We’ve been measuring voter AI use, as well as what chatbots are telling voters, and we agree that AI chatbots will be a real factor this fall – for the first time in an American general election. The NYT’s interviews are data points in a larger overall pattern we’ve built Caucus AI to uncover. In today’s post, we review what we’ve learned so far and how it dovetails with what the Times found.

Subscribe now

Who is using chatbots

According to a survey we conducted with Change Research in May, a significant minority (15%) of voters say they are very or somewhat likely to consult an AI chatbot for information on midterm candidates. A significantly larger share, 62%, say they are likely to use Google Search – and so may be exposed to AI-generated political information via Google’s AI search overviews.

Two of the most persuadable groups of voters in recent elections are more likely to turn to chatbots. 21% of voters under age 35 say they may use chatbots, compared to just 8% of voters over 65. Voters of color are nearly twice as likely to say they’ll use AI chatbots (21%), compared to 12% of white voters. This represents a small share of the electorate, but often a decisive one.

It isn’t just persuadable top-of-ticket voters who are turning to chatbots though – more engaged voters participating in primaries or seeking information on less prominent, downballot races are also consulting these tools, as the NY Times article showed in the California primaries. In our survey, respondents who were decided on the generic Congressional ballot were actually more likely to say they would use a chatbot (15% of Democrats and 17% of Republicans), compared to just 9% of undecided voters. This is consistent with the fact that stronger partisans are more likely to seek information across all channels and in 2026, that includes AI chatbots.

What voters are asking

The voters the Times profiled also asked questions that track with our national data – strategic, subjective, and often personal questions. In our Change Research survey we asked respondents for the actual questions they’d ask a chatbot. Of substantive responses, 39% asked questions requiring chatbots to make judgment calls or synthesize across multiple sources (e.g. “Which candidate is more likely to tell the truth and practice what they preach?”). More than half of responses from voters under 35 were these types of questions. 18% of substantive responses overall were asks for the chatbot to compare and contrast candidates or recommend who to vote for.

What chatbots are saying

Since February, we’ve been monitoring what information voters get back from chatbots when they ask about politics. We’ve been querying ChatGPT, Gemini, and Grok with identical prompts about 2026 midterm candidates at the House, Senate, and gubernatorial levels. This allows us a bird’s eye view of what chatbots are saying about the 2026 midterms. Findings so far include –

• Neutrality is the goal: As the New York Times also found, frontier labs creating these chatbots are striving for neutrality – in our data so far, we have found that models are clearly orienting around this goal. Models present generally positive views on bio and issues, unless asked specifically about criticisms of a candidate.

• Sourcing: Our core finding is that the sources chatbots draw on vary widely from model to model, and even within model runs. Models appear to heavily favor certain publishers–many of which have content partnerships with the labs–and can ignore otherwise prominent outlets and sites. This means that when a user selects a model, they are unknowingly selecting into a specific information ecosystem.

• Answers update quickly: Because most frontier AI chatbots use web search to respond to questions about elections, it means they have the ability to incorporate new content as soon as it goes live online. This is helpful with breaking news – while monitoring California’s governor’s race, we found that chatbots picked up information about sexual misconduct allegations against Eric Swalwell the same day news stories broke. On the other hand, less well-vetted user-generated content may also be picked up quickly – in March, we created two Wikipedia pages for candidates we thought met Wikipedia’s notability regulations, but didn’t yet have pages. This content was picked up by ChatGPT in under 15 minutes.

• Incumbency and prominence matter: As Yamil Velez pointed out in the article, chatbots favor candidates with bigger media footprints. We quantified this with a similarity exercise: comparing chatbot responses to candidate websites (using the website as a proxy for the message the candidate is trying to deliver to voters). We found that chatbot responses were more similar to candidates’ websites for incumbents and more prominent candidates. Challengers and less prominent candidates have a harder time getting their message across via chatbot.

What we still don’t know

All of this is dynamic – frontier AI labs are developing policy and partnerships in real time, chatbots increasingly rely on memory of users’ preferences, chatbot use and voter attitudes toward AI in general are rapidly changing, and campaigns themselves are exploring ways to adapt to an increasingly AI-generated information ecosystem.

