Who's winning the midterms? Depends which AI you ask
We asked five AI assistants the same 80 questions voters are asking about the 2026 U.S. midterms, from Senate toss-ups to candidates' stances on tariffs, on three consecutive days. Then we read every answer, counted which candidates each assistant names, and traced every source behind them. The assistants often disagree on who is ahead and change their calls overnight, Google's AI Mode mostly declines to say, Gemini stopped calling races altogether by the third day, and three in five citations come from traditional newsrooms.
Ten things we learned
Perplexity is the most decisive, and leans Democratic
Asked the 26 race questions on each of three days, Perplexity named a favorite or a toss-up in 76 of 78 answers and never refused. It called 40 for Democrats and 14 for Republicans, against 32 and 16 for ChatGPT and 13 and 5 for Gemini. When the assistants explicitly recommend candidates, Democrats make up 73% of Perplexity's picks, 71% of ChatGPT's and 70% of Gemini's.
Google AI Mode mostly refuses to pick a favorite
Asked who is favored in 16 Senate and 10 governor races on each of the three days, ChatGPT and Perplexity named a favorite or called a toss-up in 71 and 76 of 78 answers. Google AI Mode did so in 18, between five and seven a day. In most other cases it listed the candidates and told users to consult news outlets, and 14 times it declined outright. In Maine it said such predictions "cannot be verified."
Gemini stopped talking about the candidates
On 28 September Gemini named a favorite or a toss-up in 22 of 26 race questions. On 29 September it did so in 6, and on 30 September in none. It still answered every question, but 24 of its 26 race answers that day did not name a single candidate: they pointed users to Ballotpedia, pollsters, local news or official election sites instead. Its average answer shrank from 1,450 characters on the first day to under 800, and James Talarico, named in 11 of its answers on the first day, appeared in 3 and then 4.
The same question gets different answers, and they keep changing
ChatGPT, Gemini and Perplexity disagreed on whether anyone has the edge in six of the 11 battleground Senate races on 28 September, in ten on 29 September and in all eleven on 30 September, largely because Gemini stopped calling races. Verdicts moved from day to day too: over the three days, 12 of ChatGPT's 26 race questions got more than one verdict, and 8 of Perplexity's. In Alaska, ChatGPT went from a toss-up to naming Republican Dan Sullivan the favorite and back to a toss-up, while Perplexity went from Democrat Mary Peltola to a toss-up and back to Peltola. In Texas, Perplexity said Democrat James Talarico is favored on the first two days and called it a toss-up on the third; ChatGPT called it a toss-up all three times.
The press still writes the AI answer
Three in five citations (6,992 of 11,616) point to news outlets and independent journalism. When we could classify the page type, 71.6% of cited pages were news articles. Explainers and guides made up 11.6%, and landing pages 4.8%.
A small club of sources dominates
The ten most-cited domains account for 22.3% of all citations, and the top 25 for 39.5%. The New York Times leads with 350 citations, followed by The Washington Post (334), Ballotpedia (331) and NBC News (322).
Race trackers and forecasters are the reference layer
AI leans on rankings and aggregators for "who is favored" questions: Ballotpedia, 270toWin, Cook Political Report and The Hill's Senate rankings. Their race ratings travel straight into AI answers.
Pollsters are the names AI repeats
Among organizations named inside the answers, pollster Siena appeared in 2.2% of answers, and YouGov and Marist in 2.0% each. The political group named most often was MAGA Inc., with 36 mentions and a 6.6% share of organizational mentions.
Social media barely registers
For all the campaign content on social platforms, AI assistants rarely cite it: 210 citations in total. Facebook (99) and YouTube (83) lead. Reddit was cited six times and X once.
Campaigns and advocacy groups are mostly spoken for
Sites run by campaigns, PACs and advocacy organizations drew 360 citations over three days. AI describes candidates through what reporters and trackers write about them, and much of it is recent: heavily cited pages include Reuters, The Washington Post and The 19th stories published in September 2026.
