A strategic brief for the organizations that publish the record, the candidates who are described by it, and the officeholders who cannot edit it — in the era when voters ask AI who to trust, and AI answers.
When a voter asks ChatGPT whether their senator is good on the environment, something specific happens. Depending on the system, the AI assembles an answer out of sources it retrieves at query time — Wikipedia, Ballotpedia, press coverage, and material published by established organizations — selected by retrieval logic the platforms do not publicly document.1 If the League of Conservation Voters has given that senator a lifetime score of 94%, that number may appear in the answer. If Emily's List has endorsed her, that endorsement may be cited. If the NEA has given her an A rating, that rating may shape how the AI characterizes her relationship with educators and public schools. Whether any of that actually happens for a given senator on a given platform is measurable, and measuring it is what an audit is for. It is not something to assume.
The reason to care about being retrievable is that retrieval is where these systems fail. In a fourteen-day evaluation of six commercial answer engines against same-day BBC reporting, retrieval failures — landing on the wrong source in the first place — accounted for more than 70% of all errors; when a model retrieved a correct source, it usually extracted the correct answer from it.2 Organizations that have built decades-long track records of rigorous, methodology-backed scoring and endorsement work — and that publish that work in accessible, indexed form — are in principle the kind of source such a system can land on. Whether any specific organization's content is in practice reaching AI systems is a question that requires testing, not assumption.
"Your scorecard is not just a member communication. It is a primary source document that AI systems consult when voters ask who to trust."
Most progressive organizations have not understood this yet. They have continued treating their digital presence as a member communication and donor engagement platform — which it must remain — without recognizing that it has simultaneously become the upstream document that shapes AI outputs for millions of voters who will never visit the site, never read the scorecard, and never see the endorsement press release. They encounter the conclusion the AI drew from it.
Consider what the following organizations already produce — and how AI systems evaluate it:
Annual scorecards rating every member of Congress on environmental votes. Lifetime scores. Detailed methodology. Decades of consistent data. This is structured, authoritative, verifiable data that AI systems tend to weight heavily when answering questions about any legislator's environmental record.
Candidate endorsements with documented rationale. Fundraising support that creates a paper trail. A 40-year record of identifying and backing winning candidates. An Emily's List endorsement is a credibility signal AI systems recognize — and cite when voters ask about a candidate's reproductive rights record.
Legislative ratings, endorsements, and issue positions on education policy with the credibility of the nation's largest professional associations. When voters ask AI about a candidate's education record, union ratings and endorsements are among the most authoritative signals in the ecosystem.
Legislative scorecards, endorsements, and issue research backed by millions of members. Labor union ratings are among the most consistently structured political data in the information ecosystem — exactly what AI systems look for when synthesizing answers about labor policy and worker protection.
Every one of these organizations is already producing the raw material that AI systems value. The gap — and it can be a significant one — is that most are not yet producing it with the deliberate structure, formatting, and maintenance discipline that maximizes its visibility to AI retrieval systems. Even the most digitally sophisticated organizations in this space have generally not treated AI optimization as a distinct communications objective. The underlying assets are strong. The AI-facing architecture typically is not.
The credibility that comes from decades of rigorous, methodology-backed scoring and endorsement work is a genuine asset in the AI information environment, but only if the content carrying it can be retrieved at all. Scorecards that exist primarily as PDFs, endorsement archives reachable only through navigation designed for human readers, and research published without any summary a machine can parse are harder for a retrieval system to land on than the same content published as indexed web pages. Google states that no special markup and no machine-readable file is required to appear in AI Overviews or AI Mode, and also that a page must be indexed and eligible for an ordinary Search snippet before it is eligible for those features at all.3 On that evidence the floor is indexability, not schema, and whether a given organization clears it is something only measurement establishes.
When your scorecard or endorsement appears in an AI answer, the person reading that answer is not always a voter. The same AI information environment that shapes how voters understand candidates also shapes how every professional researcher in the political ecosystem understands them — and those professionals have outsized influence on campaign outcomes.
