Your paid media strategy assumes a voter who starts with an open mind. But a growing share of voters are forming their first impression of your candidate somewhere else — in a conversation with an AI answer engine that answered their question before your ad had a chance to run. This paper is about that conversation, what shapes it, and what campaigns can do about it.
Every campaign budget is built around a model of how voters form opinions. They see an ad. They visit a website. They read a mailer. They watch a debate. They talk to a neighbor. The model is imperfect — campaigns have always had limited insight into which of these contacts actually moved people — but the channels were knowable, and investment in them was trackable. You bought media and you measured reach. That was the game.
The model now has a gap. A growing share of voters, before they see your ad or visit your site or read your mailer, are asking an AI answer engine about your candidate. Not searching and reading — asking. And the answer they receive is not a ranked list of sources they must evaluate. It is a single synthesized response, delivered in confident plain language, assembled from whatever the AI system found most credible in the information environment you have or have not built.
That conversation is happening without you. The question is what it is saying.
"The voter is not reading your press releases. They are asking a machine what it already knows. And the machine answers based on what you have — or have not — built into the public record."
The scale of this shift is measurable, though it is still early and the data is uneven across geographies and demographics. In September 2025, the UK's AI Safety Institute published survey data from the 2024 UK general election: 13 percent of voters reported using AI answer engines for political information during the campaign. That figure is from a 2,499-person survey conducted after the election. It does not tell us how much those AI-sourced answers influenced their ultimate choices. What it does tell us is that AI answer engines were a real information source for a meaningful share of the electorate — not a fringe behavior.
For comparison: a December 2023 survey by the Bipartisan Policy Center asked American voters whether they expected to use AI for election information. Only 2 percent said yes. Whether that expectation gap has since narrowed is unclear — the American data is more dated, and self-reported intent surveys routinely undercount actual behavior. What we can say with confidence is that by 2025, AI answer engines had become a genuine, if still secondary, channel for voter information in at least one major democratic election. The expectation that this dynamic stops at the US border is not well-founded.
These are commercial analytics studies and Pew research, not controlled academic experiments. The causal direction — AI Overviews causing click-through decline versus AI Overviews appearing on high-traffic queries that were always lower-click — is genuinely contested. But the structural implication is clear: when voters search for your candidate and receive an AI-synthesized answer, a substantial share of them are reading that answer and moving on. They are not clicking through to the campaign website. They are not reading the press release. They are forming an impression from the AI's answer — and that answer was built from whatever was available in the information environment at the moment.
A voter asking about your candidate is not using one answer engine. They may ask ChatGPT on a laptop, get a Google AI Overview on a phone, or read what Perplexity assembles. Managing presence here means managing six surfaces rather than one, and the number is growing.
The asymmetry this creates is significant. Your paid media is structured to reach voters who have not yet formed a strong view. AI is answering the questions of voters who are in the process of forming one. If the AI answer precedes your ad — and on many queries, it will — it is setting a frame that your ad then has to either confirm or overcome. The campaign that understands this and manages it is playing a different game than the campaign that does not.
Every dollar of political digital spend falls into one of two categories. The first is rented reach: paid impressions that exist while the budget runs and stop when it stops. Display ads, pre-roll video, social media promotion, paid search — these are the mechanisms by which campaigns buy access to voter attention. When the money runs out, the impressions stop. The reach was real while it lasted; it does not accumulate.
The second category is built authority: content and infrastructure that persists in the information environment after it is created, accrues credibility through corroboration over time, and continues to influence what voters find — and what AI systems return — without additional spend to maintain it. A well-structured policy page. A Wikipedia article that accurately reflects the candidate's current record. A Ballotpedia entry completed through the Candidate Connection process. A series of substantive LinkedIn articles. Press coverage that is archived, linked, and findable. These do not stop working when the budget cycle ends.
AI presence falls almost entirely into the built authority category. Retrieval-augmented generation — the architectural approach underlying most major AI search and answer engine systems — works by pulling external documents into context at query time. The systems retrieve content from the web, weight it by signals that include authority, corroboration, and recency, and use it to generate an answer. Content that is well-structured, authoritative, and consistently documented across multiple indexed sources tends to appear in AI answers. Content that exists only in targeted ad inventory does not.
