No one has produced a rigorous causal estimate. This paper explains why that is not a reason to wait, and what the adoption data already tells us about the scale of the problem.
Every election cycle generates a fresh set of benchmarks: television reach, digital ad impressions, social media engagement. Campaigns need to know where voters are paying attention, and a whole market of researchers, pollsters, and media buyers exists to tell them.
Prospective Kyanos clients ask what AI says about their candidate mostly out of curiosity. They need to be prompted to ask how to fix it. And there is a lot of confusion underneath the question. People think you need to open ChatGPT or Claude or Grok to get an AI answer. They do not realize that google.com is now AI-synthesized answers by default, not the list of blue links we had for twenty years, most of them advertiser supported. The list of links as a response is largely gone. Google snuck AI in without making a big deal of it. Tell a prospect this and you get a sort of surprise. Just about everyone gets AI answers now. Most people do not know it.
Just because we do not have precise measurements, it does not mean that managing controlled surfaces, doing everything we can to help AI chatbots get the framing right, is futile. It is common sense that some voters go to AI for answers. The ROI holds up even with the most pessimistic assumptions.
The absence of a benchmark is not evidence that the channel is small. It is evidence that the measurement infrastructure has not been built.
Pitch AI monitoring to a campaign manager and the likely response is that AI chatbots just aren't on the list of things to do. Good luck. Prove it works, show me how it really influenced an election. It is a fair response. Campaigns run on too little time and money, and a season behind is where the incentives put everyone; podcasts looked skippable too, until male voters under 45 swung thirteen points between 2020 and 2024 while one campaign sat for the podcast interviews and the other mostly did not.2 The DNC's own 2024 post-mortem put numbers on that miss: the other ticket simply showed up in the new media far more often, while the party argued with its own diagnosis.3
The history of campaign media is the history of that mistake being punished. In 1952 a presidential campaign spent $1.5 million on television spots when a third of American households had a set and no one could measure what an ad did to a vote.4 When Google began selling search ads, buying your own candidate's name put your message at the top of the page for a fraction of what broadcast cost, and the campaigns that bought early owned the first thing voters saw. Obama's 2008 campaign organized on social media while opponents debated whether it was serious. In 2016 the Trump operation put nearly half its outside money into digital while the Clinton side put over seventy percent of the race's television dollars on the air.5 None of these movers had definitive proof. Eisenhower, Obama, and Trump won those elections.
Meanwhile the proof campaigns demand never quite arrives, even for the channels they trust. Seventy years after Eisenhower's first spots, the best modern measurements of television's persuasive effect find it real but small: fifty-nine randomized experiments on thirty-four thousand people found small average effects on vote choice, and the cleanest natural experiment puts a full standard-deviation advertising advantage at about half a point of vote share.6 What the research does suggest is that negative framing, repeated until it is inescapable, has an effect.7 Which is the part that should worry anyone running for office: negative TV ads take facts and frame them in the most negative way possible, and chatbots do the same thing if left to their own devices.
Campaigns were newspapers, radio, and pressing the flesh, and then they were television, decades ahead of any proof. The measurement followed the behavior. It is following it again.
There is one way this channel is different, and it makes the timing harder, not easier. Money pours into campaigns close to election day if the race is close, and late money goes where it can be spent fast: advertising is the easiest thing to pour buckets into, and the measured effects of ads decay within days of airing, which is why late saturation buys so little.8 AI presence management is the opposite shape. Its impact is cumulative. It has to start as long before election day as possible, and it cannot be bought in late October with a big check and a hope. The channels campaigns know how to fund are the ones that absorb money badly at the end. The channel they have not learned to fund is the one that only works if you start early.
To understand what AI does to voter information, it helps to understand what search already did, because AI did not arrive in a vacuum. It arrived into an information environment that web search had been reshaping for twenty years.
In 2000, fewer than half of American adults used the internet regularly, and online political activity was limited to a small early-adopter segment. By 2004, that had changed. Pew Research Center surveys from that cycle documented that roughly a quarter of American adults reported going online to get news and information about the campaigns, a figure that had roughly doubled from 2000. The Howard Dean campaign's online fundraising breakthrough in the Democratic primary that year was the first widely noticed signal that the internet was becoming a structural feature of campaigns, not a novelty.
By 2008, the shift was decisive. Obama's campaign used the web not just for fundraising but for organizing, volunteer coordination, and earned media amplification. Post-election surveys found that the share of Americans getting campaign news online had grown substantially, and for younger voters, the web had become the primary information source. The television-first model of campaign communication was still dominant, but it was no longer uncontested.