The same holds up and down that ecosystem. A beat reporter who asks ChatGPT about a state senate challenger and finds your endorsement, with the rationale and the legislative record attached, arrives at the interview with a more accurate frame. Peer funders and portfolio analysts run the same quick due diligence, and an endorsed candidate carrying a known scorecard reads as investment-grade rather than as a risk. Major donors ask what voters ask. Party staff deciding which races to resource sit in the same information environment.
Properly structured, the asset reaches everyone whose decisions shape whether an endorsed candidate wins, and only when AI can find it.
Everything that follows depends on which of three things you are, because the three face different problems and the remedies do not transfer. Kyanos encodes the distinction in code rather than in prose: the audit pipeline refuses to run without knowing whether its subject is an organization, and throws rather than assuming, because a defaulted noun mislabels half the subjects it describes.5 Three separate schema vocabularies exist for the same reason.6
You publish the evidence other people are judged by — scorecards, endorsements, ratings, research. When an answer engine characterizes a senator's environmental record, it may be reaching for your number. Your schema captures legislative scorecards, organizational endorsements, member constituency and funding relationships.
Your problem is being retrievable. The record exists and you control it. Whether a retrieval system can find and parse it is the open question, and it is measurable.
The record about you is mostly written by other people: Ballotpedia, local press, opponents, and the organizations above. Your schema captures a policy platform, endorsements received, campaign background and comparison against opponents.
Your problem is that a record may barely exist. A first-time challenger can have almost nothing an answer engine can retrieve, and absence is not neutral — it is answered from whatever else is available.
Votes, committee assignments, sponsored legislation and constituent service already exist in government databases whether or not you ever publish them. Your schema captures voting records, legislative achievements, committee roles and official statements.
Your problem is interpretation. The facts are retrievable already. What an answer engine makes of a vote — cast years ago, in a context the record does not carry — is the exposure.
"An organization asks whether AI can find what it published. A candidate asks whether there is anything to find. An officeholder asks what AI concluded from a record they cannot edit."
These are not three intensities of one problem. They are three problems, and the work each implies is different in kind. An organization invests in making an existing archive machine-legible. A candidate invests in creating a record where none exists — which is closer to publishing than to optimization. An officeholder invests in context around facts already in circulation, because the facts themselves cannot be withdrawn.
A sitting officeholder running for re-election is both a public record and a candidate. An organization that endorses becomes part of a candidate's record. Kyanos's own vocabulary treats entity kind as what something is and everything it is doing as a separate relationship, precisely because a single label cannot express an incumbent who is also a challenger for a different seat.7 Read the three cases below as roles that a single subject may occupy at once, not as a category to be sorted into.
Understanding how AI systems actually process organizational content requires letting go of the mental model of a reader — a person who visits your site, reads your materials, and forms a judgment. AI systems do not read in that sense. They parse. They evaluate structure, authority, consistency, and verifiability. What they reward is categorically different from what human communication has traditionally optimized for.
| Content Type | Why AI Systems Trust It | What Your Organization Needs to Do |
|---|---|---|
| Scorecards and ratings | Structured, numerical, methodology-backed, updated on a consistent schedule. AI systems can verify scores against voting records and find consistency across sources. | Publish with schema markup, machine-readable data formats, and consistent URL structure. Make the methodology explicit and linkable. Maintain historical scores in structured archives. |
| Endorsements | Named entity (organization) endorsing named entity (candidate) for specific office — exactly the kind of structured relationship data AI knowledge graphs are built to represent. | Publish endorsements with structured data identifying the office, the cycle, the rationale, and links to supporting evidence. Maintain a canonical endorsement archive that AI crawlers can access. |
| Issue research and policy positions | Authoritative organizations' stated positions on specific policy questions are weighted as expert opinion, especially when consistently maintained over time. | Structure issue positions as explicit, attributable claims with supporting citations. Avoid vague advocacy language — AI systems parse claims better when they are specific and verifiable. |
| Legislative analysis | Bill-by-bill analysis with named votes, specific provisions, and documented outcomes is the most traceable content in the political information ecosystem. | Publish legislative analysis in a format that connects votes to outcomes, names to positions, and claims to evidence. This is content AI systems can verify — and will trust more as a result. |
| Candidate comparisons | Side-by-side candidate evaluations on specific issues are exactly the format voters ask AI to produce — making your comparison the natural source for the AI's answer. | Structure comparisons as machine-readable tables, not narrative prose. Use consistent category labels across election cycles so historical comparisons remain useful. |
The digital content that most progressive organizations invest most heavily in — email campaigns, social media posts, event announcements, member stories, donation appeals — is largely invisible to AI information retrieval systems. This is not because it lacks value for member engagement; it has considerable value there. But it is not the content AI systems draw from when answering voter questions. Understanding the distinction is essential for allocating communications resources in the AI era.