This is not a claim about how any specific AI system ranks or retrieves content — those mechanisms are proprietary and not publicly documented. It is a structural observation about what paid media can and cannot reach: ad inventory and AI retrieval are different information channels, and they do not overlap in ways that campaigns can currently purchase.
| Attribute | Rented Reach (Paid Impressions) | Built Authority (AI Presence) |
|---|---|---|
| Durability | Ends when the budget stops; no residual effect | Persists as long as source pages remain indexed and credible |
| Perceived authority | Labeled as sponsored content; audience applies appropriate skepticism | Presented as AI-synthesized answer; audience skepticism is lower by design |
| Timing of effect | Active during the buy; concentrated in campaign sprint | Compound over time; early content continues influencing later responses |
| Reach mechanism | Determined by targeting parameters and platform algorithms | Covers answer engine queries that retrieve those sources |
| Marginal cost curve | Linear: more spend equals more impressions | Sub-linear: structured content serves many queries once built |
| Audience receptivity | Interrupts; voter is engaged in something else | Answers; voter actively sought information |
The table above describes structural properties, not measured outcomes. Attribution between ad impressions and AI citation influence is not directly measurable with current tools — we cannot run a controlled experiment where some voters receive an accurate AI answer and others receive an inaccurate one and measure the vote-share difference. What we can measure is whether AI systems are drawing from campaign-controlled sources or from other sources when they answer questions about the candidate. That measurement is the starting point for managing the channel.
"You are not being asked to stop buying media. You are being asked to recognize that you have been buying into only one of two information environments. The other one is not for sale — it has to be built."
The practical question for a campaign manager or chairman is not whether to eliminate paid media — it remains the primary vehicle for reach and persuasion at scale, and nothing in this paper argues otherwise. The question is whether the current allocation reflects an accurate model of where voter opinions are being formed. A campaign that spends heavily on display ads to reach voters who are actively searching for the candidate, and nothing on ensuring that the AI answer those voters receive before clicking is accurate and favorable, has a gap in its strategy that rented reach alone cannot close.
The investment required to build authority in the AI information environment is different from the investment required to buy media. It is more front-loaded, more time-dependent, and less immediately measurable. It is also, structurally, more durable. A dollar spent building a well-documented policy record that AI systems can retrieve continues working in month six. A dollar spent on a pre-roll ad in month six works only in month six.
The case for built authority gets stronger when you account for who else researches your candidate through answer engines. Major donors check before they write the cheque. PAC research directors allocating across a competitive field run the same queries, faster, because it is their job. Journalists prepare background before they report, and what they read shapes the frame they bring. Party committee staff assessing viability do the same. That is the multiplier: a dollar of built authority reaches voters directly, and simultaneously shapes what the people deciding whether to fund, cover or recruit the candidate find when they look.
The research on AI and political information is growing rapidly, but it is uneven. Some findings are well-documented. Others are preliminary. A few that circulate widely in political communications circles do not hold up well under scrutiny. This section reviews what is actually established, where the caveats matter, and where the research is genuinely thin.
The UK AI Safety Institute's survey data — 13 percent of UK voters reporting AI answer engine use for political information in the 2024 general election — is the most rigorous published figure on actual voter behavior during a campaign. It is a self-report survey, which means it measures what voters recall doing, not what they actually did. It covers one election in one country. Generalizing it to American races of varying levels and geographies requires care.
What the figure establishes is that AI answer engines are not a hypothetical future channel. They were a real information source for a non-trivial share of the electorate in a major democratic election in 2024. Whether equivalent figures for American voters would be higher, lower, or similar is genuinely unknown — the comparable US data from the same period does not exist at the same methodological quality.
A December 2025 study published in a peer-reviewed journal, conducted by researchers at RAND and the University of Toronto's Rotman School, ran a randomized experiment in which participants answered questions about candidates before and after interacting with an AI answer engine that provided political information. The study found that the AI interactions moved candidate preference by approximately 3.9 percentage points on average.