80 questions in seven categories
Every question was phrased neutrally and asked of each assistant once a day, on 28, 29 and 30 September. The share of Democratic mentions shows how the named candidates split by party within each category (Republican share is the remainder; independents excluded).
| Category | Qs | Example question | Most-named candidates | Dem. share |
|---|
Five assistants, five answering styles
ChatGPT writes the longest answers and names the most candidates per answer. Perplexity names the widest range of people and calls almost every race. Google AI Mode makes few race calls, Gemini's answers halved in length after the first day, and by the third it answered election questions mostly with links to other sources, and Google AI Overviews did not appear at all for 146 of 240 questions.
How each assistant called the key Senate races, day by day
Answers to "Who are the candidates in the 2026 [state] Senate race and who is favored to win?" asked on 28, 29 and 30 September. Where an assistant reached the same verdict on all three days, one label covers them; where it changed its call, each verdict is shown with the days it held. Hover a label for the sentence it rests on.
| Race | ChatGPT | Gemini | Perplexity | Google AI Mode | AI Overviews |
|---|
Which assistant leans toward which party
Three measures, each split by party, over all three days. On plain mentions all five assistants are close to even. The gap opens when they judge. Every assistant recommends more Democrats than Republicans, and every assistant that calls races calls more of them for Democrats. Perplexity leans furthest among the assistants that answer consistently. Google AI Mode mostly declines to judge at all, and from the second day on Gemini joined it.
1. Candidates mentioned
2. Candidates recommended
Gemini, Google AI Mode and AI Overviews recommend few candidates at all, so their shares rest on small numbers.
3. Race calls
How to read this. A tilt in these numbers is not proof of partisan bias. Many of 2026's most competitive Senate seats are held by Republicans, so questions about flips, underdogs and momentum naturally surface Democratic challengers. Several governor races in the sample, such as New York, California and Pennsylvania, have a Democrat whom all the answering assistants describe as a clear favorite. The assistants repeat what polls, forecasters and news coverage report. The differences between assistants asked identical questions on the same days, and within one assistant from one day to the next, are the finding.
The candidates AI names most
Number of answers (out of 1,053 with text, over three days) that name each candidate, across all five assistants. Office-holders not on the 2026 ballot, such as Donald Trump and J.D. Vance, are excluded.
Top 15 candidates by answers naming them
Who each assistant names first
The first three candidates of the relevant party that appear in each answer, in order, on each of the three days. Order of appearance is a signal of prominence, not an endorsement, and a name can appear as context (for example, the seat a candidate is running for).
| Question | Day | ChatGPT | Gemini | Perplexity | Google AI Mode |
|---|
Citation sources by category
Share of all citations, by source type
Share of classified pages, by content type
The 20 most-cited domains
Every campaign, PAC and advocacy site in the study, taken together, was cited about as often as The New York Times alone.
The individual pages AI cites most
Race-overview pages win. The single most-cited URL is CBS News' list of 12 key Senate races, ahead of NBC News' midterms hub, a second CBS News list of races to watch and a Brookings analysis of the primaries. Most of the rest are "races to watch" lists and race ratings, the format that answers "who is favored" in one place.
| # | Page | Type | Citations |
|---|
Implications for the people shaping the race
Campaigns
Your own website is a minor input. What AI says about your candidate is built from news coverage, Ballotpedia profiles and race ratings. Keep those accurate and current, and give reporters clear, quotable policy positions.
Newsrooms
Race explainers and "races to watch" lists are being read by machines as much as by people. Clear dates, structured candidate lists and plain statements of who is favored make coverage easier for AI to cite correctly.
Civic and voter-info groups
Nonpartisan guides compete for a small number of slots dominated by large publishers and Ballotpedia. Earning mentions in coverage from those sources is the most direct route into AI answers.
