answer engines are the uncontrolled last mile of every campaign channel. Every dollar you spend on TV, mail, canvassing, and digital now partially depends on what happens when a voter goes to verify. Here is how to think about the math.
Campaigns spend money to move voters. They buy television to reach them, mail to remind them, canvassers to persuade them, and digital to target them. The channel mix changes every cycle. The logic doesn't: reach a voter, persuade them, and hold that persuasion through election day.
That logic now has a gap.
When a voter is moved by a campaign contact — a door knock, a TV ad, a mailer — a growing fraction will, within hours, open their phone and verify what they heard. They Google the candidate. They ask ChatGPT. They check Perplexity or Google AI Overview. What they find in that moment either confirms the contact and hardens the persuasion, or contradicts it and unravels it.
This is the last mile. Campaigns have always known it existed in some form — a voter talks to a neighbor, sees a counter-ad, reads an article. What has changed is the medium intercepting them. answer engines are now the first thing many voters encounter when they go to learn more. And unlike a neighbor or a newspaper, an answer engine is available instantly, responds in full sentences, and — as December 2025 research established — is roughly four times more persuasive than a political advertisement.
The central argument. answer engines are not a new campaign channel. They are the uncontrolled last mile of every existing channel. Every dollar a campaign spends on television, mail, canvassing, and digital now partially depends on what AI platforms say when a curious voter goes to verify. Most campaigns are not monitoring this layer at all.
AI exposure happens through two separate, additive layers. Understanding the difference matters because campaigns are typically tracking only one of them.
In 2024, 11% of voters reported using ChatGPT or Google Gemini specifically to research candidates.3 This is the number campaigns hear about. Applied to a competitive congressional district of approximately 300,000 likely voters, it represents roughly 33,000 intentional AI queries about the candidates in the race.
Google AI Overview is a different phenomenon entirely. It is not something voters choose. It appears above the first organic search result — before any news article, any campaign website, any human-curated result — for every voter who Googles a candidate's name for any reason. They saw your television ad and wanted to know more. They received a mailer and searched to verify a claim. They heard the name in a local news story. They are not "using AI." They are Googling. The AI intercepts them.4
Most of those voters do not know they are reading AI-generated content. They believe they are reading a search result.
These two layers are not overlapping. They are additive. A voter who opens ChatGPT to research a candidate is a different event from the same voter Googling the candidate's name. For a competitive congressional race:
| Layer | How It Reaches the Voter | Est. Congressional Volume |
|---|---|---|
| Dedicated AI tools | Voter deliberately opens ChatGPT, Gemini, or Perplexity to research the candidate | ~33,00011% × 300K likely voters — CCI 2024 |
| Google AI Overview | Voter Googles candidate name for any reason; AI-generated answer appears before any organic result | ~90,000–180,000Est. 30–60% of likely voters conducting at least one candidate name search |
| Total organic AI encounters | ~120,000–210,000 |
Google AI Overview volume is the author's estimate, informed by survey data on where Americans get election news and by the architecture of Google AI Overview as a pre-result layer on informational queries. It is not a measured figure and is not drawn from any published study.8 For a Senate race, multiply by 10–15×.
The Google AI Overview layer is likely three to five times larger than the dedicated tool layer. The 11% figure — the one campaigns reference — describes only the visible, intentional slice. The larger exposure happens passively, at scale, and is currently invisible to virtually every campaign.
Edison Research, Infinite Dial 2026. Doubled in two years.
Center for Campaign Innovation post-election survey, n=1,500.
Appears above first organic result on all informational queries.
In December 2025, two studies published simultaneously in Nature and Science — involving real voters in four countries across three national elections — established something campaigns should treat as a material fact about the current information environment.5
The mechanism is not emotional manipulation or psychological tricks. It is factual density. These answer engines generate large numbers of specific claims rapidly, giving them the persuasive weight of a well-briefed expert in a one-on-one conversation. When researchers blocked the AI from citing facts, the persuasive effect dropped substantially. The AI persuades the way a prepared advocate does — not the way a :30 spot does.
Two implications follow directly from this finding.
An unfavorable AI answer does not neutralize a persuasion and leave the voter at neutral. It moves the voter in the wrong direction, and recovering that voter means overcoming both the lost persuasion and an active counter-persuasion. An earlier version of this paper priced that recovery at 2–4× the original persuasion, derived from the relative persuasive magnitude of answer engines versus advertising. That derivation does not hold — a ratio of effect sizes is not a ratio of recovery costs, and it double-counts a loss the model already captures — and it has been removed. The cost of a reversal is treated here as the cost of the persuasion lost.
