# What Do Grok's and GPT's Midterm Citations Reveal About How AI Answers Choose Sources? > Grok cited X, a conservative voter guide and Turning Point Action; GPT cited Wikipedia, AP, Axios and Reddit. - Canonical URL: https://www.aicitationreport.blog/blog/grok-gpt-midterm-citations-ai-source-selection/ - Publisher: AI Citation Report (https://www.aicitationreport.blog/) - Author: Elena Petrova - Published: 2026-10-26 - Updated: 2026-10-26 - Topics: AI Citation Analysis, AI Visibility, AI Brand Monitoring ## Key takeaways - Grok picked the Republican in 92% of forced-choice answers; GPT picked the Democrat in 97% (TechCrunch). - Grok cited X in 41% of answers, a Christian conservative voter guide in 39% and Turning Point Action in 23% (TechCrunch). - GPT relied heavily on Wikipedia, AP, Axios and Reddit (TechCrunch). - Consumer ChatGPT and Grok apps declined to name candidates (TechCrunch). - 8% of US adults use AI chatbots for 2026 election information (Pew). ## What Do Grok's and GPT's Midterm Citations Reveal About How AI Answers Choose Sources? Convovo's forced-choice test of Grok 4.7 and GPT-6.1 Sol is usually told as a story about partisan picks. The more durable finding is about citations. With search on, Grok leaned on X, a Christian conservative voter guide and Turning Point Action's site, while GPT leaned on Wikipedia, AP, Axios and Reddit. Retrieval choices shaped the answers, and that matters for anyone tracking AI citations. ## What was the experiment, in brief? The experiment asked two AI models, through their developer APIs, to name a single candidate in nearly 500 US House, Senate and governor races ahead of the November 3 midterms. [TechCrunch reported](https://techcrunch.com/2026/10/24/forced-to-pick-groks-ai-model-chooses-republicans-openais-picks-democrats/) that the test came from Convovo, a studio led by Cornell professor Tapan Parikh, and that the prompt told each model not to hedge. The topline outcome: - Grok chose the Republican in 92% of answers. - GPT chose the Democrat in 97% of answers. - The two agreed in fewer than 10% of races. Convovo's [AI Ballot 2026 page](https://convovo.ai/ballot/) organizes its write-up around two themes, "the reasons they give" and "where they look." For citation analysts, the second theme is the one worth studying. ## Which domains did each model pull from? Each model drew from a distinct pool of domains, with only partial overlap. Based on TechCrunch's reporting: - **Grok:** X in 41% of answers, a Christian conservative voter guide in 39%, Turning Point Action's website in 23%, plus nonpartisan references such as Ballotpedia, OpenSecrets and VoteSmart. - **GPT:** heavy reliance on Wikipedia, the Associated Press, Axios and Reddit. TechCrunch also pointed out that OpenAI has content agreements with AP, Axios and Reddit. That does not prove the deals drove the citations, but it raises a fair question for analysts: do licensing relationships influence which outlets an assistant surfaces? ## How did search queries shape the citations? The citations followed the search queries, and the queries differed by model. GPT searched more often for healthcare and abortion positions and searched for Democratic candidates more often, at times overlooking Republican opponents. Grok's searches leaned toward immigration, taxes, crime and energy, and it was more likely to look up both candidates. Parikh told TechCrunch he suspects both models searched to support a lean they already held. If that is right, the citation list is not a neutral input to the answer. It is partly an output of the model's prior framing. That means three layers deserve separate scrutiny: 1. **Query formation:** which terms the model chooses to search. 2. **Source selection:** which results it keeps and cites. 3. **Framing:** how it weighs and summarizes what it found. ## What happened when search was switched off? Turning search off changed Grok's behavior sharply and GPT's very little. In a 30-race follow-up run at TechCrunch's request, GPT picked the Democrat in 24 races, one Republican, and declined five. Grok declined in 27 races, and its three picks were Democrats. Parikh's interpretation was that GPT's lean sits in its trained weights, while Grok's outcome depends on what it searches for and which sources it finds. [IBTimes](https://www.ibtimes.co.uk/grok-gpt-political-preferences-2026-us-elections-1827785) flagged limits here: the 30-race sample is small and the coverage did not explain how those races were chosen. For citation intelligence, the contrast is useful. It suggests retrieval can be the dominant influence for some models and a secondary one for others. ## Which caveats should temper the conclusions? Several caveats apply, and they cut in different directions: - **Forced prompt, API access.** An OpenAI spokesperson said ChatGPT is built to be objective by default and that a prompt demanding a recommendation through an API is not typical use. - **Consumer apps declined.