Whether a voter got an AI answer depended on the tool, the language and the wording of the question. Google's AI Overview answered — of election searches, while Google AI Mode produced an answer for —.
We ran the same election questions on five AI search tools several times a day, across 55 House, Senate and governor races, in English and Spanish. These figures count whether an AI answer appeared. We have not yet checked whether the answers were accurate.
An AI answer means generated text. Bing passages quoted word for word from one web page, and Google AI Mode's "request limit" messages, are not counted.
For "Republican position on [issue] 2026", Google showed an AI Overview in — of searches. For the same question about Democrats, it did so in —. The gap is concentrated in a few issues: on the economy it was — against —.
When we searched for both candidates in the same race at the same time, Google gave only the Democrat an AI Overview in — pairs and only the Republican in — pairs. On Bing, only the Democrat got a Copilot Answer in — pairs and only the Republican in —.
Each issue is one wording asked about 26 times, so the gap rests on three issues: the economy, abortion and healthcare. One possible cause is the word itself: "Democratic" is also an ordinary adjective, which a topic filter may treat differently from "Republican". We are testing other wordings.
On the same candidate comparisons, asked in the same collection rounds, Bing's Copilot Answer appeared on — of English searches and — of Spanish ones. Copilot chat and Google AI Mode answered both languages.
Spanish searches used Spanish-language settings (Google hl=es, Bing market es-US). Each Spanish question is matched with its English twin for the same race.
Google's AI Overview answered — of "who won" searches. On election night, Google's own answer panel filled — of result pages. Google AI Mode produced an answer — of the time, and Copilot chat —.
We have not yet checked whether these answers named the right winner, or whether they came before the Associated Press called the race.
All Google searches, each page counted once by its most prominent feature. Election night pulls asked mainly "who won" questions.
Copilot chat never answered "where should I get information about who to vote for": 0 answers in — searches without a location and — with one. It answered every time we asked whether AI search results about candidates can be trusted.
| Question | AI tool | AI answered | Searches |
|---|
Most citations go to sites that Media Bias/Fact Check does not rate. Among the — it does rate, left and left-center sources outnumber right and right-center ones on every tool.
The news supply itself leans left on these ratings, and regular search results have not yet been rated for comparison, so this does not show that the AI tools choose left-leaning sources.
Types are assigned from each site's address and display name. Sites were not visited.