The Source Gap: Early Findings on How AI Chatbots Inform Voters

There is a trove of data to dig into with our chatbot tracker, and we will release more in-depth research findings in the coming weeks (have ideas? Drop us a note!). Below, we highlight a few initial patterns and findings that we’ve found interesting from the tool so far:

One of the most interesting differences across the three models1 we’ve tested is the sheer variation in web domains each model cites. In particular, while all 3 models rely heavily on Wikipedia, each model also appears to have a “home base” of sources it trusts, with sometimes little overlap between models.

• GPT 5.2 disproportionately cites the Associated Press, Yahoo, Axios, the Washington Post, and the NY Post.

• Grok cites social media. In our dataset so far, it is the only model to cite X, Instagram, Facebook, and LinkedIn regularly.

• Gemini cites more including Grokipedia: Gemini responses return significantly more sources per response (7.2 vs. GPT’s 4.4) and citations vary more across prompt repetitions. Gemini is also responsible for nearly all references to Grokipedia, Elon Musk’s “anti-woke” alternative to Wikipedia.

• EMILY’s List candidate bio pages are the most commonly-cited explicitly partisan source, other than candidate websites, although they make up less than 1% of all sources cited.

We have several great examples of candidates’ own brands breaking through, although there can be variation across models.

• All three model responses for Mary Peltola feature her “Fish, Family, Freedom” slogan prominently – the clearest example of a candidate’s brand breaking through in LLM results that we have found.

• Kelly Ayotte’s “Don’t Mass up New Hampshire” slogan and Josh Shapiro’s “Get S— Done” (censoring courtesy of Gemini) also pop – showing that prominent brands surface in AI results even when they cite disparate sources.

Some candidates share names with other prominent people – the models all handle this differently, sometimes returning bios for several people at once. GPT 5.2 is particularly reluctant to guess at what the user wants, especially for lesser-known candidates – roughly 25% of GPT’s House responses are requests for name disambiguation.

There is more work to be done here to understand if users querying while physically located in a given district have the same disambiguation issues or if location matters for response quality. Regardless, users are able to easily clarify who they mean in actual interactions.

Models vary in how often they cite a candidate’s own web presence. Grok cited either a campaign or official website in 89% of candidate responses, Gemini did so 65% of the time, and ChatGPT did so less than half the time. ChatGPT cited campaign websites just 22% of the time, and only 13% of the time for U.S. Senate candidates. All three models were more likely to cite official government sites than campaign sites, and more likely to cite campaign sites for lesser-known House candidates than for well-known senators. For more prominent candidates, models may have more third-party sources at their disposal, or they may “know” a candidate from their training data, and skip the candidate sites themselves.

In future research, we are interested in exploring more deeply here – what makes a campaign site more or less likely to be surfaced by AI chatbots? Are more prominent candidates or established legislators less able to influence search results due to the sheer volume of other content about them on the internet?

In the coming weeks and months, we’ll be going deeper on these topics, as well as longitudinal analysis as we collect more data over time. Please reach out if you have ideas, we’re open to collaborating.