More voters now ask an AI assistant who is running and who is ahead before they read a single article. The five assistants we tested gave different answers to the same question, changed their answers from one day to the next, and two of them mostly declined to say who is ahead. What they say is built almost entirely from news coverage, race ratings and Ballotpedia, not from campaigns. Anyone who wants to be described accurately by AI has to be described accurately in the sources AI reads.
Mateusz RzeteckiFounder, GetScorra.com & Organicly.agency
Headline facts
- Asked who is favored in 26 Senate and governor races on 28, 29 and 30 September, Google AI Mode named a favorite or a toss-up in 18 of 78 answers and explicitly declined in 14. ChatGPT and Perplexity did so in 71 and 76.
- Gemini called 22 of 26 races on 28 September, 6 on 29 September and none on 30 September, when 24 of its 26 race answers named no candidate at all and only pointed to other sources.
- Across the 78 race answers, Perplexity called 40 for Democrats and 14 for Republicans; ChatGPT 32 and 16; Gemini 13 and 5.
- Of the candidates each assistant explicitly recommends, Democrats make up 73% for Perplexity, 71% for ChatGPT and 70% for Gemini. In plain mentions the split is close to even (51–56%).
- The candidates AI names most: James Talarico (185 answers), Abdul El-Sayed (162), Ken Paxton (158), Mike Rogers (142) and Susan Collins (128).
- 60.2% of 11,616 citations point to news media. The New York Times (350), The Washington Post (334) and Ballotpedia (331) are cited most.
Source: Scorra and Organicly, "Who's Winning the Midterms? Depends Which AI You Ask", September 2026. 1,200 answers from ChatGPT, Gemini, Perplexity, Google AI Mode and Google AI Overviews to 80 questions (US, English), collected 28–30 September 2026.
How we measured it
- Questions
- 80 neutral, informational questions a voter might type. They cover Senate races in 16 states (including North Carolina, Maine, Ohio, Michigan, Texas and Georgia), governor races in 10 states, at-risk House seats, retirements, primaries, fundraising, polling, and candidates' positions on the economy, immigration, abortion, tariffs, healthcare, AI regulation and public safety.
- Engines
- ChatGPT, Gemini, Perplexity, Google AI Mode and Google AI Overviews, U.S. market, English.
- Dates
- Three measurement runs with identical questions, on September 28, 29 and 30, 2026. Unless a figure names a day, it adds up all three. AI answers change daily, and many verdicts changed between runs; treat these figures as a three-day snapshot, not a trend.
- Classification
- Each cited domain was grouped as press (earned media), aggregator/directory, research or public body, campaign or advocacy site, social platform, or unclassified. Content-type shares exclude pages that could not be classified.
- Candidate mentions
- Scorra's answer analysis read all 1,200 answers and recorded every person named, with party, office and state, and the role each plays in the answer: recommended, neutral, background context or criticized. Name variants are merged into one person. Gemini answered all 240 questions, but from 29 September most of its election answers pointed to other sources instead of discussing candidates. Google AI Overviews did not appear for 146 of its 240 questions, leaving 1,053 answers with text. Party shares count Democrats and Republicans only; independents and people without a party are left out.
- Race calls
- For each "who is favored" question, the whole answer was read for its conclusion: a named favorite, a toss-up, no call (candidates listed without judging, or only sources given), or an explicit refusal. A heading such as "Who is favored?" is not a verdict. Each verdict is stored with the sentence it rests on. Every question was asked on all three days, so each assistant has 78 verdicts.
- Organization mentions
- Share of voice counts pollsters, PACs, party committees and advocacy groups named in answers. News outlets and directories are excluded from that measure. Individual candidates were not scored.
- Scope
- The study measures which sources AI assistants rely on. It does not assess whether individual answers were accurate or balanced, and it takes no position on any candidate or party.
This report was prepared by Scorra, which tracks what AI assistants say when people ask them questions, and Organicly (organicly.agency), an organic growth agency.
Is your brand in the answer?
Scorra measures how ChatGPT, Gemini, Perplexity and Google's AI describe your brand, every day.
Run a free scan