominees have much of one and little of the other, and without the four races that carry the Senate result neither measure is clear of noise.
Each dot is a nominee. Across: of their progressive posts, the share the reader judged clearly progressive. Up: the share of their whole feed that leans progressive. The two rise together. Nine nominees sit at zero across, with no post judged clearly progressive; Adam Hamilton and Everett Wess posted nothing progressive and are not shown.
The four races that carry the Senate result sit further across and higher up than most of the other nominees. Three of them gave more of their feeds to progressive content than any other nominee, and all four are among the five with the most of it judged clearly progressive. Of the 604 posts the reader judged clearly progressive, 473 came from these four.
The shaded areas hold the two kinds: one nominee as far left as one of the four but with less of the feed progressive than any of them, and two with more of the feed progressive than Flanagan, one of the four, but a far smaller share of their progressive posts judged clearly progressive. Three races cannot settle the question.
Measured with the same controls as the Senate result, a feed with a larger share of progressive posts correlated with running further behind. How far left, which counts only the 26 nominees with progressive posts, was weaker and within noise. Without the four races, each measured on its own, neither is clear of noise.
Hover over or tap any dot for that nominee's numbers and a typical post.
Data: Context's crawl of candidates' public posts (Oct 1 to Nov 2, 2026), AP results via Fox News (Nov 8, 04:30 UTC), FEC filings for fundraising through Sep 30, FEC and Wikipedia for 2024 presidential results, The Downballot for 2024 House district results, Sabato's Crystal Ball race ratings. Thanks to G. Elliott Morris, whose website-based scale we compared ours against.