A test grading 7,449 answers from ChatGPT, Gemini, Google AI Mode, and Perplexity against official government data found an overall false-answer rate of 8.2 percent, roughly one false answer in 12. Perplexity was wrong most often at 22.6 percent, followed by Gemini at 6.2 percent, ChatGPT at 4.4 percent, and Google AI Mode at 2.9 percent, according to Tom's Guide reporting of the Searchable study.
Searchable tested 106 underlying questions across five metropolitan areas and their respective states, checking answers against official government data. The testing took place in September and October 2026, and about 62 percent of the 43,079 webpages accessed during the test were federal, state, or local government pages. That grounding matters: the study measured answers against authoritative records, not against another model opinion.
The scope is also the caveat. These rates apply to these questions and locations, not to every election question or product version. Still, a sevenfold gap between the highest and lowest false-answer rates is too large to dismiss as noise, and it gives users a concrete reason to cross-check political answers rather than trusting any single assistant.
The study does not by itself explain the gap, but the pattern fits what is known about retrieval design. Assistants that ground answers tightly in official pages and show their sources tend to do better on factual civic questions, while assistants that synthesise more freely across the open web have more room to drift. Google AI Mode, built on Google search retrieval, had the most direct path to the government pages that dominated the test corpus.
For users, the lesson is to prefer answers that cite primary sources and to click through on anything that affects a vote, a deadline, or a location. A cited government page can be verified in seconds; an uncited paragraph cannot. Publishers of civic information should note the corollary: being the cited primary source is worth more than ever, because the assistant that cites you is also the one most likely to get the answer right.
For SEOs, the study reinforces a simple hierarchy: primary sources win. Pages that republish official data with clear structure, dates, and jurisdiction labels are the easiest for retrieval systems to ground on, and the hardest for lookalikes to imitate convincingly. Structured data, clear headings, and verbatim official figures all help an assistant choose your page over a rewritten copy.
It also argues for humility in AI visibility reporting. A brand can be cited frequently and still be part of wrong answers if the citation is decorative rather than grounding. The metric that matters is not citation count alone but citation quality: is your page the source the answer was built on, or just one of several links appended afterwards?
The practical advice for voters is simple and does not depend on which assistant scored best. For anything that affects a vote, such as polling places, hours, deadlines, or ballot measures, verify the answer against an official government source before acting. The test itself shows why: about 62 percent of the pages accessed were government pages, which means the correct information was usually retrievable, and the failures were in synthesis and selection rather than in availability.
For publishers of civic information, the study is an argument for publishing official data in the most retrievable form possible: plain-language pages, clear jurisdiction labels, up-to-date dates, and structured tables that an assistant can quote without paraphrasing away the precision. A polling-place page that buries the address in a paragraph is harder to ground than one that states it in a table, and on civic questions that difference can decide whether a voter gets the right answer.
Sources: NeoTeo, November 1, 2026, reporting Tom's Guide coverage of the Searchable study.
Linkbaits helps you create content that earns citations and backlinks. Start with the free tools and build your first linkable asset this week.
Start free trial