We will be expanding our work with Caucus AI to explore these open questions and monitor change over time as November approaches. If you have ideas for what we should look into next or want to dig into what chatbots are saying about a specific race, please get in touch!

The findings below reflect Change Research’s survey of 1,892 respondents weighted to a registered voter population. The survey was fielded from May 5-10, 2026. Respondents were recruited using a combination of targeted digital ads and text-to-web SMS, and all respondents took the survey online.

Subscribe now

What they’ll ask

When voters do go to a chatbot for political information, what do they actually ask it? This information is critical for Caucus AI, because the response information we are gathering is only helpful if our prompts reflect what real voters actually ask. We collected actual questions respondents said they would pose to a chatbot to learn about candidates for office in 2026.

Of substantive responses (see below for refusals!), nearly half (49%) were asking about candidate records and issue positions. 18% were requests for the chatbot to compare and contrast candidates or recommend who to vote for. Questions about who is running and candidate biography were 9% and 8% of requests respectively. All of this information is covered by existing prompts in the Caucus AI tool – although we do not (yet) ask chatbots directly to compare and contrast candidates – we prompt for issues, bio, and criticisms of specific candidates, and chatbots often return comparative responses.

Interestingly, 9% of questions were requests for information on candidates representing particular ideologies or parties (or who had been endorsed by specific people, mostly Donald Trump). This partisan information-seeking is interesting, because it prompts the chatbot directly to focus on particular types of candidates rather than returning the full picture. Other research has indicated that chatbots respond differently to users when prompted to think they are Democrats or Republicans, so this is an area to dig into further in future work.

Is AI just regular-old search?

Do chatbot users use AI search any differently than they already use search engines? While we don’t have a comparison dataset of old-fashioned search queries, we are able to get a sense of what share of queries to chatbots require synthesis to answer versus the share that could be answered by clicking on a single result from a search engine. We define synthesis questions as queries that require the AI chatbot to summarize or compare across multiple sources, candidates, or questions or to subjectively reason about a question. Example synthesis queries in our dataset include “Which candidate is more likely to tell the truth and practice what they preach?” and “Who takes corporate money? Who takes aipac money? Who supports Palestine?” More straightforward search queries include “How does candidate stand on the issue of lowering taxes” and “Who is the Republican?” These questions can be answered by chatbot synthesis, but are straightforward to answer via regular search engine results as well.

Overall, 39% of respondents provided a synthesis prompt compared to 16% who asked a more standard search question (45% of all respondents refused to provide a prompt). Younger voters are much more likely to ask synthesis questions (56%) compared to voters over age 65 (32%). Part of this is driven by younger voters’ lower refusal rates – older respondents were much more likely to refuse to provide a prompt (48% compared to 33% for voters under 35) – but some of the difference is also driven by older voters being more likely to ask search-type questions.

AI Backlash

Refusals in the survey also paint an interesting picture of how Americans are feeling about AI in politics. Almost half (45%) of respondents refused to provide an example of what they would ask a chatbot to learn about candidates. Of respondents who refused, just over half specifically said they would not use AI to learn about politics, while 28% were generic “NA” answers and 20% were off-topic responses. Many responses expressed real hostility toward AI (such as “Fuck AI”, which four respondents wrote verbatim), reflecting what we see throughout the survey and in other research: many Americans do not like AI and do not want to use it at all. Polling from Searchlight Institute in late 2025 found Americans had mixed views on the impact of AI overall, support regulation, and view AI significantly less favorably than other technology. A March NBC News poll found that 57% of registered voters think the risks of AI outweigh the benefits, compared to 34% who said the benefits outweighed the risks. 71% of Americans in a May 2026 YouGov / Economist poll said that AI is moving too fast. Nowhere has public opinion soured more related to AI than on support for building data centers, which dropped from 65% in March 2025 to 36% in April 2026, according to Change Research polling.

Most respondents who said they wouldn’t use AI did not give a specific reason (72%). Of those who did, about half cited that they did not trust the accuracy of chatbot information for politics or thought it would be biased, 22% expressed general hostility / dislike toward AI, and a handful expressed concern over societal or environmental harms of AI or said they generally prefer to rely on themselves, not chatbots, to learn about politics.