This is a laboratory finding, not a field study. The effect observed in a controlled experiment — where participants were explicitly interacting with an AI tool for research purposes — may not replicate in real-world conditions where AI answer engine use is one of many information sources and is not the study's explicit focus. The 3.9-point figure is a directional signal that AI answer engine interactions can move political attitudes, not a prediction about how much they will move attitudes in a given real race. It should be read with that caveat clearly in view.
A 2024 study by the Reuters Institute at Oxford University tested Perplexity and ChatGPT on 100 questions about the UK election — biographical facts, policy positions, voting records, endorsements. Perplexity answered accurately 83 percent of the time. ChatGPT answered accurately 78 percent of the time. Inaccurate answers were more common on nuanced policy questions than on factual biographical queries.
A separate study by the Tow Center for Digital Journalism at Columbia University, published in February 2025, tested citation accuracy across multiple AI search tools and found error rates ranging from 37 percent to 94 percent depending on the tool and query type. The Tow Center study focused specifically on whether AI systems correctly attributed claims to the sources they cited — a different question from whether the underlying claim was true. Both types of errors — factual errors and citation errors — represent ways a voter can receive inaccurate information through AI.
AI systems do not make errors randomly. They tend to fill gaps in the information record with whatever they can find — which means that candidates with thin, inconsistent, or outdated information records are more likely to be represented inaccurately than candidates with well-documented, consistent ones. The accuracy gap is not a fixed property of AI systems. It is, in part, a function of the information environment the campaign has or has not built.
This does not mean that building a strong information record guarantees accurate AI answers — it does not. AI systems are black boxes and can produce inaccurate answers regardless of source quality. But thin, inconsistent records are reliably correlated with higher error rates in what we have observed. Building a stronger record reduces the surface area for errors. It does not eliminate it.
When a voter asks an AI answer engine about your candidate and receives an answer, that answer was assembled from somewhere. Understanding where — in the aggregate, since the major answer engines do not publish their retrieval logs — is the practical foundation of AI presence management.
Research from the Tow Center and the Reuters Institute, taken together with what is publicly documented about how retrieval-augmented systems work, produces a consistent picture: AI systems draw heavily from third-party sources rather than from campaign-produced content. A February 2026 commercial study by OtterlyAI, tracking citations across major AI platforms, found that approximately 95 percent of AI citations came from third-party sources rather than the subject's own official content. This is a single commercial study and has not been independently replicated — but it is consistent with how retrieval architectures are documented to work. Official campaign websites are one source among many, weighted alongside Wikipedia, Ballotpedia, local news archives, academic databases, and whatever else the system found credible and indexed.
The implication is significant: a campaign cannot control its AI presence by managing only its own website. It has to manage the broader information environment — the third-party sources that AI systems weight heavily, the Wikipedia and Ballotpedia records that appear to function as primary references for many candidate queries, and the news archive that provides both positive and negative corroboration of the candidate's record.
Research from Seer Interactive, a commercial analytics firm, found that approximately 65 percent of AI citations came from content published within the past year, and approximately 89 percent from content published within the past three years. This is a commercial study, not peer-reviewed research, and should be read as a directional signal rather than a precise figure. But it suggests that recency matters in what AI systems draw from — which has direct implications for how campaigns think about content freshness.
A candidate whose policy positions were documented extensively three years ago but whose website has not been updated since may find that AI systems are drawing from older content rather than the current record. A challenger who has produced substantial new content in the past twelve months — LinkedIn articles, structured policy pages, a completed Ballotpedia entry, locally archived press coverage — may have better AI visibility than an incumbent with a longer but more stagnant record.
For campaigns, the practical implication of the LinkedIn research, if directionally accurate, is that a candidate's LinkedIn presence is not adequately served by a completed profile. It is served by published, substantive articles — pieces that give AI systems something to retrieve and cite. A candidate who publishes a 1,000-word piece on housing policy on LinkedIn is giving AI systems a citable, authoritative, first-person document on that issue. That document does not exist in the information environment if it is not written.
Kyanos measures what we call Grounding Share: the percentage of AI-generated responses to queries about a candidate that cite campaign-controlled or campaign-favorable sources. A high Grounding Share means the AI is building its answer from sources the campaign has shaped. A low Grounding Share means the answer is built from whatever the AI found — local news, Wikipedia stubs, old opposition research, or no identifiable sources at all.
