Agentik Momentum's win probabilities for six 2026 Senate races, built only from how AI answer engines talk about each candidate, shown next to three prediction markets and the RealClearPolling average.
| Race | Senate SignalChance to win | KalshiChance to win | PolymarketChance to win | PredictItChance to win |
|---|---|---|---|---|
| Loading the latest numbers… | ||||
Market prices include the gap between buy and sell prices, so a market's two sides can add up to slightly more than 100%.
Senate Signal combines two readings from our AI tracking, which puts a fixed set of questions about each race to ChatGPT, Gemini and Google AI Overviews every day. Each reading becomes the Democrat's share of that signal, between 0 and 100 percent. We blend the two with fixed weights and turn the result into a win probability. Senate Signal uses AI data only. The RealClearPolling average and the three prediction markets sit beside it for comparison and never feed into it.
Sentiment share. For each candidate, net sentiment is the percent of AI answers that are positive about them minus the percent that are negative, over the last 7 days, including today. Senate Signal is recalculated every hour as new answers arrive. The answers come from issue questions that name one candidate at a time, and both candidates get the same questions. We leave out what the answers say about polls, fundraising and campaign momentum, so the reading reflects how AI describes each candidate and not how the race is going.
Visibility share. How often AI answers bring up each candidate in questions that name no one, such as "Which Michigan Senate candidate would do more to fight inflation?" None of these questions asks about polls or who is winning. In these races AI names both candidates at similar rates, so visibility moves Senate Signal very little.
Senate Signal is built to be about as uncertain as Senate polling, which has missed the final margin by 5.4 points on average since 1998. A modest edge in how AI describes a candidate becomes a modest win chance, not a sure thing. We will score Senate Signal against the November results.
The Polling view also shows Senate Signal as an estimated vote share, so it can sit next to the polling average. We turn the win chance back into a margin with the same uncertainty as the forecast and split it evenly on either side of 50%. The estimate covers the two-party vote only, so it leaves out third parties and undecided voters, which the polling average includes.
Senate Signal's forecast locks at 7:00 PM ET on November 3, when the first polls close in these six states, in Georgia. The market prices and the RealClearPolling average freeze at the same moment, so the board shows what each source said before any votes were counted. A race counts as called when the Associated Press calls it. Each called race shows the winner, and a gold outline marks the races where Senate Signal had the winner ahead.
AI answers reflect the news, polls and commentary those engines draw on. They don't ask voters anything. Senate Signal measures how the information environment leans, as a reading separate from polls and markets. If a reading is missing for a race, its weight is shared out among the other readings, and the card says so.
Tone is the only reading with a direction, favorable or unfavorable, and it is the closest AI equivalent to a favorability rating. Every candidate gets the same questions in matched pairs, so a gap between them comes from how AI describes them. Favorability tracks vote share only loosely, which is why the scale is wide.
Visibility shows which candidate AI brings up when nobody is named. It mostly reflects incumbency and name recognition, and in these races it barely separates the candidates. It keeps a small weight so that a real gap would still register.
The scale and the reasons behind each weight come from published research on election forecasting. None of the weights is fit to this year's polls or markets. Research: Gayo-Avello on social media election forecasts, Graefe et al. on combining forecasts and Nate Silver on Senate poll error (FiveThirtyEight, 2018, author's republished copy).