The 2012 and 2016 cycles extended the pattern. Social media accelerated information circulation. Search became the default starting point for voter due diligence: background checks on candidates, issue research, fact-checking during debates. By 2016, surveys consistently found that a majority of Americans were turning to the internet first when they wanted to learn more about a candidate or a race.
As search became a primary voter research channel, a parallel infrastructure grew up around it. Campaigns hired digital directors. SEO firms developed political practices. Ballotpedia and Wikipedia became understood as critical assets. Campaigns could audit what voters found when they searched, study which sources ranked prominently, and invest in the content that shaped those results.
The key characteristic of this environment: it was legible. A campaign could see what Google returned for a candidate's name. The information environment was navigable because the results were visible, to voters, and to the campaigns managing their presence in them.
By 2020, voter reliance on search for candidate research was deeply embedded. Not universal: television, direct mail, and social media still reached voters who did not seek out information actively. But for the persuadable, research-oriented voter who is disproportionately influential in competitive races, web search was the default tool for forming an initial impression of a candidate before any campaign communication reached them. Campaigns had adapted to that reality over twenty years of iteration.
That is the baseline. The shift described in this paper is not happening in a pre-digital information environment. It is happening on top of a search infrastructure that campaigns already understand and depend on, which is precisely what makes it difficult to see clearly. The platform that campaigns learned to manage is now delivering different answers than it used to. And the tools campaigns built for the old version of that platform do not fully work on the new one.
Every previous technology shift in political communication required voters to change their behavior. To get political news on social media, voters had to join social media. To watch a campaign's YouTube ads, voters had to be on YouTube. The new channel had to earn adoption: it had to compete for attention against established habits and convince voters to do something differently.
The AI shift is different in a way that has no precedent in political communication history. Voters did not have to change anything they were already doing.
The URL bar has not changed. The habit has not changed. The search box that American voters have been typing into for the better part of twenty-five years, google.com, accessible from every browser, built into the address bar of every phone, is the same address it has always been. What changed is what comes back.
Jane Smith is a Democratic candidate for State Senate in District 14. She served on the Riverside school board from 2014 to 2018, during which the district faced a budget controversy involving a bond measure she supported. She has focused her campaign on education funding and local economic development. In 2016, she was involved in a public dispute with the teachers union over contract negotiations.
The voter in both panels typed the same thing into the same address bar. In 2018, they got a ranked list of sources and made their own judgments. In 2025, they got an answer, assembled from sources the AI selected, weighted by criteria not publicly documented, delivered without a byline. The campaign's official website is still there, below the fold. But for a substantial share of voters, the impression was already formed before they scrolled to it.
This is why adoption metrics like the Edison 65% figure understate the actual exposure. That figure counts Americans who consciously used a standalone AI chat platform. It does not count the much larger number who used Google.com and received an AI Overview, which requires no change in behavior, no new account, no deliberate choice to use AI. The answer simply appeared where the blue links used to be.
There is a second reason to expect this exposure to grow regardless of any organic shift in voter preference: the companies building these systems have strong commercial incentives to expand usage. OpenAI, Google, Meta, and Microsoft have each made multi-billion-dollar bets on AI as the next dominant computing platform. Usage growth is a primary metric they report to investors. These companies are actively integrating AI into the tools people already use, browsers, phones, operating systems, productivity software, in ways that require no deliberate adoption decision. The voter does not need to choose AI. AI is increasingly the default layer on top of tools they already use. As usage expands across everyday queries, political queries follow as a matter of course. The growth trajectory is not contingent on voters deciding to change their behavior. The platforms are making that decision for them.
Absent a direct measure of chatbot influence on voters, adoption data is the closest available proxy, and the adoption data does not point in an uncertain direction.
Weekly usage: Edison Research (May 2026) found that 65% of Americans used an AI chat platform in the preceding week, roughly 175.5 million adults, up from 52% in February 2026.1
Global search layer: Google AI Overview, the AI-generated summary that appears above organic search results on the world's most-used search engine, reaches more than one billion users globally.9
Conversational scale: ChatGPT has crossed 900 million weekly active users.10
These figures do not tell us how many of those users are asking political questions. They do not tell us what answers those users are receiving, or how those answers compare to what campaigns intended to communicate. But they establish the outer boundary of potential exposure at a scale that is no longer speculative. And the trajectory matters as much as the snapshot: two years ago, the Edison figure was roughly half of what it is today, and there is no evidence of deceleration ahead of 2028.
More than half of American adults are now in regular contact with systems that will answer their questions about candidates, answers those candidates did not shape.
The structural difference in those answers matters as much as the scale. When a voter received a search results page, they saw sources. They made a judgment about which to trust, clicked through, and formed an impression shaped by what they chose to read. The source was at least nominally visible. A campaign could audit the results page and see what voters were finding.
