01
The New Room
There is a conversation happening in political communications right now, and it has a vocabulary problem. On one side: technical people who have spent the last three years watching a new information channel emerge and who use terms like "entity disambiguation" and "training corpus" as casually as you use "earned media" and "persuasion universe." On the other side: experienced campaign professionals who built careers on the old map and who are very good at what they do — but who occasionally feel a flicker of exposure when the conversation turns to AI.
This paper is for the second group.
Not because they are behind — though the territory did change faster than anyone expected, and anyone who claims they saw all of it coming is probably selling something. But because the vocabulary of AI presence is not actually that complicated once you strip away the jargon, and because the professionals who can move between both worlds — who can take a briefing from a technical vendor and translate it into a campaign decision — are going to be worth considerably more than those who cannot.
Think of this as a phrasebook. You don't need to be fluent. You need enough words to ask good questions, understand the answers, and not give away that this is the first time you've heard any of it. The goal is not to fool anyone. The goal is to be genuinely useful in a room where this is the subject.
The short version of what changed: Voters used to form their first impressions of a candidate through media, mail, ads, and doors. Now a meaningful and growing share of them start with a question to an AI answer engine. The answer engine answers. That answer shapes the impression. Nobody bought that answer. Nobody approved that message. Nobody may even know it happened. That is the new room.
The rest of this paper gives you the vocabulary to operate in it.
02
How AI "Knows" Things
Before you can influence what an AI answer engine says about your candidate, you need a basic model of where its knowledge comes from. It is not magic, and it is not a search engine. The mechanics matter because they explain why certain strategies work and others do not.
Large Language Model (LLM)
The engine under the hood of every AI answer engine
ChatGPT, Gemini, Perplexity, Copilot — all of these are built on large language models. An LLM is a system trained on enormous amounts of text. It learned the patterns of language, the relationships between ideas, and a great deal of factual content by reading — in a loose sense of the word — a significant portion of the internet, along with books, articles, and other sources. When you ask it a question, it is not searching the web in real time. It is drawing on what it absorbed during training. The practical implication: what an AI says about your candidate reflects what was written about your candidate before a certain date, weighted by how much of it existed and how authoritative it appeared.
Training Data
The text the AI read before it started talking
Training data is the body of text — web pages, articles, books, databases — that an LLM was exposed to during its development. If your candidate was written about extensively and accurately before the training cutoff, the AI has a richer, more accurate picture of them. If your candidate was not widely covered, or was covered inaccurately, or confused with someone who shares their name — that is what the AI learned. The training data is the source. Everything downstream flows from it.
Knowledge Cutoff
The date the AI stopped reading
Every LLM has a knowledge cutoff — a point in time after which it has no information. A model with a cutoff of early 2024 does not know about anything that happened after that date unless it has a retrieval system layered on top. For campaigns, this creates a specific problem: a candidate who filed last month, pivoted on an issue, or changed their party affiliation may not be reflected accurately in an AI's response — not because the AI is wrong about what it knows, but because what it knows is old. This is one of the most common sources of AI misinformation about candidates, and it is entirely addressable.
Crawling
How AI systems go out and collect new information
Some AI systems — particularly those with real-time or near-real-time search capability, like Perplexity — supplement their base training with active web crawling. A crawler is an automated program that visits web pages, reads their content, and feeds that content back into the system. Pages that are well-structured, clearly written, and authoritative are more likely to be read and used. Pages that are thin, ambiguous, or poorly formatted are more likely to be skipped or discounted. Publishing the right content in the right format is how you get new information into a crawling AI system.
Retrieval-Augmented Generation (RAG)
When AI looks something up before answering
RAG is a technique where an AI system, before generating a response, goes and retrieves current information from the web or a database. Perplexity uses this heavily — you can see the sources it cited in real time. Google AI Overviews also use retrieval. The practical implication: for RAG-based systems, freshness and source quality matter enormously. A well-structured, recently published page about your candidate can influence what Perplexity says today. For pure LLM responses, you are working with older training data and the levers are different.
The key insight: AI does not have opinions about your candidate. It has patterns — built from what was written, when it was written, and how much of it existed. Changing what AI says means changing the patterns. That starts with changing what exists to be read.
03
How AI Thinks About People
This is the section most campaign professionals find both surprising and immediately actionable. AI systems do not process information about people the way a human reader does. They build structured representations — and those representations can be correct, incomplete, or catastrophically wrong. Understanding how they work tells you exactly where to intervene.