Across the models tested, the more persuasive a model was, the less accurate the information it provided; the methods that raised persuasiveness most — post-training and prompting — systematically reduced factual accuracy. Separately, answer engines instructed to advocate for right-leaning candidates made more inaccurate claims than those advocating for left-leaning candidates in all three countries, and the two U.S. answer engines were not equally effective: the pro-Harris answer engine moved likely Trump voters 3.9 points, while the pro-Trump answer engine moved likely Harris voters 1.51 points.6
Persuasiveness and factual accuracy are distinct properties of a model's output, and the research finds them in tension. Both describe how these systems behave rather than anything a campaign can set. What a campaign can affect is the record those systems draw on, and what it can know is what they are currently saying.
The Nature/Science studies used chatbots explicitly instructed to persuade. Organic AI answers — the ones voters encounter when they Google a candidate or ask an AI a question — are informational, not persuasion-optimized. Their persuasive effect is real but lower than the figures above. A separate Yale study (PNAS Nexus, March 2026) reported that unprompted AI summaries of historical events moved opinions, with modest effects its authors suggest could compound over repeated interaction.10 The 4× multiplier should be read as the upper bound on AI persuasive power; organic exposure operates somewhere below that ceiling.
Two studies complicate this picture and are worth stating plainly. A UK AI Security Institute study found that political knowledge rose about equally for people researching with answer engines and with search engines, with belief in true information increasing and belief in false information decreasing to the same extent for both.12 And a PNAS experiment using a retrieval-grounded bot built on official party platforms raised users' knowledge of party positions by roughly 13 points while barely moving their party evaluations or vote preference.13 Neither measures the question this paper is about — how accurately a named candidate's record is represented across engines — but both point the same way. An informational AI encounter is a weaker persuasive event than a persuasion-optimized one, and the reversal rate used in Section 04 should be read as an upper bound rather than a central estimate.
The persuasion value at risk from unmonitored AI answers can be estimated. The math requires three inputs: how many voters the campaign persuades, what share of those voters encounter AI content after the contact, and what share of those AI encounters contain unfavorable content.
The first input is knowable. The second is estimable. The third is unknown without active monitoring — and it is the most consequential variable in the calculation.
For every 1,000 voters a campaign successfully persuades:
| Step | Estimated Rate | Result per 1,000 Persuaded Voters |
|---|---|---|
| Persuaded voters who go online to search or verify | 15–30% | 150–300 voters |
| Those searches that encounter AI content | 50–70% | 75–210 AI encounters |
| AI encounters with unfavorable framing or error | 15–35%Unknown for unmonitored candidates — this is the number the diagnostic produces | 11–74 unfavorable encounters |
| Voters whose persuasion is reversed | 25–50% | 3–37 reversals |
The wide range reflects genuine uncertainty at each step. The point is not a precise figure. It is that the loss is real, proportional to persuasion spend, and invisible to campaigns that are not monitoring the layer.
A fully-loaded door-knock persuasion costs approximately $570 per durable shifted voter.1 A television or digital-driven persuasion in a competitive federal race runs $300–600. The figures below apply that cost directly to the reversals produced by the chain of loss above. The $570 figure is carried across from turnout field experiments rather than persuasion experiments, and is used here as an approximation.
| Race | Persuasion Spend | Estimated Persuasion Value at Risk | As % of Persuasion Spend |
|---|---|---|---|
| Competitive State House | $150K | $400–$5,500 | <11% |
| Competitive Congressional | $1.5M | $4,200–$55,000 | 0.3–3.7% |
| Competitive Senate / Statewide | $8M | $22,500–$294,000 | 0.3–3.7% |
Value at risk figures are the author's estimates derived directly from the chain-of-loss model above. A 2–4× recovery multiplier applied in the previous version has been removed as unsupported, which is why these figures are roughly a third of those published before. The percentage band is identical across tiers because the model is proportional to spend. All inputs are explicitly speculative; actual values depend on the unfavorable answer rate, which is unknown without monitoring.
The unfavorable answer rate — what share of AI responses about an unmonitored candidate contain material errors or unfavorable framing — is the most consequential input in this model. Published research has begun to measure it, though not yet for a US race. An audit of four leading answer engines during two 2026 Irish byelections found wrong answers about who was on the ballot and where to vote, markedly different source selection across systems, and attention concentrated on frontrunners while candidates further down the field were largely ignored.11 No equivalent US figure has been published, and the rate varies by candidate, platform, and how much structured information exists about the candidate online.