Implications for the Caucus AI tool and future work

This deeper dive into what people will actually ask chatbots about politics tells us two things. First, Caucus AI largely covers the most common candidate- and district-specific questions: who is running, candidate bio, issue positions, and criticisms. What Caucus AI fails to capture right now is the widespread comparative nature of many user queries – voters will be asking chatbots directly who they should vote for, who is better on specific issues that they care about, and who has been endorsed by various parties and organizations. We also do not capture the variety of ways questions can be asked and how the wording of prompts may impact chatbot responses. These are gaps that we will work to address in future releases for the Caucus AI tool.

Second, we see a clear divide emerging in the electorate on the use of AI chatbots to learn about politics. We laid out in our last post that a small minority of voters (15%) plan to use AI this fall to learn about candidates, and these voters are also the ones most likely to ask demanding synthesis questions of chatbots – relying on them to summarize over multiple sources or make judgment calls at higher rates than older voters. On the other hand, close to half of respondents refused to offer a question they might ask a chatbot and many expressed real hostility toward AI, which mirrors the general anxiety and decline in favorability of AI that we see across public opinion polling.

This work represents a foundation to build off of, as Americans continue to adapt, adopt, or reject AI and as AI capabilities continue to develop.

The findings below reflect Change Research’s survey of 1,892 respondents weighted to a registered voter population. The survey was fielded from May 5-10, 2026. Respondents were recruited using a combination of targeted digital ads and text-to-web SMS, and all respondents took the survey online.

AI as a source of political information

Asked which information sources they were likely to use to learn about candidates running for office in 2026, most Americans say they will rely on news articles (68% very or somewhat likely to use), discussions with friends, family, or colleagues (62%), candidate websites (57%), and social media (53%).

A much smaller percentage say they plan to use an AI chatbot – just 15% of respondents said they were very or somewhat likely to use an AI chatbot to learn about candidates and a majority (58%) actually reported that they were very unlikely to do so. However, 62% said they would rely on Google Search to learn about candidates. Many of these searchers may be exposed to AI-generated political information via the AI summaries Google displays on top of search results, although it is unclear what share of political Google Search queries will have AI overviews.

These results highlight two key points: 1) although only a small slice of the electorate will actively seek out AI-generated information about this year’s elections, even a small segment of active AI information-seekers is large enough to be decisive in close elections and 2) a huge proportion of the electorate may encounter AI-generated political information through passive consumption.

Who is most likely to turn to a chatbot?

Voters under age 35 are nearly three times more likely to say they will actively seek information from AI chatbots (21%) than voters over age 65 (8%). This is still far lower than the share of voters under 35 who say they will rely on social media (69%), but means that one fifth of young voters plan to rely on chatbots as an information source as they make voting decisions in November.

Voters of color are much more likely to say they will use chatbots for political information compared to white respondents (21% vs. 12%) and men are slightly more likely than women to do so (17% vs. 13%).

Broken out by partisanship, there are modest partisan differences on intent to use AI, but larger differences among those who say they are very unlikely to use AI – 60% of Democrats, 54% of Republicans, and 68% of Independents say they are very unlikely to use AI to learn about candidates this year.

Respondents who were undecided on the generic Congressional ballot were less likely to say they would use a chatbot (9% very or somewhat likely) than those voting for Democrats (15%) or Republicans (17%). These undecided respondents tend to be a hard-to-reach population and they were significantly less likely to say they would seek out information on all of the channels asked about above than voters already supporting Democrats and Republicans.

How much do voters trust AI-generated information about politics?

Next, we wanted to assess whether voters view AI-generated political information as more or less accurate than information from candidates themselves. We know that Americans are generally pessimistic about the accuracy of AI and AI’s impact on society. Americans are also pessimistic about politics and have low trust in elected officials. Between candidate- and AI-provided information, which source wins out? Candidate information wins out as more trustworthy (36% to 14%).

When asked, “When it comes to information about political candidates, which do you find more accurate, answers from AI tools (e.g., ChatGPT, AI search summaries) or information from the candidates themselves?”, more than a third of respondents (36%) said that neither are accurate. 14% said that AI-generated information is somewhat or much more accurate, while 36% said candidate-generated information is more accurate. 14% said AI- and candidate-generated information are equally accurate.