Entity
A person, place, or thing that AI treats as a distinct, named thing in the world
In AI terms, an entity is a discrete, identifiable thing — a person, an organization, a place, a concept. "Seth Bodnar" is an entity. "The University of Montana" is an entity. When an AI builds its understanding of the world, it organizes knowledge around entities and the relationships between them. Your candidate is an entity. The question is whether the AI has a clear, accurate, well-populated representation of that entity — or a confused, thin, or conflated one.
Entity Disambiguation
The process of sorting out which person is which when a name belongs to more than one person
If your candidate shares a name with a city councilman in another state, a retired professor, or a minor local official, AI systems may conflate them. Disambiguation is the process — both technical and content-based — of making clear to AI systems which entity is which. It involves publishing structured, authoritative content that clearly identifies your candidate's unique attributes: full name, role, geography, biography. Without deliberate disambiguation, AI is making its best guess — and its best guess may be wrong in ways that are not immediately obvious.
Identity Collapse
When AI describes the wrong person in answer to a question about your candidate
Identity collapse is the most severe form of entity confusion. It occurs when an AI system, asked directly about your candidate, responds by describing a different person with the same name. The response is not hedged — the AI presents fabricated or misattributed information with confidence. In a campaign context, this can mean a voter asking "Is [Candidate] a Democrat?" and receiving a confident wrong answer drawn from a different person's record. Identity collapse is measurable, addressable, and genuinely dangerous. It is also more common than most campaign professionals realize, particularly for candidates who filed recently or who have a common name.
Ghost Candidate
A candidate who effectively does not exist in AI
A ghost candidate is one for whom AI has so little information that it either cannot answer questions about them at all, or answers with so many hedges and gaps that the impression left is one of thinness and obscurity. "I don't have detailed information about this candidate" is the AI equivalent of a blank yard sign. In a field where voters increasingly start with an AI query, invisibility is not neutrality. It is a deficit.
Knowledge Graph
AI's internal map of how people, places, and facts connect to each other
A knowledge graph is a structured representation of entities and the relationships between them. Think of it as a web of facts: this person attended this institution, held this role, is affiliated with this party, is associated with these issues. AI systems use knowledge graphs — both internal ones built during training and external ones like Wikidata — to answer questions about people. The more populated and accurate a candidate's position in the knowledge graph, the more consistent and correct AI responses tend to be. An empty or conflated node produces inconsistent, unreliable answers.
04
The Surfaces
One of the concepts that trips up communications professionals most reliably is the idea that "AI" is one thing. It is not. There are multiple distinct AI platforms, each with different architectures, different training data, different update cycles, and different behaviors. A campaign that is doing well on Perplexity may be failing on ChatGPT. The vocabulary for navigating this starts with the concept of surfaces.
Surface
Each distinct AI platform where a voter might encounter a description of your candidate
A surface is any AI-powered system where a user can ask a question and receive a synthesized answer. The current major surfaces for political content are: ChatGPT, Perplexity, Google AI Overviews, Gemini, Microsoft Copilot, and Brave Search. Each has its own behavior. Perplexity retrieves in real time and shows its sources. ChatGPT draws primarily on training data and tends to be more confident — and more confidently wrong. Google AI Overviews appear at the top of standard search results — increasingly the first thing a voter sees before they click on anything. Treating these as a single channel is like treating television, radio, and newspapers as a single media buy.
AI Presence
The overall picture of your candidate that emerges across AI surfaces
AI presence is the aggregate of what AI platforms say about a candidate when asked. A candidate with strong AI presence is described accurately, consistently, and with sufficient detail across the major surfaces. A candidate with weak AI presence is described inconsistently, inaccurately, or not at all. AI presence is measurable — you can query the surfaces systematically, score the responses, and produce a baseline that you can improve against. That baseline measurement is the starting point for all active management of AI presence.
Citation
When an AI response names a specific source for what it just said
Some AI systems — particularly Perplexity and Google AI Overviews — attach citations to their responses. A citation is a named source that the AI used to construct its answer. Citations matter for two reasons. First, they indicate which sources AI considers authoritative for your candidate — useful intelligence. Second, they signal that the information came from somewhere current and retrievable, which is generally more favorable than information drawn from older training data with no source attached. High citation counts on factual queries about your candidate are a positive signal.
A note on pace: The surface landscape is changing. New AI-powered search experiences are being introduced by Google, Microsoft, and others at a pace that outstrips any fixed list. The vocabulary here — surfaces, citations, AI presence — is designed to remain useful as specific platforms evolve. The question "what does this surface say about my candidate?" is the right question regardless of which surface is currently dominant.