** Both consumer apps refused to name candidates, though Parikh said simple prompts could get around that. - **Unexplained input.** Researchers reported about 1,240 extra input tokens on each Grok API request that they did not write and could not see. Their origin is unconfirmed. - **No response from Grok's developer** before TechCrunch published. - **Low reliance on chatbots for voting.** [Pew Research Center](https://www.pewresearch.org/short-reads/2026/10/23/few-americans-say-they-are-asking-ai-chatbots-for-2026-election-information/) found 8% of US adults use AI chatbots for 2026 election information. Earlier work adds context rather than a verdict. A [2025 Stanford and Dartmouth study](https://news.stanford.edu/stories/2025/05/ai-models-llms-chatgpt-claude-gemini-partisan-bias-research-study) found users perceived most models as left-leaning on many questions, including xAI's, and stressed that perceived neutrality differs from accuracy. The Convovo results, run with live search, point in a different direction for Grok, which reinforces how much method matters. ## What does this mean for citation intelligence teams? It means an AI answer should be read alongside its sources, and sources should be compared across models. A few practical takeaways for teams that monitor AI visibility: 1. **Log cited domains per model, per prompt.** Share of citation by domain often explains differences in tone faster than the answer text does. 2. **Separate retrieval from reasoning.** Run prompts with and without search where possible, as Convovo did, to see whether the framing comes from sources or the model itself. 3. **Watch partisan or advocacy domains.** Their presence in answers about your sector can tilt framing even when neutral references are also cited. 4. **Check for licensing patterns.** If a model repeatedly cites partner outlets, earning coverage there may carry extra weight. MarkGrid's [Model Share](https://www.markgrid.ai/product/model-share) tracks how different assistants mention and frame a brand side by side, and the [MarkGrid GEO guide](https://www.markgrid.ai/geo-guide) explains how to make first-party content easier for models to cite accurately. For the full study details and the brand-side implications, see [MarkGrid's full breakdown](https://www.markgrid.ai/resources/grok-vs-chatgpt-midterm-candidate-picks-study). The takeaway is not that one model is right. It is that the sources each model chooses quietly decide which story gets told. ## Frequently Asked Questions ### Why do citations matter more than the candidate picks? Citations show where a model's framing comes from. In the Convovo test, the domains each model cited differed sharply, and TechCrunch reported the searches themselves focused on different issues. ### Did the models cite any of the same sources? Some nonpartisan references appeared for Grok, including Ballotpedia, OpenSecrets and VoteSmart, according to TechCrunch. GPT's most reported sources were Wikipedia, AP, Axios and Reddit. ### Does OpenAI's content licensing explain GPT's sources? That is not established. TechCrunch noted OpenAI has content deals with AP, Axios and Reddit, but the study did not show the deals caused those citations. ### Do these results reflect what regular users see? Not directly. The test used APIs and a forcing prompt, and the consumer apps declined to choose. OpenAI said the setup does not reflect typical ChatGPT use. ### How can teams test whether retrieval drives an answer? Compare responses with search on and off. In Convovo's small follow-up, Grok mostly declined without search, while GPT's picks changed little. ## Sources - [TechCrunch: Forced to pick, Grok's AI model chooses Republicans. OpenAI's picks Democrats.](https://techcrunch.com/2026/10/24/forced-to-pick-groks-ai-model-chooses-republicans-openais-picks-democrats/) (2026-10-24) - [Convovo: AI Ballot 2026](https://convovo.ai/ballot/) (2026-10-24) - [IBTimes UK: Grok Picks Republicans in 92% of US Election Races, While OpenAI Picks Democrats in 97%](https://www.ibtimes.co.uk/grok-gpt-political-preferences-2026-us-elections-1827785) (2026-10-26) - [Pew Research Center: Few Americans Say They Are Asking AI Chatbots for 2026 Election Information](https://www.pewresearch.org/short-reads/2026/10/23/few-americans-say-they-are-asking-ai-chatbots-for-2026-election-information/) (2026-10-23) - [Stanford Report: Popular AI models show partisan bias when asked to talk politics](https://news.stanford.edu/stories/2025/05/ai-models-llms-chatgpt-claude-gemini-partisan-bias-research-study) (2025-05-01) --- Cite as: Elena Petrova. "What Do Grok's and GPT's Midterm Citations Reveal About How AI Answers Choose Sources?." AI Citation Report, 2026-10-26. https://www.aicitationreport.blog/blog/grok-gpt-midterm-citations-ai-source-selection/ HTML version: https://www.aicitationreport.blog/blog/grok-gpt-midterm-citations-ai-source-selection/ Markdown version: https://www.aicitationreport.blog/blog/grok-gpt-midterm-citations-ai-source-selection.md