Grounding Share is not a sentiment metric. It does not measure whether AI says nice or hostile things. It measures the information provenance of the answer — whose sources is the AI using when it constructs its view of your candidate? That is the question that tells you whether you have work to do.
The source question is where AI presence management meets traditional political communications infrastructure. Press secretaries have always known that the record a candidate runs on is built from what gets written — by the campaign and by others — and accumulated over time. What AI adds is a new mechanism through which that accumulated record gets synthesized and delivered: not a journalist reading the archive and writing a story, but a retrieval system reading the archive and generating an answer. The record that matters is the same record. The mechanism through which it reaches voters has changed.
Political finance research has long documented that early money in a campaign is not just helpful — it is structurally different from late money in its effects. A 2022 study by Rachel Porter and David Steelman in a peer-reviewed political science journal found that early fundraising success has a statistically significant independent effect on future fundraising, above and beyond what candidate quality alone would predict. A follow-on analysis by Porter in 2025 extended this to general election outcomes: early financial credibility — demonstrated through FEC reporting in the first quarters of a race — was associated with better organizational capacity and larger donor networks later in the cycle. Early money signals viability. Viability attracts more money.
The same structural dynamic applies to AI presence — and the mechanics are worth understanding clearly, because they are different from the dynamics of paid media.
AI retrieval systems appear to weight corroborated records — content that has been indexed across multiple sources over time — more heavily than recently published content in a single location. A policy position documented on the campaign website in January, echoed in a local news story in February, reflected in a Ballotpedia update in March, and cited in a LinkedIn article in April has had four months of indexing cycles and cross-source corroboration by the time a voter asks about it in May. A policy position published for the first time in October has had none of that — and is racing to establish a record in the compressed sprint before Election Day.
"Earlier is not just faster. Earlier is structurally more effective. An AI information record built over six months is different in kind from one built in six weeks."
This is not a claim about how any specific AI system handles indexing or corroboration — those mechanisms are proprietary. It is a structural observation about information environments: records that have existed longer, been referenced more widely, and been confirmed by multiple independent sources are more deeply embedded in retrieval indexes than records that are new. That structural advantage can only be built over time. It cannot be purchased at scale in a late sprint.
For campaigns thinking about early money allocation, the AI presence argument parallels the traditional case for early infrastructure investment. Campaigns that build field programs early do not just get more volunteer hours — they build organizational capacity that functions differently under pressure than organizations stood up late. Campaigns that establish credible AI presence early do not just get more favorable AI answers in the first quarter — they build a corroborated, indexed record that compounds in authority as Election Day approaches.
The practical implication of this for a treasurer or campaign chairman making early allocation decisions: AI presence investment is not a line item that can be deferred to the sprint phase without cost. Deferring it does not save money — it changes what money can buy. A dollar invested in structured content and information infrastructure in the first quarter of a campaign does different work than a dollar invested in October. The October dollar is buying recency in a compressed window. The Q1 dollar is buying time — the time for a record to be indexed, referenced, and embedded across the information environment before the voter is actively making up their mind.
An early AI presence investment is not primarily a technology purchase. It is a content and infrastructure investment. The work involves: completing the candidate's Ballotpedia Candidate Connection profile; establishing or updating the Wikipedia article with accurate, cited information; publishing substantive policy content on the campaign website and on LinkedIn; ensuring press coverage is archived and linkable; and implementing basic structured data markup so that AI systems can parse the campaign site's key claims clearly.
None of this is expensive relative to the media budget at any race level. For an emergent campaign — a challenger entering a district or state race — the investment required to establish a credible AI information record in the first quarter of the campaign is modest. The cost of not doing it, measured in the quality of AI answers voters receive during the closing sprint, is harder to quantify but structurally real.
The argument in this section is an extrapolation. No peer-reviewed study has documented that political action committees use AI to research candidates during investment decisions. No publicly available data confirms that PAC research staff are querying ChatGPT or Perplexity when evaluating emergent campaigns. What follows is a reasoned inference from how investment decisions are made, what AI tools are now accessible, and what the absence of a coherent AI presence looks like to a researcher using those tools. It should be read as a hypothesis — one that we believe is worth taking seriously — not as an established finding.