When a voter receives an AI-generated answer, the source is obscured. The answer arrives without a byline. It does not say "this characterization came from a 2019 local news article that may no longer reflect the candidate's current record." It sounds like a neutral, knowledgeable source, and it can produce characterizations of a candidate's record, including inaccurate ones, with apparent confidence.
Campaign communication strategies are designed for an information environment where sources are attributable, messages are addressable, and the channels voters use are at least partially legible to campaigns. Conversational AI operates differently on all three dimensions. Sources are aggregated and obscured. The message cannot be addressed because the campaign did not participate in its creation. The channel's internal weighting and retrieval logic is proprietary.
Optimizing a campaign website for Google's organic rankings is a different task from ensuring that chatbots like ChatGPT and Google AI Overviews draw accurate characterizations from the broader information record. Both matter. They require different capabilities.
Across Kyanos's monitoring of the major chatbot platforms, ChatGPT, Google AI Overview, Perplexity, Microsoft Copilot, Gemini, and Claude, a consistent pattern emerges: voters asking about candidates receive answers that were not shaped by those campaigns in any direct sense. The campaigns had no input into the indexed sources. They had no review of the response before it was delivered. In many cases, they have no awareness that the exchange occurred.
The question "how many voters will be influenced by chatbots in 2026 and 2028?" cannot currently be answered with precision, for a structural reason, not a data-collection reason. "Influence" is a causal concept, and establishing causal relationships between AI-delivered information and vote choice would require controlled research designs that do not yet exist at political-campaign scale.
What can be measured is exposure: how many voters are receiving AI-generated information about candidates, what that information says, and how it compares to what campaigns intend to communicate. That measurement is available now. The absence of a causal model is not a reason to defer monitoring. It is a reason to begin it.
Political professionals have developed a productive habit of treating each cycle as preparation for the next: television disciplines built in the 1980s inform digital strategy today; social media lessons from 2008 and 2012 shaped 2016 and 2020. In this frame, it is tempting to treat chatbot influence as a 2028 problem. That framing is mistaken. More than half of American adults are already using AI platforms weekly. Google AI Overview is already the first thing more than a billion users encounter when they search. The information environment of the 2026 midterm cycle is the AI environment. Campaigns not monitoring what AI says about them in this cycle will arrive at 2028 with two cycles of compounded ignorance rather than one.
The question is not whether chatbots will influence voters in 2026 and 2028. The question is whether campaigns will have any visibility into what voters are being told, and any capacity to respond when the answers are wrong.
Earlier monitoring is not just faster. It is structurally more useful. A campaign that begins in January establishes a baseline against which drift can be detected and corrections measured. A campaign that begins in October is still establishing a baseline, with no time to act before voters go to the polls.
Systematic AI answers monitoring does not resolve the causal question. It creates the conditions under which the causal question becomes answerable, and it provides immediate operational value independent of that longer-term research objective.
Chatbots like ChatGPT and Gemini produce inaccurate characterizations of candidates' records, positions, and backgrounds. Identifying those inaccuracies requires systematic querying across platforms: the same query posed to multiple systems, the responses evaluated against a verified factual baseline. Without this, campaigns learn about AI-generated inaccuracies reactively, typically through press coverage, by which point the information has already reached voters.
These chatbots are updated continuously. A response that accurately characterized a candidate's position in January may not do so in October. Monitoring establishes the baseline against which drift can be measured, enabling campaigns to identify deterioration in what AI says about them before it affects a significant portion of the electorate.
The corrections available to campaigns, structured content, consistent cross-platform documentation, corrections to the third-party sources retrieval chatbots draw from, require knowing precisely what the problem is before it can be addressed. Monitoring is the diagnostic step that makes remediation possible. Without it, corrections are made without evidence of what needs correcting or when.
Each of these capabilities requires a baseline: a documented record of what chatbots like Perplexity and Microsoft Copilot were saying before any intervention was made. That baseline can only be established by monitoring, and monitoring can only establish a useful baseline if it begins early enough to observe the information environment before the campaign sprint compresses all decisions.
The question posed in this paper's title does not yet have a precise answer. No one has produced a rigorous causal estimate of chatbot influence on voter attitudes or behavior in the 2026 or 2028 cycles, and the data infrastructure to produce one does not yet exist at scale.
But the adoption data establishes the outer boundary of potential exposure at a scale that makes the absence of a precise answer a problem, not a reason for patience. More than half of American adults are already in regular contact with AI platforms that will respond to questions about candidates, responses those candidates did not shape, cannot monitor without dedicated effort, and cannot correct without first knowing what they say.