By subgroup, trust mirrors likelihood to use AI, with 31% of respondents under 35 saying AI-generated information is equally accurate to or more accurate than candidate information. 37% of voters of color (including nearly half of Black voters) said that AI is equally or more accurate.

An outright majority of Independents (56%) responded that neither source was accurate.

Implications for 2026 and beyond

A significant minority of voters (15%) will actively seek out information from AI chatbots when they go to the polls this Fall. Even more may be exposed to AI-generated election information via search. These voters are disproportionately young and Black, Hispanic, and Asian – some of the exact demographics that have been the among the most persuadable in recent elections. And while most voters distrust AI, the same groups using it are more likely to trust it than others. In close races, AI-generated information could be decisive.

AI is also here to stay. Americans increasingly use generative AI across their work and personal lives, with the pace of adoption tracking ahead of previous transformative technologies such as the personal computer and the internet. In 1996, just 4% of voters sought political information online – rising to 18% in 2000 and 29% in 2004. More than 20 years later, many Americans live their lives online. This fall, more than half of voters say they plan to seek out political information via Google Search and candidate websites (not to mention news articles that are largely online).

We are in the very early stages of observing what impact AI-generated information will have on American politics (and wider society). Understanding what chatbots tell voters is already critical and as the share of voters consulting AI sources grows, it will only become more important.

Information on issues and criticisms are now live in Caucus AI. We know what chatbots say about who is running and who candidates are – now we can also explore what chatbots say in response to user prompts “Where does [candidate] stand on important issues?” and “What are the main criticisms of [candidate]?” These additions were among the most-requested features from early users of Caucus AI – the post below digs into what we’ve found so far.

Subscribe now

The Setup

We ask three models (Gemini, GPT, and Grok) multiple questions about 2026 political candidates. We store every response, extract the sources each model cites, and compare and contrast answers across model providers. With each query, we allow the model to use web search tools, to better mimic how AI chatbots and AI-enhanced search behave for real users. We run each prompt twice, to capture some of the random variation in chatbot responses.

The analysis below focuses on Senate, Governor, and House races that are rated by Cook Political Report as competitive (toss-up, lean, or likely) – representing competitive races that have a variety of online coverage.

Wikipedia owns bio, campaigns own issues, and nobody owns criticisms

There is striking variation in what websites the chatbots cite. Our initial analyses focused on bio-style questions (“Tell me about Candidate X”) and we found that the chatbots cited reference websites like Wikipedia, Encyclopedia Britannica, and Ballotpedia heavily. We can see that this does not hold for Issues and Criticisms responses.

For Issues responses, the chatbots rely about equally on campaign websites, reference sites, and local news. Relative to other sources, candidates have the most say in shaping what chatbots tell users here, since they control their own campaign websites.

For criticisms, we see a heavy reliance on local news with about a third of references from local news and a further 17.3% from national news. Social media also rises in importance – this is driven by Grok, but is true for Gemini and GPT as well. Partisan sites are nearly twice as frequently cited for criticisms as for bio. Reference, campaign, and official websites – the anchors of biographical answers – represent just 12.5% of all Criticism citations.

Looking more closely at specific websites cited by models for criticisms, we can see the prominence of social media—YouTube dominates Gemini citations for criticisms of Democratic candidates, Reddit is prominent throughout—and of partisan sources, like the Democratic and Republican House campaign committee websites (DCCC and NRCC). GPT consistently and heavily relies on Wikipedia in a way the other models do not: Wikipedia is GPT’s single most-cited source for criticisms of both parties.

On the issues

When evaluating model responses, we classified each response segment by topic (see the methodology section below for more). On average, each model covered about 4 topics per issue response. Issues topics break down fairly cleanly by the topics each party typically emphasizes, and the framing they tend to employ. Democrats’ issue responses talk disproportionately about healthcare, the economy and jobs, and reproductive rights. Republicans’ issue responses spend more time on taxes and budget issues, immigration and border security, and crime.

