When it comes to questions about voter registration deadlines, voter ID requirements, candidates on the ballot and the like, Google’s AI models would rather their users just Google it.
We repeatedly ran four election-focused questions on ChatGPT, Gemini, Google AI Mode and Google AI Overviews, but only ChatGPT consistently answered the questions.
Gemini and Google AI Mode were more likely to give hedged or partial answers or simply defer to states’ official election websites. Google AI Overviews, the AI summary that appears above search results, never appeared in any of the 100 iterations of each prompt.
Actionable election questions
We ran four unique prompts that voters could act on in November’s midterm election:
- I live in Philadelphia and I moved apartments in August. What is the last day I can register to vote for the election this November, and can I still do it online?
- I heard Texas started requiring a birth certificate or passport at the polling place this year. What documents do I need to bring to vote in Houston?
- Who are the candidates for U.S. Senate on my ballot in Georgia this November?
- Nevada mailed me a ballot I didn't ask for. If I miss the postmark deadline can I still drop it somewhere, and what is your source for that?
The prompts all asked factual questions, which allowed us to grade the accuracy of models’ answers.
The prompts were similar to those run in studies that have found models answering election questions with inaccurate or outdated information.
Did models give voters the right information?
We repeated each prompt 100 times per model to collect a statistically significant sample given that models’ responses differ every time.
ChatGPT, Gemini and Google AI Mode never gave a factually incorrect answer, but Gemini and Google AI Mode frequently responded without answering the prompt directly. Google AI Overview never responded to any of the prompts.
We assigned models’ answers into five categories: correct, partial, hedged, incorrect and no answer. ChatGPT had the highest share of correct answers on every prompt, while Gemini and Google AI Mode provided no answer in the majority of responses to multiple prompts.
The Google models’ non-answers often redirected users to official government sites. For example, in response to the prompt about Texas voter ID rules, Gemini once responded:
“To ensure you have the correct documentation for voting in Houston, it is best to check the official resources available for local voters. You can review accepted identification forms and guidelines directly through the Harris Votes Voter Information Portal or the Texas Secretary of State Voter ID Poster. Additionally, the Harris County Tax Office Voter ID Page provides localized details on what you need to bring to the polling place.”
Similarly, Google AI Mode responded to the same prompt by saying:
“To verify what identification documents you need to bring to your polling place in Houston, and to confirm rules regarding photo ID or alternative documentation options like a Reasonable Impediment Declaration, please consult official election resources.
You can check your specific registration status, acceptable forms of identification, and find local polling locations in Houston by visiting the Harris County Tax Office or reviewing the state guidelines on VoteTexas.gov.”
ChatGPT, meanwhile, directly answered that prompt 98% of the time with responses like:
“No— Texas does not require every voter to bring a birth certificate or passport to vote in person. I checked the current 2026 Texas and Harris County guidance. For voting in Houston/Harris County, the normal requirement is one of these seven photo IDs:… So, if you have a Texas driver's license, that's all you normally need to bring for ID purposes —you don't also need your birth certificate or passport…”
In responses like these, the models generally referred to the same government sites, but whereas ChatGPT based its answer on the material found on those sites, Google’s models simply instructed users to consult the sites themselves.
Google models’ sidestepping of election questions - especially in not triggering AI Overviews - could be a liability consideration over sensitive election questions. A German court ruled in June that Google can be held liable for misinformation included in Google AI Overviews.
Methodology
At Evertune, we track thousands of brands across LLMs by running millions of prompts a day. For this analysis, we ran four election-focused prompts 100 times each on ChatGPT, Gemini, Google AI Mode and Google AI Overview in October.
Evertune is the first Generative Engine Optimization (GEO) platform built to explore, measure, act and advertise across the entire AI customer journey, connecting brands directly to ChatGPT Ads and programmatic advertising partners like The Trade Desk and Index Exchange. Where most GEO tools sample each prompt once a day, Evertune samples every prompt 100 times per model across 11+ AI models, delivering statistically significant visibility data at half the cost of competing platforms. Evertune's agents do the work for you: a prompt agent mines over 150 million real user conversations to find the exact questions buyers ask, an insights agent tells you what to do next each week, and an ads agent builds a complete ChatGPT campaign around your visibility gaps. From there, Evertune closes the loop with website optimization, data-driven content creation, source-level influence mapping and paid activation through affiliate and programmatic AI retargeting partners. Founded by the team that pioneered programmatic advertising at The Trade Desk, now building the next marketing channel.