05
The Signals
Once you understand how AI builds its picture of a candidate, the next question is: what can you actually do about it? The answer involves a set of content and technical tools that are collectively called signals. But not all signals are created equal — and more importantly, not all of them are yours to control.
Before getting into specific tools, it helps to sort signals into three categories. Where a signal falls shapes your strategy, your timeline, and your expectations about what is possible.
Signals you control
Structured data on your campaign website. The llms.txt file. Your candidate biography. FAQ markup. Press kit copy. These are fully within your authority. You publish them, you write them, you update them. They are also the signals AI systems can most easily attribute to the campaign itself — which means they are useful, but they carry the credibility of a primary source that is also a self-interested party. AI systems appear to treat them as one input among many.
Signals you have some influence over
Earned media — news coverage, profiles, interviews, op-eds — falls here. You pitch it, reporters decide whether to write it, editors decide whether to run it, and the publication decides how to frame it. You do not control the output. But you can influence the volume and the framing through sustained press strategy. Coalition statements, endorsement content, and third-party surrogates also fall here: you cultivate those relationships, but you do not write the copy. Wikipedia is a partial exception — you can submit edits, but volunteer editors make the final call.
Signals others control
Opponent content. Negative coverage in news archives. Ballotpedia entries written and maintained by others. Commentary from advocacy groups. These signals exist whether you engage with them or not. You cannot delete them or rewrite them. The only tools available are volume (drowning a bad signal with better signals) and accuracy (submitting corrections where correction mechanisms exist).
The hard truth about technical markup: Structured data, llms.txt, and corrective copy are valuable — and they are necessary. But a campaign website with technically perfect AI markup and no press coverage will still have thin AI presence. The AI information environment is built primarily from third-party sources: news coverage, organizational endorsements, Wikipedia, and the accumulated record that reporters and researchers have created over time. Technical signals from the campaign itself are one input. Earned media tends to be a far larger one. Sometimes the most effective AI presence intervention starts not with a developer, but with a press strategy.
Earned Media
News coverage, profiles, and third-party reporting that AI systems treat as authoritative source material
Earned media is the traditional PR term for coverage you did not pay for — news stories, interviews, profiles, op-eds, and mentions in credible publications. In the context of AI presence, earned media plays a structural role that paid and owned media cannot replicate. AI systems appear to weight third-party sources — news publications, civic databases, established organizations — more heavily than content published by the campaign itself, for the same reason a citation from a neutral analyst carries more weight than a company's own press release. A strong earned media record does more to shape AI's picture of a candidate than almost any technical optimization. Campaigns that treat AI presence as a purely technical problem — and skip the press work — tend to find that the AI environment reflects the coverage gap accurately.
Structured Data
Machine-readable labels embedded in a web page that tell AI exactly what it is looking at
A regular web page contains text that humans read and infer meaning from. Structured data is a layer of markup — invisible to casual readers — that tells machines explicitly what the content represents. A structured data block on a campaign website might say, in machine language: "This is a Person. Their name is [X]. Their political affiliation is Independent. Their current role is Senate Candidate. Their prior role was University President." AI crawlers and search engines read this markup directly. It is one of the clearest and most authoritative signals you can send. The standard vocabulary for structured data in political contexts comes from Schema.org, a shared framework used by Google, Bing, and others.
What it actually looks like: Structured data is expressed in a format called JSON-LD — a compact block embedded in a page's code. The same information appears twice: once for humans, once for machines.
What the page says (prose)
Seth Bodnar is an independent candidate for U.S. Senate in Montana. He previously served as President of the University of Montana.
What the structured data says (JSON-LD)
{
"@type": "Person",
"name": "Seth Bodnar",
"jobTitle": "U.S. Senate Candidate",
"affiliation": "Independent",
"description": "Former President,
University of Montana"
}
The prose version is written for persuasion. The JSON-LD version is written for machines. Both live on the same page. When AI systems crawl the page, they read the machine version — and the explicit labels reduce the chance that AI misclassifies the candidate, conflates them with someone else, or gets the party affiliation wrong.
llms.txt
A file on a website that speaks directly to AI systems about who the candidate is
An llms.txt file is a plain-text document, placed at the root of a campaign website, written specifically for AI language models rather than human readers. Think of it as a briefing document addressed to the AI: "Here is who this person is. Here is their role. Here is what is accurate and what may be confused with other people." It is a relatively new convention — an emerging standard, not yet universal — but several major AI companies have indicated they attend to it. In a field where every signal matters, it is inexpensive to produce and worth having.
