From U.S. Elections to Taipei’s Debate What Shapes AI’s View
of the Candidates?

01 · Sources behind the answers

AI's source rankings do not mirror online attention

Across the three waves, CNA accounted for 26.2% of all ChatGPT citations, exactly four times the count for second-place TVBS. United Daily News led Gemini's media ranking with 7.5% of all citations, with a narrower gap over the outlets that followed.

Only four outlets appeared in both models' media top ten. Each model's leading outlet stayed the same across all three waves. Even when covering the same election, the two models regularly drew on different sources.

News volume and AI citations tell different stories

The difference becomes clearer when we compare the rankings with the wider online picture. In KEYPO's news-volume ranking, which attributes articles to their original outlet, CNA ranked 13th both before and after the debate, yet remained ChatGPT's most-cited source in all three waves. The same outlet held very different positions in news volume and AI answers.

We saw a similar pattern in social discussion. On the day of the debate, Threads accounted for 36.2% of the posts and articles observed by KEYPO, making it the largest source platform. Yet the AI source lists contained no direct Threads URLs in any of the three waves. The busiest places for discussion do not necessarily appear as direct links in AI answers.

This adds another way to examine media influence. Alongside the volume and reach of coverage, we can track which content AI repeatedly cites when helping readers understand people and issues.

Profiles and research can also lead the rankings

The citation lists extend beyond news. After excluding media outlets and news platforms, Taipei City Government ranked first for both models, accounting for 12.3% of all ChatGPT citations and 5.8% of all Gemini citations. Official websites, encyclopedias and forums all form part of this information mix.

Looking at individual pages makes this easier to understand. For ChatGPT answers to questions about Chiang Wan-an, the most-cited page was QSearch's analysis of social discussion and polling. It appeared in 10.7% of answers. For questions about Puma Shen, his Legislative Yuan profile led the ranking, appearing in 21.5% of answers. Neither leading page was a news report.

Gemini's candidate-related page rankings showed a different mix. CommonWealth's debate analysis led the pages related to Chiang Wan-an, while Chinese Wikipedia led those related to Puma Shen. Commentary on a current event and established biographical information both appeared among AI's citations.

The rankings also give us a reason to look more closely. When AI cites an outlet, is it using an interview, a policy explanation or coverage of a controversy? A citation can frame a person or issue in very different ways.

02 | How AI compares candidates

The sources matter. So does the ability you ask about.

Sources show us what material AI draws on. To understand how it compares candidates, we also need to read the question. Asking whether Chiang Wan-an or Puma Shen is better suited to be mayor produces a different assessment from asking who communicates better.

For the question about suitability to be mayor, some ChatGPT answers favored Chiang Wan-an because they prioritized administrative experience and continuity in government. For communication skills, such as answering directly and handling follow-up questions, 80% of pre-debate answers were more positive about Puma Shen. Each assessment came with conditions and referred to a different ability.

After the debate, more answers to the communication question offered a two-sided comparison: 53% in the first post-debate wave, rising to 60% in the follow-up wave. Some answers praised Shen's quick responses while also crediting Chiang's explanations of policy context. The tone used to describe the two candidates also became more similar.

Gemini gave a two-sided comparison for both questions across all three waves. To understand an AI assessment, it helps to ask one more question: Which ability does it mean by "suitable"? And what conditions does it attach when it speaks positively about someone?

03 | Which information makes it into answers

When new information appears, which issues stay in the answers?

Different comparison criteria produce different assessments. Over time, the issues discussed in the answers also change. Beyond the source rankings, we repeatedly asked, “What are the two candidates mainly arguing about?” to see whether debate topics appeared in the answers.

KEYPO recorded debate exchanges on child abuse, food safety and city government meetings. In Gemini answers, mentions of child abuse rose from 7% before the debate to 87% in the first post-debate wave, and stayed at 87% in the follow-up wave. Meanwhile, mentions of previously frequent "plagiarism"-related disputes fell from 73% to 0% in both later waves.

As some issues appeared more often and others less often, the emphasis of the answers shifted. Food safety and school lunches appeared much more often in Gemini answers and continued to appear frequently in the follow-up wave. ChatGPT was already discussing these issues frequently before the debate, so the models had different starting points.

In this short-term comparison, some debate topics appeared in AI answers and remained there in the next wave. To understand that change, we need to track more than which topics are gaining attention online. We also need to see how the answers describe them, and whether they still appear the next time we ask.