Kyanos · Model Addendum · March 2026

The Last Mile Problem:
The Model

Full chain-of-loss inputs, scenario tables, break-even arithmetic, upside model, and sensitivity analysis behind the main paper. Every assumption is labeled. Substitute your own numbers to stress-test the conclusion.

Ed Forman Founder, Kyanos
March 2026
Companion to:
The Last Mile Problem
Contents
A · Input Variables

Every assumption,
labeled and sourced.

The main paper presents ranges for value at risk without showing all the intermediate steps. This section disaggregates the model into its component inputs, labels each one as measured, estimated, or inferred, and states the rationale for the conservative, base, and aggressive values.

A model is only useful if a skeptical reader can find the assumptions they disagree with and understand what happens when they are changed. That is the purpose of this section.

A1 — Base Population: Likely Voters
Input type: Measured (public record)
The number of likely voters in the race. This is the denominator for everything that follows.
Conservative
250K
Smaller competitive congressional district
Base
300K
Typical competitive congressional
Aggressive
3M+
Senate / statewide race
A2 — Share of Persuaded Voters Who Search Online After Contact
Input type: Estimated (Pew 2024, BPC 2022)
Not all voters who receive a persuasion contact will go online to verify or learn more. This rate captures the fraction who do, within a window relevant to the persuasion event (typically the same day or next day).
Conservative
15%
Lower-engagement voters, older demographics, less competitive race
Base
22%
Midpoint of 15–30% range; consistent with general online verification behavior
Aggressive
30%
High-engagement races, younger electorate, high-salience contest
A3 — Share of Searches That Encounter AI-Generated Content
Input type: Estimated (Google product documentation; industry analysis)
Of voters who search online after a contact, what fraction will encounter AI-generated content — either dedicated tools (ChatGPT, Perplexity) or Google AI Overview? Google AI Overview appears on virtually all informational queries about a named individual; dedicated tool usage is the smaller fraction. Combined reach is the key figure.
Conservative
50%
Lower AI tool adoption, older electorate, non-Google search mix
Base
60%
Midpoint — Google AI Overview penetration is the dominant driver
Aggressive
70%
High AI adoption, younger electorate, Google-primary search behavior
A4 — Unfavorable Answer Rate
Input type: Unknown without monitoring — this is what the diagnostic measures
Of AI encounters about the candidate, what fraction contain material errors, unfavorable framing, missing context, or inaccurate claims? This is the most consequential and least knowable input. It varies by candidate, platform, and information environment. The diagnostic produces this number for a specific candidate. Without it, only a range can be stated.
Low
10%
Well-documented candidate; accurate, comprehensive information record online
Base
20%
Typical candidate with gaps; some errors; limited structured digital presence
High
35–50%
Poorly documented candidate; active opposition; contested record; sparse structured data
A5 — Persuasion Reversal Rate
Input type: Estimated (derived from Nature/Science persuasion research, December 2025)
Of voters who encounter an unfavorable AI answer, what fraction have their persuasion materially reversed? Not every unfavorable answer undoes a persuasion — some voters are committed, some encounters are minor. This rate reflects the fraction who were genuinely moved by the AI encounter in the wrong direction.
Conservative
25%
High-commitment voters; strong prior persuasion; AI encounter brief or low-engagement
Base
35%
Typical persuadable voter; meaningful AI encounter; consistent with 3.9pp shift research
Aggressive
50%
Low-commitment voter; extended AI conversation; multiple unfavorable claims
A6 — Cost per Durable Persuasion
Input type: Measured (Gerber and Green field experiments; Tech for Campaigns 2024)
The fully-loaded cost of one durable shifted voter via campaign contact — door-knock, digital, or mixed channel. The reversal recovery cost is 2–4× this figure, reflecting the cost of overcoming both the lost persuasion and an active counter-persuasion (per the Nature/Science research on AI persuasive effect magnitude).
Low
$300
Digital-heavy campaign; efficient targeting; low-cost media market
Base
$570
Door-knock blended cost from Gerber-Green, adjusted for current labor
High
$800+
High-cost media market; difficult targeting environment; late-cycle spend
A7 — Recovery Cost Multiplier
Input type: Inferred (derived from Nature/Science persuasion magnitude finding)
Recovering a reversed voter costs more than the original persuasion — the campaign must overcome both the original lost ground and an active counter-persuasion. The 2–4× multiplier is derived from the relative persuasive magnitude of AI chatbots versus political advertising (4× per the December 2025 research). 2× represents a conservative recovery assumption; 4× represents the upper bound implied by the research.
Conservative
Voter was weakly persuaded by AI; quick recovery via additional contact
Base
Midpoint; consistent with research on persistence of AI-induced attitude change
Aggressive
Upper bound from persuasion research; extended AI conversation; entrenched reversal
How These Inputs Combine