The appropriate response to that uncertainty is not to discount the model. It is to measure: the act of monitoring is precisely how the unfavorable answer rate for a specific candidate becomes known. The diagnostic produces this number. Without it, the value at risk estimates above remain ranges. With it, they become a specific dollar figure for a specific race.
The chain-of-loss model in the previous section is the defensive case — persuasion value lost to unfavorable AI answers. There is an equally important offensive case that most campaigns are missing entirely.
The 120,000–210,000 organic AI encounters in a competitive congressional race include voters the campaign never contacted. They Googled the candidate because they saw an ad — perhaps a competitor's ad — heard the name in a news story, or received opposition mail. They had no contact with the campaign. What the AI said was the first impression they formed.
If those encounters are favorable and accurate, they are doing persuasion work for free, at a scale the campaign did not pay for. If they are unfavorable or inaccurate, they are doing damage the campaign cannot see.
At a midpoint estimate of 150,000 organic AI encounters in a congressional race: if a candidate's information record is associated with a 10-percentage-point improvement in favorable answer rate — 15,000 additional favorable encounters. At 2% of those encounters producing durable persuasions, that is 300 additional voters moved, with an equivalent advertising cost of $171,000.
Against a $10,000 remediation investment, that is a 17:1 return — before counting the defensive value of catching unfavorable answers before they reach voters at scale.
Every number in the two paragraphs above is an assumption chosen to show the scale of the exposure, and none of them is a finding. The 10-percentage-point improvement is assumed, not predicted: Kyanos does not claim that any action produces a given improvement in how an answer engine describes a candidate, and no published research establishes such a figure. The 2% conversion from a favorable AI encounter to a durable persuasion is likewise unmeasured, and two studies cited earlier in this paper point below it. The $570 cost per voter is carried across from turnout field experiments. Treat the $171,000 and the 17:1 as order-of-magnitude illustrations of what the encounter volume could be worth if these assumptions held, not as a return anyone should budget against.
Remediation does not guarantee favorable AI answers. These answer engines are black boxes; their responses cannot be controlled or predicted. It is worth being explicit about what the research does not support. Where the December 2025 studies examined persuasiveness and factual accuracy together, they found the two in tension, with the more persuasive models providing less accurate information. Nothing in that work shows that a better information record makes an AI more persuasive about a candidate. The arithmetic above is conditional on an improvement in favorable answer rate, and the 17:1 figure is illustrative, not a guaranteed return. It is presented to establish the structural scale of the opportunity, not to make a specific performance promise.
Whether the ambient AI layer is currently working for or against a given candidate is unknown to most campaigns. The diagnostic answers that question. Remediation addresses what the diagnostic finds. Ongoing monitoring tracks changes over time — including the shifts MIT researchers observed in model responses across the 2024 cycle.7
Competitive campaigns typically invest in several intelligence services to monitor the information environment. These services are well established, and their costs are treated as standard line items:
| Service | What It Tracks | Typical Cost9 |
|---|---|---|
| Earned media monitoringCision, Meltwater, comparable | Press coverage, broadcast mentions, online news | $2,000–$10,000/month |
| Opposition research | What opponents may say; vulnerability identification | 3–8% of total budget |
| Tracking polls | Voter preference movement over time | $5,000–$15,000/cycle |
| Digital ad monitoring | Competitor digital creative and spend patterns | $500–$2,000/month |
None of these services track what AI platforms say about a candidate. That is not a criticism of these tools — they were built before answer engines became primary information intermediaries. It is simply a gap in what the standard intelligence stack currently covers, at a moment when the December 2025 research suggests AI may be the most persuasive channel voters encounter.
The research and the model point toward three distinct services, each addressing a different phase of the problem. What follows is a recommendation on scope and rational budget, derived from the break-even analysis above. These are vendor-agnostic ranges — they describe what the work should cost based on the value at risk, not what any specific provider charges.
Every campaign spending meaningful money on federal persuasion should know what AI platforms say about their candidate before significant spend begins. A diagnostic should cover the major platforms where voters are likely to encounter AI-generated political content: at minimum ChatGPT, Google AI Overview, Perplexity, Meta AI, Microsoft Copilot, and Grok. It should test responses across question types — candidate background, policy positions, opponent comparisons, and general searches — not just a single prompt. A single prompt is not representative of the range of ways voters actually query a candidate's name.
The output should give the campaign a platform-by-platform picture of what is accurate, what is missing, what is framed unfavorably, and what is factually wrong. That picture is the input the chain-of-loss model requires to become a specific dollar figure rather than a range.
Ranges are directional. They reflect 2026 market conditions derived from the break-even model above; actual costs vary by vendor, race scope, and what the diagnostic finds.