Corrective Copy
Content written specifically to address and replace inaccurate AI narratives
Corrective copy is not a press release or a campaign bio. It is content engineered to correct a specific inaccuracy in AI-generated descriptions of a candidate. If AI platforms are consistently calling your candidate a Democrat when they filed as an Independent, corrective copy makes the filing date, the party status, and the reason for the independent run explicit, factual, and prominent — in forms that AI systems can find, read, and integrate. It comes in two registers: a factual version for AI-facing content (structured data, llms.txt, FAQ markup) and a campaign-voice version for human-facing content (the candidate biography, the website About page).
Framing
The narrative context in which a candidate's record is presented — and which AI tends to absorb and repeat
AI systems do not just repeat facts. They absorb narrative patterns. If the bulk of coverage about a candidate frames them as "a Democratic-adjacent independent" or "a longshot challenger," those framings tend to appear in AI-generated summaries even when they are reductive or outdated. Deliberate framing — publishing content that consistently uses a specific phrase, narrative arc, or identity signal — is how you shape the pattern over time. This is not spin. It is the recognition that answer engines tend to repeat what they find, and that what it finds is shaped by what you publish.
06
This Is Not SEO
If you have been in digital communications for any length of time, you have navigated the SEO conversation. The instinct — entirely reasonable — is to treat AI optimization as a new version of SEO. It is not. The differences are specific and they matter for how you think about the work.
| The question |
SEO |
AI Presence |
| What are you influencing? |
A ranked list of links |
A synthesized answer |
| What does the voter see? |
Options — they choose what to click |
A conclusion — AI has already done the synthesis |
| How many dominant surfaces? |
Primarily one (Google) |
Seven or more, each behaving differently |
| Where does failure show up? |
Page 2 — low visibility |
Confident misinformation — active damage |
| What does success look like? |
High ranking for relevant queries |
Accurate, consistent, detailed answers across surfaces |
| Can you see how it works? |
Mostly no — also a black box |
No — output is measurable, mechanism is opaque |
The critical difference is in what voters receive. In search, a voter gets a list and makes choices. The agency is theirs. In AI, a voter gets an answer. The AI has already done the synthesis and made the call. If that answer is wrong — if it calls your candidate a member of the wrong party, or describes a different person entirely — the voter does not see the error. They see a confident statement from a source they perceive as authoritative and objective.
The practical implication: SEO failure means being hard to find. AI failure means being actively misrepresented. They are not the same problem, they do not have the same solutions, and they should not be managed by the same strategy or measured by the same metrics.
Go deeper → For the full strategic and financial case — who else is querying AI about your candidate and why it matters beyond voters — see
The Channel You're Not Buying.
07
The Black Box
Here is where intellectual honesty becomes a professional asset. Anyone who tells you they know exactly why an AI said something, or can promise what it will say after a specific intervention, is either mistaken or overselling. AI systems are, in the technical phrase, black boxes.
Black Box
A system whose internal workings you can observe the inputs and outputs of, but not the mechanism in between
You can query an AI and see what it says. You can publish content and observe whether the AI response changes over time. What you cannot do — with any current AI system — is inspect the internal process that produced a given output and say with certainty: "this content caused this response." The relationship between inputs and outputs appears to exist and appears to be influenceable. But it is probabilistic, not deterministic. A vendor who promises "publish this structured data and your AI score will go from 43 to 82 within six weeks" is not working from evidence. Be skeptical of anyone who claims certainty about mechanisms that are not publicly documented.
The box gets tuned. The engineers behind these models make internal adjustments frequently — changing how information is weighted, how sources are prioritized, how entities are resolved — much the way Google has always made internal changes to how its ranking algorithms work. Those changes are not announced. They are not documented. The only way to detect that something has shifted is to measure what the AI is saying, notice that it changed, and respond. We are early in the development of these systems. Between now and the 2026 election, it is reasonable to expect that the major platforms will make significant internal changes — some of which will affect what AI says about your candidate without any action on your part. This is another reason the work is a cycle, not a project. The box keeps moving.
AI Presence Score (APS)
A measured baseline of how accurately and completely AI describes a candidate today
An AI Presence Score is a structured measurement of AI outputs — queries run across surfaces, responses scored for factual accuracy, entity correctness, narrative consistency, and citation quality. It establishes a baseline: this is what AI says about this candidate right now, at this moment in time, before any intervention. The score by itself tells you where the problems are. What it cannot tell you — until after you intervene and re-measure — is whether your interventions changed anything. The before-and-after comparison is the only honest measure of effectiveness in this field.
A well-designed presence score aggregates multiple underlying dimensions into a single number — so that a campaign team can track direction without reading a technical report every cycle. The useful question the score answers is not "what is our number?" but "is it moving?"