Value at risk = (persuasion spend ÷ cost per persuasion) × post-contact search rate × AI encounter rate × unfavorable answer rate × reversal rate × recovery multiplier × cost per persuasion

Simplified: Value at risk ≈ persuasion spend × (post-contact search rate × AI encounter rate × unfavorable answer rate × reversal rate × recovery multiplier)

At base values: $1.5M × (0.22 × 0.60 × 0.20 × 0.35 × 3) = $1.5M × 0.028 ≈ $42,000 base-case value at risk for a congressional race. The wide reported range ($16K–$168K) reflects the full conservative-to-aggressive sweep across all inputs simultaneously.

B · Scenario Table

Value at risk as the
key unknown moves.

The unfavorable answer rate is the single input that campaigns know the least about and that matters most. This table holds all other inputs at base values and varies only the unfavorable answer rate. It shows what the model produces across the four most relevant scenarios.

All other inputs held at base: post-contact search rate 22%, AI encounter rate 60%, reversal rate 35%, recovery multiplier 3×, cost per persuasion $570.

Congressional Race ($1.5M persuasion spend)
Unfavorable Answer Rate Reversals (Base Model) Value at Risk As % of Persuasion Spend Interpretation
10%Well-documented candidate 12 reversals $21K 1.4% Diagnostic break-even at 3 reversals — program pays for itself even if it catches a quarter of modeled exposure
20%Base estimate (unknown) 24 reversals $42K 2.8% Meaningful — $42K at risk against a $5,500 diagnostic that breaks even at 3 prevented reversals
35%Poorly documented candidate 43 reversals $73K 4.9% Material — program cost ($5,500–$29,000) is well below value at risk at any scenario
50%Active opposition; sparse record 61 reversals $104K 6.9% Significant — full program at $29,000 is less than 30% of value at risk

Formula: reversals = ($1.5M ÷ $570) × 0.22 × 0.60 × unfav rate × 0.35. Value at risk = reversals × $570 × 3× recovery multiplier. All inputs other than unfavorable answer rate held at base. At 20%: 2,632 × 0.22 × 0.60 × 0.20 × 0.35 = 24 reversals × $1,710 = $42,000.

Senate / Statewide Race ($8M persuasion spend)
Unfavorable Answer Rate Reversals (Base Model) Value at Risk As % of Persuasion Spend Interpretation
10% 65 reversals $111K 1.4% Measurable even for a well-documented candidate; diagnostic still breaks even at 3–9 reversals
20% 130 reversals $222K 2.8% Over $200K at risk against a full program cost of $33,000–$80,000
35% 227 reversals $388K 4.9% Material to any race budget; program cost is under 20% of value at risk
50% 324 reversals $554K 6.9% Over half a million at risk; full Senate program at $80,000 is under 15% of value at risk

Same methodology scaled to $8M persuasion spend and $570 cost per persuasion.

On the 50% Scenario

The 50% unfavorable answer rate is not a prediction — it is the model's stress test. It represents a candidate with an actively contested online record, sparse structured information, and meaningful opposition research targeting AI-indexed sources. Whether any candidate is in this situation is an empirical question — and the diagnostic is the instrument that answers it. Presenting the 50% case is not a claim that this is typical; it is transparency about what the model produces under adverse conditions.

C · Break-Even Table

How many reversals
must a program prevent
to justify its cost?

Break-even is the simplest test: given the cost of a diagnostic, remediation, or monitoring program, how many persuasion reversals does it need to prevent to pay for itself? At $570 per persuasion and a 3× recovery multiplier, each prevented reversal avoids $1,710 in recovery cost.

