How we work
We asked AI products election questions daily and in repeated runs, kept what they said and the sources they used, and classified them the same way for each product.
What we tested
We tested ChatGPT, Claude, Gemini, Grok, Perplexity and Google's AI Overview in two ways. Through the web products, driven in a real browser, free and paid, logged in and logged out where each exists. And through each company's developer API, using the model and settings closest to each product, with web search on where it is offered. Not every product was tested both ways. We recorded which version gave each answer.
What we asked
Candidate lists for U.S. Senate, governor and competitive House races came from Ballotpedia. Two main sets of questions ran on them:
- Candidate comparisons, asked from Oct 4 until polls began closing on Election Day, daily from Oct 10, which show what sources each product uses. API versions were asked about all 123 races, and 14 kinds of question about 6 of them; web versions about those 6 races, and some web versions missed days, including ChatGPT's in the last week of October.
- The pick experiment, a bare "Candidate A or Candidate B?" with the names in both orders, repeated in separate runs to see whether results hold.
We also ran shorter tests of question wording and account tiers in October, and a wider set of election questions from the night of Nov 2 through Nov 3, including ballot measures and state legislative races. These count toward our totals but not our source charts.
How we classify
An AI model sorts each pick answer into one of seven outcomes: picks a candidate, conditional pick, horse race, neutral comparison, asks a question, declines, or other. That model is from Anthropic's Claude family, and Claude is also one of the products we tested. Websites are sorted into types by the same model, along with a maintained list, automatic rules and corrections from our own audit. Both are imperfect, which is why signed-in visitors can read the answers behind every figure. We recovered 86 Claude web answers that our tool had only partly captured, using saved pages. Seven more stopped at a permission prompt with no answer, and we do not count them.
How to read the results
- AI answers vary. The same question can get a different answer when asked again, so we report shares across many answers.
- Products change without notice. Results describe what a product did on the dates shown.
- Figures on these pages are computed once a day and carry their dates.
- Most of the source links we recorded come from API versions, so source figures mostly describe those versions rather than the web apps.
- ChatGPT web link counts are partial: our tool captured only some of the links shown. In our source charts, the Grok row comes from SuperGrok, and its link counts are partial.
The method notes have more detail, including why counts differ from page to page.
The survey
The figures on how many Americans use AI for election information come from a separate two-wave national survey by SSRS, not from our testing. Full question wording, toplines and methodology are in the survey report.
Access to the data
Summaries are public. Answer-level data and downloads are available to anyone who signs in with a Google account. We keep a record of which signed-in visitors viewed which pages.