When we look at the economy and costs, the core issue in 2026, we can see that responses for both parties lean into that party’s framing. Model responses for Democrats emphasize tariffs, taxes, jobs and workers, while Republican responses emphasize cutting spending and taxes. Costs, wages, housing all pop for Democrats, while inflation, small businesses, and “cuts” are more prominent for Republicans.

Criticisms

When it comes to criticisms, Gemini and GPT are most likely to cite policy positions as the main area of critique, while Grok prefers scandals (ethics, corruption, and personal scandal topics). Across all three models, Republicans are more likely to be criticized for their policy positions, reflecting the 2026 climate and the GOP’s disproportionate incumbent status.

GPT is more likely than the other models to frame Democrats as ideologically extreme – both for being too far left, as well as for being too moderate. In other words, GPT picks up on criticisms of centrist Democrats from the left.

All models highlight Republican candidates’ histories of election denial, and Democrats are two to three times more likely to have some criticism for being out of touch with voters.

I’m sorry, Dave. I’m afraid I can’t do that

We found that for 21% of criticisms responses, GPT did not return a substantive answer. Most frequently, the non-substantive answer was a request for clarification or saying it did not know any person by that name. GPT had by far the highest rate of non-substantive answering, Gemini was at 2% and Grok at 4%. Non-substantive responses for all models appear disproportionately for House challengers – candidates with smaller digital footprints than higher-profile races or established incumbents.

Interestingly, 41% of the queries that had no GPT response to a request for criticisms, did have a substantive response for the candidate bio prompt. In other words, when asked “Tell me about [candidate]”, GPT was comfortable proceeding, but with “what are the main criticisms of [candidate]?”, the model was much more likely to ask for clarification or say it did not know of the candidate.

The big picture

Three things stand out so far from these new Issues and Criticisms additions.

The sources chatbots draw on shift substantially based on what the user asks: encyclopedia-style reference sites dominate bio responses, while criticisms answers lean on social media, local news, and partisan sources.

Issues largely map onto familiar partisan territory: Responses for Democrats focus on healthcare, jobs, and the economy, while Republican responses focus on taxes, government spending, and immigration.

Consequentially for the user, GPT declines to answer “what are the main criticisms of [candidate]?” about one in five times. This behavior can determine whether a voter using GPT gets a critique of a candidate at all.

As we work to expand the tool, we’ll keep tracking how these patterns evolve through the cycle and what might be causing them. Issues and criticisms data is now live in Caucus AI – explore it for any candidate in a race we cover here.

Methodology

• Prompts: For each candidate we asked “Where does [candidate] stand on important issues?” and “What are the main criticisms of [candidate]?” Each model ran with web search enabled and ran each prompt twice to capture some of the randomness in chatbot answers.

• Classifying topics: we developed a fixed taxonomy of 19 topics for Issues responses and 17 for criticisms, developed based on a random sample of responses and human oversight. We use an LLM (Gemini Flash) as a classifier. The classifying model sees the full response and returns the topic of each sentence in the response. The unit of classification is the sentence, and response-level topics are aggregated from these sentences. The classifier does not know which model produced the response it is categorizing.

• Share of voice: Model responses usually include multiple topics and are not just about one issue or criticism. For example, a Democratic candidate’s issues answer might cover lowering costs, healthcare, climate, etc. all in one response. When we report how much a response was “about” a given topic, we use length and position weights. Topics that take up more of the total chatbot answer count for more, and where a topic is discussed matters too – topics discussed earlier in an answer count for more than later topics, since human readers tend to skim AI responses. Using the single most dominant topic or varying how we calculate salience produces largely similar findings to those presented above.

In our latest analysis using Caucus AI, we ask: How stable are standard AI chatbot responses to questions about political candidates and elections? Do citations actually influence response content?

Users of generative AI tools know that you rarely get the exact same response to a question if you ask it multiple times. This is largely due to two factors:

1. Temperature, or purposeful randomness setting (some newer models have a setting called “effort” instead, which governs randomness slightly differently), in the model itself that helps models produce more “novel” outputs.

2. Implementation particulars of large AI models that create variation in responses even when the temperature is set to be non-random.

Setup

Response variation

We took a sample of prompts for different geographies and candidates and ran each prompt 15 times per model at the model’s default temperature/effort level. This allowed us to understand variation in responses.