If the diagnostic reveals material errors, significant gaps, or systematically unfavorable framing, remediation is the appropriate next step. Remediation means building or correcting the structured information record that retrieval answer engines draw on — official biography, policy positions, legislative record, endorsements, public statements — in formats that answer engines like Perplexity and Google AI Overviews can accurately retrieve and represent.
Not all AI problems are equally addressable. An error in a widely-indexed source (Wikipedia, a major news archive) is harder to correct than a gap in the candidate's own structured online presence. A reputable vendor should be able to distinguish between these cases and scope remediation accordingly. Be skeptical of remediation proposals that do not begin with a clear diagnosis of which specific problems are addressable and how.
Ranges are directional. Derived from 5–15% of estimated value at risk; actual scope and cost depend on what the diagnostic finds and which gaps are addressable.
A one-time diagnostic has a limited shelf life. The MIT and Carnegie Mellon research documented that AI model responses to identical political questions shifted — sometimes abruptly — over the course of the 2024 election cycle, without public notice from model developers.7 A campaign that runs a diagnostic in August and assumes the results hold through November is taking on unexamined risk.
Ongoing monitoring should run at a cadence sufficient to detect changes before they propagate at scale — bi-weekly is a reasonable standard for an active race. The monitoring should cover the same platforms and question types as the diagnostic to allow direct comparison over time.
The earned media monitoring comparison is a useful pricing anchor. Campaigns routinely pay $2,000–$10,000 per month for press monitoring services. AI platform monitoring involves similar infrastructure and covers a medium that — per the December 2025 research — appears to be more persuasive per encounter than the press coverage those services track. Monthly rates substantially above the earned media monitoring range are difficult to justify on the basis of effort alone.
Ranges reflect 2026 market rates for comparable earned media monitoring at each race tier.
The diagnostic is the right entry point regardless of what follows. It converts the value at risk model from a range into a specific number for a specific candidate, and it tells the campaign which problems are present and which are not. A campaign that runs a diagnostic and finds clean answers has spent a modest sum to establish a baseline and a documentation trail. A campaign that runs a diagnostic and finds systematic errors has avoided spending a full persuasion budget on top of an undetected problem.
The diagnostic also disciplines the remediation conversation. Without a clear picture of current AI representation, remediation proposals are speculative. With it, scope, priority, and cost can be evaluated against a documented baseline. This is the same logic that makes opposition research the precondition for communications strategy — the intelligence has to precede the response.
For any competitive federal race spending more than $500,000 on persuasion: commission a diagnostic before significant channel spend begins. Budget $3,000–$15,000 depending on race tier. Evaluate remediation against the specific findings — not as a standard add-on. If the race extends into the fall with active paid media running, add ongoing monitoring at $1,000–$5,000/month.
The total investment for a congressional race — diagnostic plus remediation if warranted plus monitoring through a six-month active period — is unlikely to exceed $25,000, or 1.6% of a $1.5 million persuasion budget. Whether that investment is warranted depends on what the diagnostic finds.
These ranges are offered as market guidance, benchmarked against what campaigns already pay for comparable intelligence. The defensive arithmetic in this paper does not on its own justify them. At the base-case inputs, a congressional program at the top of these ranges costs more than the modeled value at risk, and three of the model's terms — the post-contact search rate, the reversal rate, and the cost per persuasion — are unmeasured or carried across from a different literature. A reader who accepts only what is measured should read the spend guidance as a market comparison and not as a return calculation.
The argument for measuring does not rest on reversal arithmetic, and it survives every caveat above. Four leading answer engines asked about two 2026 byelections gave wrong answers about who was on the ballot, drew on markedly different sources from one another, and concentrated their attention on frontrunners while largely ignoring candidates further down the field.11 For a challenger or a down-ballot candidate, the finding is not unfavorable framing but absence. Analysis of real assistant conversations finds that when people bring politics to an AI they overwhelmingly ask for explanation and analysis — only about 3% of political requests seek advice or a recommendation.14 They are asking who this person is. And what comes back is shaped by what is crawlable and structured rather than by what is authoritative, in an environment where campaign and party sites remain open to crawlers while independent news organizations increasingly block them.15 None of that requires a persuasion model to matter. It only requires that voters ask, which they do.