PACs that invest in emergent campaigns — challengers in winnable districts, candidates in open-seat races where the path to victory requires outside support — are making bets on viability. Their research process involves the same questions a voter might ask: Who is this candidate? What is their record? What do they stand for? Is there enough of a public profile here to suggest a campaign that can close in the sprint?
For decades, that research was done through briefing books, direct calls with campaign staff, reviews of FEC filings, and conversations within the political network. All of that still happens. What has changed is that AI answer engines are now a tool available to every researcher — inside PACs, inside journalism, inside opposition research — and those tools return answers based on the information record that exists, not the one a campaign wishes existed.
"A PAC researcher who asks Perplexity about a challenger and receives a thin, uncertain answer is receiving a signal — not about what the AI thinks of the candidate, but about the state of the candidate's public information record."
A candidate with a well-structured, consistently documented public record — accurate AI answers across multiple surfaces, Wikipedia entry current, Ballotpedia complete, substantive LinkedIn content, local press coverage archived and findable — presents differently in an AI-mediated research query than a candidate whose AI presence is thin, inconsistent, or assembled from outdated sources. Whether that difference currently influences PAC investment decisions is unknown. That it could, given the accessibility and ubiquity of these tools, seems like a reasonable inference about a future that may already be arriving.
The practical implication, if this extrapolation has any validity: the AI information record a campaign builds is not only a voter-facing asset. It is increasingly the same record that institutional funders, political networks, and journalists encounter when they research the candidate quickly. Campaigns that treat AI presence as a voter communication tool only may be underestimating its value as a credibility signal across the full ecosystem of stakeholders who make investment decisions during a campaign cycle.
We state again clearly: this is not a finding. It is a reasoned hypothesis about how the value of AI presence may extend beyond direct voter contact, grounded in what we know about how research happens and how AI tools are used. The absence of a strong AI presence record may not yet be disqualifying in any funding decision we are aware of. But campaigns that establish strong records now are building an asset that has value in multiple contexts — and they are building it while it is still relatively uncommon. The first-mover dynamic in political communications has historically rewarded the campaigns that moved before a channel was crowded.
The argument in this paper does not require a campaign to rebuild its digital infrastructure or invent a new staff function from scratch. It requires a set of specific, prioritized actions — most of which are accessible to any campaign, regardless of race level or budget — taken in an order that reflects how AI presence actually accumulates.
Before changing anything, understand what AI systems currently say about your candidate. Open ChatGPT, Perplexity, and Google. Ask the questions a voter in your district would ask — not biographical queries, but issue queries: Where does she stand on housing? What has he done on public safety? What is her record on education funding? Read the answers carefully. Note what is accurate, what is missing, what is wrong, and what sources are being cited. Do this across all six major AI surfaces. Write it down. This is your baseline — and it will tell you where the gaps are before you spend a dollar on fixing them.
Ballotpedia covers candidates at every level of office and does not impose the notability standards that Wikipedia requires for new articles. The Candidate Connection questionnaire creates a structured, AI-visible record of the candidate's background, positions, and campaign. It is free to complete. It is indexed by search engines and appears to be weighted by AI retrieval systems when they construct answers about down-ballot candidates. For any candidate who does not already have a complete Ballotpedia profile, this is the first thing to do — accessible, low-cost, and measurable in its effect on what AI returns about the candidate.
If the candidate has a Wikipedia article, read it carefully against the current record. Outdated information, missing context, and framing inherited from prior news cycles are common. Work within Wikipedia's documented guidelines to update it — Wikipedia has well-established rules about sourcing and conflicts of interest for campaign staff edits, which should be followed. If the candidate does not yet have a Wikipedia article, assess whether she meets notability requirements; for many state and local races, she will not yet, and Ballotpedia is the more accessible starting point. Do not attempt to create an article that does not meet Wikipedia's standards — a declined or deleted article is worse than no article.
The campaign website should function as the authoritative record for every factual claim the campaign wants AI systems to reflect accurately. Policy pages should be specific — not talking points, but documented positions with enough context that an AI system can accurately summarize them. The candidate biography should be accurate, complete, and consistent with what appears on Wikipedia and Ballotpedia. Press coverage should be archived and linked in ways that make it findable. This does not require rebuilding the site; it requires a deliberate review of what is there, updated with the understanding that a machine is also reading it.