The honest position — and the credible one: "We measured where we are. We implemented specific interventions. We will re-measure. We will know what changed." That is a defensible, professional approach to AI presence work. It is also, in a field full of overclaiming, fairly rare — which makes it a competitive differentiator for the professionals who adopt it.
08
The Work Cycle
AI presence work is not a one-time deliverable. It is a cycle — because AI surfaces update, because campaigns evolve, because new information enters the information ecosystem and displaces old information. Understanding the shape of the cycle matters because it is how you scope, budget, and evaluate the work.
The Baseline Measurement
What does AI say about your candidate right now?
The first phase is measurement. A defined set of queries — identity queries, policy queries, competitive queries, biographical queries — is run across the major AI platforms and each response is scored. The output is a report showing which surfaces are performing well, which are failing, where the specific inaccuracies are, what the ghost risk looks like, and which findings are critical versus manageable. The measurement tells you where you are. It does not fix anything. It makes the decision about what to fix an informed one rather than a guess.
The Intervention
What you do about what the measurement found
The second phase is response. It has two parts. The first is strategic: the campaign team reviews the findings, makes decisions about which problems to address and how aggressively, and designates who has final approval. The second is operational: a digital director works through the approved actions — publishing corrective copy, adding structured data, updating reference pages, submitting Wikipedia and Ballotpedia corrections, publishing issue content. Intervention does not guarantee changed AI behavior. It publishes the signals. The re-measurement phase determines whether the signals landed.
The Re-measurement
Did anything change?
The third phase re-runs the same measurement methodology after interventions have been in place long enough for AI systems to potentially have incorporated them. It answers the before-and-after question. It also surfaces new issues — because the information landscape continues to evolve, because opponents produce content, because events happen. The re-measurement is not a victory lap. It is the next baseline. The cycle then repeats: new findings, new decisions, new interventions, new measurement. This is not unlike the ongoing work of any communication function. It is just operating on a channel that most campaigns have not been managing.
The standing team: Because this cycle repeats, the most effective campaigns treat AI presence management as an ongoing function with a standing review team, not a one-time project. The team that reviews findings and makes decisions in cycle one is the same team in cycle two and cycle three. The vocabulary, the process, and the decision-making get faster with repetition.
09
Quick Reference
The page to dog-ear. One sentence per term.
| Term |
Plain English |
| Large Language Model (LLM) | The AI engine trained on massive amounts of text — the technology behind ChatGPT, Gemini, and the rest. |
| Training Data | The text an AI read during development — the source of everything it "knows." |
| Knowledge Cutoff | The date after which an AI has no information — anything that happened after this date may not be reflected. |
| Crawling | The automated process by which AI systems collect new content from the web. |
| RAG (Retrieval-Augmented Generation) | When an AI looks up current information before answering — Perplexity does this heavily. |
| Entity | A distinct, named person, place, or thing in AI's internal model of the world. |
| Entity Disambiguation | Making clear to AI which person is which when a name is shared by more than one individual. |
| Identity Collapse | When AI confidently describes the wrong person in response to a question about your candidate. |
| Ghost Candidate | A candidate who is invisible or nearly invisible to AI — present enough to be asked about, not present enough to be described accurately. |
| Knowledge Graph | AI's structured map of how entities connect — who holds what role, where, affiliated with what. |
| Surface | Each distinct AI platform where a voter might ask about your candidate — ChatGPT, Perplexity, Gemini, Copilot, Google AI Overviews, Brave. |
| AI Presence | The overall picture of your candidate that emerges when you query AI platforms systematically. |
| Citation | A source named by an AI in its response — signals currency and retrievability. |
| Structured Data | Machine-readable labels embedded in a web page that tell AI exactly what it is looking at. |
| llms.txt | A plain-text file on a campaign website written specifically to brief AI crawlers on who the candidate is. |
| Corrective Copy | Content engineered to replace a specific inaccuracy in AI-generated descriptions. |
| Framing | The narrative context in which a candidate is consistently described — AI absorbs and repeats it. |
| Black Box | A system whose outputs you can observe and measure but whose internal mechanism you cannot inspect. |
| AI Presence Score (APS) | A single aggregated number tracking how accurately and consistently AI describes a candidate — useful for direction, not precision. |
| Baseline Measurement | Phase one of the work cycle — systematic queries across AI surfaces, scored and reported before any intervention. |
| Intervention | Phase two — strategic decisions about what to fix, followed by operational execution: corrective copy, structured data, reference updates. |
| Re-measurement | Phase three — re-running the same methodology after interventions to answer: did anything change? |