[1] Alan S. Gerber and Gregory A. Huber, "Getting Out the Vote Is Tougher Than You Think," Stanford Social Innovation Review, 2016. Note that this source synthesizes turnout (GOTV) field experiments, not persuasion experiments, and its headline finding is that mobilization effects are considerably smaller than previously believed. The ~$570 cost per durable shifted voter used in this paper is carried across from that turnout literature and should be read as an approximation. ↩
[2] Edison Research with SSRS, February 2026: 52% of Americans 18+ used an AI chatbot weekly (SSRS). Edison's May 2026 wave put the figure at 65%. ↩
[3] Center for Campaign Innovation, "2024 Post-Election National Survey," fielded November 3–7, 2024, n=1,500. 11% of voters used ChatGPT or Google Gemini to research candidates. ↩
[4] Google AI Overviews: company-disclosed reach, reported at 2 billion monthly users on Alphabet's Q2 2025 earnings call and above 2.5 billion at Google I/O 2026. The "1B+" figure in the previous version of this paper was already out of date when published. ↩
[5] Hause Lin, Gabriela Czarnek, Benjamin Lewis, Joshua P. White, Adam J. Berinsky, Thomas Costello, Gordon Pennycook and David G. Rand, "Persuading voters using human–artificial intelligence dialogues," Nature 648, 394–401, December 4, 2025 (nature.com, subscription). Companion paper: Kobi Hackenburg et al., "The levers of political persuasion with conversational artificial intelligence," Science 390(6777), December 4, 2025 (science.org, subscription; free preprint at arXiv:2507.13919). Accessible summaries: Cornell Chronicle; MIT Technology Review. ↩
[6] Cornell University Chronicle, December 4, 2025. On the study's 0–100 candidate-preference scale, the pro-Harris answer engine moved likely Trump voters 3.9 points; the pro-Trump answer engine moved likely Harris voters 1.51 points. The same release reports that the more persuasive a model was, the less accurate the information it provided, and that models advocating for candidates on the political right made more inaccurate claims in all three countries. ↩
[7] Sarah Cen et al., MIT CSAIL, MIT Sloan and MIT LIDS, 2025 (lead author now at Carnegie Mellon University). Nearly daily queries across 12 models on more than 12,000 prompts between July and November 2024, producing over 16 million responses; candidate–trait associations shifted over the period, in some cases tracking news events. MIT CSAIL; reported in Tech Brew, October 6, 2025. ↩
[8] Pew Research Center, "Americans' Views of 2024 Election News," October 10, 2024. Pew reports that 8% of US adults name search engines as their most common source of political and election news. The 30–60% candidate-name-search estimate used in the table above is the author's own and is not a Pew finding; no published measurement of post-contact candidate search behavior is known to the author. ↩
[9] Earned media monitoring pricing reflects market rates for Cision, Meltwater and comparable services as of 2025–2026. Race-tier persuasion spend estimates from Tech for Campaigns, "2024 Digital Ads Report." ↩
[10] Matthew Shu et al. (Daniel Karell, senior author), "How latent and prompting biases in AI-generated historical narratives influence opinions," PNAS Nexus 5(3), pgag022, March 2026 (open access); Yale News. n=1,912. The study tested AI-generated accounts of two 20th-century events rather than candidate information; the senior author describes the effects as modest and says they "could compound" with frequent use. ↩
[11] Audit of Claude, ChatGPT, Gemini and Grok during the Galway West and Dublin Central byelections, University College Dublin and the University of Strathclyde, May 2026. Approximately 200 election-related questions per system. UCD News; method detail at the Connected_Politics Lab; coverage in The Irish Times. ↩
[12] UK AI Security Institute, "Do chatbots inform or misinform voters?," September 30, 2025. Also reports that 32% of answer engine users, or 13% of eligible UK voters, used AI to seek political information during the 2024 election. ↩
[13] Yamil R. Velez, Donald P. Green and Semra Sevi, "Chatbot Voting Advice Applications inform but seldom sway young unaligned voters," PNAS 122(50), December 2025 (pnas.org; free full text at PMC). ↩
[14] Andy Hall and Dan Thompson, "Mapping America's Political Conversations with Claude," Free Systems, August 4, 2026. Analysis of Anthropic Economic Index data covering Claude conversations in April and May 2026. Politics accounted for roughly 0.5% of US conversations, ranking 66th of 189 topic categories; nearly half of political requests produced explanations and just 3% ended in advice or a recommendation. Published on Substack rather than in a peer-reviewed journal. ↩
[15] Andy Hall, "AI Is A Shitty Political Advisor," Free Systems, March 19, 2026, with survey data contributed by Sean Westwood and Justin Grimmer. Hall's stated central finding is that AI political advice "is driven by what is crawlable and structured — not by what is authoritative or accurate," and he documents news organizations restricting AI crawlers while party and campaign sites remain open. Published on Substack rather than in a peer-reviewed journal. The author discloses consulting relationships with a16z crypto, Forum AI and Meta Platforms, and states that his writing is independent of those roles. ↩