Based on current commercial research — with the caveats about single-vendor studies noted in Section 4 — LinkedIn appears to be among the most-cited domains across major AI platforms for professional queries. A completed candidate profile is a start. Substantive published articles — 500 to 2,000 words, on the issues that define the race, in the candidate's own voice — provide AI systems with citable, authoritative, first-person content that does not otherwise exist in the information environment. This is one of the most accessible investments available to an emergent campaign: no technical infrastructure required, no advertising spend, and the content is useful to human readers as well.
Coverage that exists but is not archived, linked, or findable does not contribute to the AI information environment in the way coverage that is. Every time a story runs about the candidate — even local coverage in a small market — it should be linked from the campaign website and, where possible, referenced in subsequent campaign materials. The goal is a paper trail that is consistent, corroborated across multiple sources, and findable by a retrieval system months later. This is not a new discipline for communications directors; it is the existing press archive practice extended with explicit attention to AI retrievability.
The diagnostic you run before you start should be repeated on a schedule — monthly for most campaigns, more frequently during the sprint. AI systems update their information based on what is newly indexed, and the record that exists in January may not reflect what was added in February. Systematic monitoring — running the same question bank across all major AI surfaces at regular intervals — is the only way to know whether the information environment is moving in the right direction. This is where specialized tools become useful for campaigns at scale; for smaller campaigns, a disciplined manual process is a reasonable starting point.
The actions above do not guarantee specific AI answers. AI systems are updated periodically, their retrieval and weighting mechanisms are not publicly documented, and there are no reliable techniques for predicting exactly how any system will respond to any query after implementing these steps. What the actions above do is reduce the surface area for inaccurate or unfavorable AI answers by ensuring that the most authoritative, current, and credible version of the candidate's record is well-documented, widely indexed, and consistently maintained. The outcome of that investment is more AI answers drawn from campaign-controlled sources and fewer assembled from gaps. Whether that translates to measurable vote-share movement is a question that the research base does not yet answer with confidence — and any vendor who claims otherwise is overstating what is known.
What we can say: a campaign that does this work is playing a more complete game than one that does not. The channel exists. Voters are using it. The record it draws from is being written — by your campaign, by your opponent, by local news, by Wikipedia editors, by whoever cares enough to write it. The question is whether you are one of those authors.
The campaigns that will look back on this period as a missed opportunity are the ones that treated AI presence as a future problem — something to address when the data was more definitive, when the platform was more mature, when the budget had room. The data is already here, the platforms are already live, and the voter who asked a answer engine about your candidate last Tuesday already formed an impression from the answer they received. The campaign that addresses this in 2026 is addressing it before the majority of their competitors are. That window does not stay open indefinitely.
Sources cited in this paper
UK AI Safety Institute: Survey data, 2024 UK general election, published September 2025 (2,499-person sample). · Bipartisan Policy Center: American voter AI use survey, December 2023. · Gartner: "Gartner Predicts Search Engine Volume Will Drop 25% by 2026," Alan Antin, February 19, 2024. · BrightEdge: 2025 organic click-through rate analysis, May 2025 (commercial analytics vendor). · Pew Research Center: AI Overview click-through analysis, October 2025. · Porter and Steelman: "Early Fundraising and Campaign Viability," peer-reviewed political science journal, 2022. · Porter: Follow-on analysis, 2025. · Nature (peer-reviewed): AI answer engine political persuasion experiment, Rand/Pennycook et al., December 2025. · Reuters Institute, Oxford University: AI election accuracy study, 2024 (100 UK election questions, n=2 platforms). · Tow Center for Digital Journalism, Columbia University: AI citation accuracy study, February 2025. · OtterlyAI: AI citation source analysis, February 2026 (commercial vendor; not independently replicated). · Profound: LinkedIn citation tracking study, March 2026 (commercial vendor; not independently replicated). · Seer Interactive: AI citation recency analysis (commercial vendor; not independently replicated).
Claims about AI system behavior are observations from measured outputs. They are not claims about internal AI architecture, which is not publicly documented. Extrapolations are labeled as such in the text.