Break-Even by Program Type
Program Congressional Budget Range Reversals Needed to Break Even Senate Budget Range Reversals Needed to Break Even
DiagnosticOne-time; establishes baseline $3,000–$8,000 2–5 reversals $6,000–$15,000 4–9 reversals
RemediationConditioned on diagnostic findings $8,000–$15,000 5–9 reversals $15,000–$35,000 9–21 reversals
Ongoing monitoringPer month, active race period $1,000–$3,000/month <1–2 reversals/month $2,000–$5,000/month 1–3 reversals/month
Full cycleDiagnostic + remediation + 6 months monitoring $17,000–$41,000 10–24 reversals $33,000–$80,000 19–47 reversals

Break-even: program cost ÷ ($570 × 3× recovery multiplier) = program cost ÷ $1,710. The break-even reversal count is fixed — it does not depend on the race's unfavorable answer rate. What changes with the unfavorable rate is how many total reversals are modeled, and therefore what fraction the break-even represents.

Break-Even in Context
Program Break-Even Reversals (Congressional) As % of Modeled Reversals at 20% Unfav. Rate
Diagnostic ($5,500 midpoint) 4 reversals 13% of 24 modeled reversals
Remediation ($11,500 midpoint) 7 reversals 28% of 24 modeled reversals
6-month monitoring ($12,000 midpoint) 7 reversals 29% of 24 modeled reversals
Full cycle ($29,000 midpoint) 17 reversals 71% of 24 modeled reversals

At base inputs (20% unfavorable rate, congressional), the model produces 24 reversals. The diagnostic breaks even at 4 — preventing 13% of modeled exposure. Note that the full cycle's 71% figure means the program needs to address most of the modeled problem to break even at this unfavorable rate; at 35% or 50% unfavorable, the same $29,000 full cycle breaks even at a much smaller fraction of modeled exposure.

The Diagnostic as Decision Gate

The diagnostic's break-even is 3–4 prevented reversals regardless of the unfavorable rate — because the break-even is a fixed cost divided by a fixed recovery unit. What changes with the unfavorable rate is how many reversals are modeled in total. At 20% unfavorable: 24 modeled, break-even at 4 (13%). At 5% unfavorable: 6 modeled, break-even at 4 (53%). The diagnostic pays for itself even when the problem is small — and it tells you whether the problem is small or large. That is why it is the right entry point regardless of what the campaign suspects.

D · Upside Model

The arithmetic of
favorable answers.

The defensive model asks: what does a campaign lose when AI answers unfavorably? The offensive model asks the inverse: what is the value of AI answers that work in the campaign's favor — including for voters the campaign never contacted directly?

A competitive congressional race generates an estimated 120,000–210,000 organic AI encounters, including from voters the campaign has not contacted. If those encounters are favorable and accurate, they are doing persuasion work the campaign did not pay for.

The Upside Arithmetic
Input Value Type
Total organic AI encountersCongressional midpoint estimate 150,000 Estimated
Favorable answer rate improvementFrom baseline to post-remediation +10 percentage points Illustrative — not a guaranteed outcome
Additional favorable encounters 15,000 Derived (150K × 10%)
Durable persuasion rateFraction of favorable encounters producing a lasting opinion shift 2% Estimated — consistent with political ad persuasion benchmarks
Additional durable persuasions 300 voters Derived (15K × 2%)
Equivalent advertising costAt $570 per durable persuasion $171,000 Derived (300 × $570)
Against a $10,000 remediation investment 17:1 illustrative return Before defensive value
This Is Illustrative Arithmetic, Not a Guaranteed Return

The upside model shows the structural scale of the opportunity, not a specific performance promise. AI systems are black boxes — their responses to specific candidates cannot be controlled or predicted. What the research establishes is an association: the most persuasive AI chatbots drew on accurate, fact-rich information, and less accurate chatbots were less persuasive. Building a more accurate, well-structured information record appears associated with better AI representation. Whether any specific campaign sees a 10-percentage-point improvement in favorable answer rate depends on factors outside the campaign's control. The 17:1 figure uses the inputs most favorable to the argument; substitute your own numbers to test it.

Sensitivity of the Upside Model
Favorable Rate Improvement Additional Favorable Encounters Additional Persuasions (at 2%) Equivalent Ad Spend
+5 pp 7,500 150 $85,500
+10 pp (base) 15,000 300 $171,000
+15 pp 22,500 450 $256,500
+20 pp 30,000 600 $342,000

Based on 150,000 base organic AI encounters, 2% durable persuasion rate, $570 per persuasion. The 2% persuasion rate is deliberately conservative — consistent with political digital ad benchmarks, which are generally lower than door-knock rates.