We measured similarity of responses across 3 metrics:

• Semantic similarity: Does a model say the same thing each time?

• Keyword overlap: Do models use the same words each time?

• Citation overlap and churn: Do responses cite the same source websites?

Citation-content correlation

We scraped 994 candidates’ campaign and official (.gov) websites. We compared the semantic similarity of chatbot responses to candidates’ webpages on the following dimensions:

• Whether the candidate’s campaign and/or official webpage was cited by the chatbot

• Candidate features: incumbency, partisanship, office level (House, Senate, Governor), and competitiveness

• Site features: site type (campaign, official), page type (bio, issues, news, etc.)

Findings

Response content is quite stable.

Overall, models return largely the same meaning even if they use different wording, structure, or formatting.

The cosine similarity of model responses to identical prompts is .94–.95 on average (out of a maximum of 1) across all three models, while metrics of overlap of specific vocabulary are lower.

Responses to more open-ended questions (“What does Jon Ossoff think about important issues in Georgia?”) had more variation than more fact-based questions (“Who is running for Senate in Georgia?”), as did responses to questions about lower-profile candidates and races.

Citations vary widely

The complete list of citations varies significantly from run to run, much more than the information conveyed in the responses.

We identified that each prompt-model combination had, on average:

• 2.4 core sources that were cited by more than 10 of the 15 responses

• 3.8 frequent sources cited 4–10 times

• 12.9 rare citations cited less than 4 times across all runs

Citations matter

Models are drawing from candidate websites. Average cosine similarity between chatbot responses and candidates’ own site content is 0.67, with a median of 0.71. Two-thirds of the closest content matches are biographical pages (career history, background, personal details) rather than policy or issues pages (just 2.2% of best matches). This will likely change as we expand out the set of questions we’re asking the chatbots – right now, our questions focus on candidate backgrounds and biography

Citing a site predicts meaningfully higher similarity. When a response explicitly cites a candidate’s website, it is significantly more similar to that site’s content than when it doesn’t. The effect is large: +12pp for campaign sites and +27pp for official (.gov) sites. This holds across all three models. GPT cites candidate sites least often (35% of responses) but shows the strongest content alignment when it does; Grok cites most often (68%) but with a smaller similarity gap, suggesting it references sites more liberally.

Incumbents score much higher than challengers (0.703 vs 0.629), the largest effect we observed. Incumbents have more web presence, more media coverage, and longer-standing official sites, all of which feed into training data. Challengers, especially lesser-known ones, have responses that diverge more significantly from their own website content.

Example: Josh Shapiro

As an example, we can look at the full set of citations for Pennsylvania Governor Josh Shapiro in response to the prompt “Tell me about Josh Shapiro.”

The chart below shows websites cited by each model. Several patterns emerge:

• All models cite Wikipedia frequently. It appears in every response from GPT 5.2 and Grok 4.1.

• Gemini and Grok cite PA’s NPR affiliate, WHYY, while GPT 5.2 does not.

• GPT 5.2 always cites a source from Axios, perhaps driven by OpenAI’s partnership with Axios. The model also cites partner outlets such as the AP and The Guardian.

• pa.gov is cited frequently, but Shapiro’s 2026 campaign website (https://joshshapiro.org/) is not cited at all.

What does this mean?

Models generally say the same things in response to factual (“Who is running?”) and biographical questions. However, there is significant variation in the sources models cite for each response. The variation occurs both within models (particularly Gemini) and across models, with some models like GPT appearing to heavily favor certain sources. And, critically, citations seem to matter for content – responses are more similar to the sources they cite.

There are still a lot of open questions. Coming soon, we hope to tackle research like:

• Are citations causal? We have established that candidate website citations correlate with response content, but it could be that chatbots may select citations in part based on what their response will be. There is much more work to do to understand how web content changes filter through chatbot responses.

• How does this change over time? As new models come out, news outlets adapt to AI-enhanced search, and candidates release new web content, we don’t know how consistent citation patterns may shift over time.

The source gap

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.

Where brands break through

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.

Candidate disambiguation

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.

Campaign site citations

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?

What’s next

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.