E · Sensitivity Analysis

Which inputs move
the answer most.

Not all inputs are equally consequential. A model is only as useful as your understanding of which assumptions dominate the output. This section ranks the inputs by their effect on the value at risk calculation and notes which the model is surprisingly robust to.

Input Sensitivity Ranking

The following rankings are based on the relative effect of moving each input from its conservative to aggressive estimate, holding all other inputs at base. A high-sensitivity input can swing the output by 3× or more; a low-sensitivity input moves it by less than 30%.

1. Unfavorable Answer Rate — Dominant driver
SensitivityVery high

Moving from 10% to 50% unfavorable answer rate swings value at risk by 5×. This is the number campaigns do not know without a diagnostic. It is also the number most likely to vary dramatically across candidates — a well-documented incumbent with decades of public record will sit at a very different point than a first-time challenger with sparse online presence. The diagnostic resolves this uncertainty entirely.

2. Recovery Multiplier — High sensitivity
SensitivityHigh

Moving from 2× to 4× recovery multiplier doubles the value at risk figure. This input is inferred from the Nature/Science persuasion research and is difficult to measure directly for organic AI encounters. Using 2× rather than 3× as the base produces a more conservative result; the conclusion (diagnostic cost is well below break-even) holds at any value in the range.

3. Post-Contact Search Rate — Moderate sensitivity
SensitivityModerate

Moving from 15% to 30% doubles this input, which flows linearly through the model. It adds 2× to the output at the extremes. Campaigns with older, lower-engagement electorates should use 15%; campaigns targeting younger, high-information voters should use 25–30%.

4. AI Encounter Rate — Moderate sensitivity
SensitivityModerate

The 50–70% range reflects genuine uncertainty in the Google AI Overview penetration rate across different query types and voter demographics. Because Google AI Overview is now the default surface for named-individual informational queries, this rate is unlikely to fall below 50% in any competitive federal race electorate. The model is moderately sensitive to this input — it adds roughly 40% to the output at the high end — but the floor is relatively well anchored.

5. Reversal Rate — Moderate sensitivity
SensitivityModerate

Moving from 25% to 50% doubles this input and flows linearly through the model. The 35% base estimate is deliberately conservative — not every voter who encounters an unfavorable AI answer is moved by it. The conclusion is stable across the full range of this input.

6. Cost per Persuasion — Low sensitivity to the break-even conclusion
SensitivityLow (to break-even)

Cost per persuasion appears in both the value at risk numerator and the break-even denominator, which means it largely cancels out in the break-even calculation. At $300 per persuasion, the value at risk is lower but so is the cost of recovery per voter — the ratio between them stays roughly constant. This is one of the reasons the break-even conclusion is robust: it does not depend on a specific cost per persuasion.

What Makes the Model Robust

The break-even conclusion — that the diagnostic cost is small relative to plausible value at risk — holds across nearly all combinations of inputs. The only scenario where it fails is a candidate with a genuinely clean, accurate, comprehensive AI record (unfavorable answer rate below ~5%) combined with a very low-engagement electorate (post-contact search rate below 10%). That is a case where the diagnostic itself confirms there is nothing to worry about — which is also valuable information. The diagnosis costs the same either way; what changes is what the campaign learns from it.

All model inputs and derivations are the author's. Base data sources: Gerber and Green field experiment syntheses for cost per persuasion; Edison Research Infinite Dial 2026 for AI usage rates; CCI 2024 post-election survey for dedicated AI tool usage; Google product documentation and industry analysis for AI Overview penetration; Nature (Lin et al.) and Science (Hackenburg et al.), December 4, 2025, for persuasion magnitude and recovery multiplier rationale; Pew Research 2024 and BPC 2022 for online verification behavior estimates. All estimated inputs are explicitly labeled as estimated or inferred in the input variable definitions above.

The author acknowledges a commercial interest in the subject of this paper. That interest is disclosed so readers can evaluate the analysis with appropriate context. The model is structured to be stress-tested; readers who disagree with specific inputs are encouraged to substitute their own values to test whether the conclusion changes.