The specific failure modes are:
Scorecards published as PDF downloads are among the most common failures. They are readable by humans who seek them out. They are largely invisible to AI retrieval systems that need web-accessible, structured text.
"Senator X has repeatedly betrayed working families" is a message. "Senator X voted against the PRO Act in 2021 (H.R. 842)" is a fact. AI systems can verify and cite the fact. They cannot do much with the message.
Referring to the same legislation as "the PRO Act," "the Protecting the Right to Organize Act," and "the workers' rights bill" across different pages creates semantic fragmentation that AI systems struggle to resolve into a coherent picture.
Content without a clear publication date, or content that looks current but reflects superseded information, gives a retrieval system no way to establish what is current. How each platform weights recency is not publicly documented, and we do not claim to know. The reason to date your content is that undated content withholds a signal, which is a cost you control.
Your organization's Wikipedia page — and the Wikipedia pages of the legislators you endorse, rate, or oppose — is in many cases more influential in AI outputs than your own website. Wikipedia's structure, citation requirements, and consistent formatting make it a preferred source for AI retrieval systems. This is not cause for despair — it is cause for strategy. Organizations that actively maintain accurate, well-cited Wikipedia pages for themselves and monitor the pages of key legislators are seeding the ecosystem more effectively than organizations that only optimize their own site.
The mirror image is more useful than the list of failures, because every item on it is something you can produce this quarter. The four properties below are what Kyanos scores answer-engine responses on, turned around to describe the material that produces them.8 None of this is a claim about what any platform does internally — that is not documented and we do not know it. It is a description of material that is retrievable and checkable, versus material that is neither.
| Property | What it looks like in practice | Who it matters most for |
|---|---|---|
| One name, everywhere | The same legislation called the same thing on every page — "the PRO Act (H.R. 842)" and not, elsewhere, "the workers' rights bill." The same for people: full name, role, jurisdiction, on first use. | Candidates most of all, since a name collision hands your record to somebody else |
| Positions stated, not implied | "Voted against H.R. 842 in 2021" rather than "has repeatedly betrayed working families." The message may be true; only the fact can be cited. | Officeholders, whose votes are already public and will be characterized by someone |
| Claims with an identifier attached | A bill number, a roll-call date, a score with its methodology beside it. A number a reader can chase is a number that can be quoted. | Organizations — this is what a scorecard already is, if it is published as text |
| Indexable web pages | An HTML page at a stable URL, dated, reachable without a search form. Google states plainly that no special markup is required to appear in its AI features — but a page must be indexed and eligible for an ordinary search snippet first.3 | Everyone. This is the floor, and the PDF scorecard is the most common way of falling through it. |
Read against the failure list, the pattern is plain. A PDF scorecard and an indexed scorecard page contain the same judgments; only one of them can be retrieved. "Betrayed working families" and "voted against H.R. 842" express the same position; only one of them can be verified. The difference in each pair is not effort or budget. It is a publishing decision made without the retrieval question in mind, usually years ago.
Take the single document your organization is best known for — the scorecard, the endorsement list, the annual research report. Check three things: is it an HTML page rather than a download; does it carry a visible date; and does it name each subject and each bill exactly once per page, the same way. If those three are true, it is retrievable material. If they are not, that is the highest-value fix available to you, and it is editorial work rather than engineering.
Issue advocacy has always been a competition over framing. Who defines the terms of a debate — who gets to say what a policy is about, what its effects are, who it helps and who it hurts — has enormous influence over how voters understand and respond to it. Progressive organizations have built sophisticated framing operations over decades. The question this moment raises is whether those operations are producing outputs in a form that AI systems can find, parse, and reproduce accurately — and whether fragmentation across organizations is undermining the coherence those systems reward.
The relevant concept is coherence — or its absence.
When a voter asks an AI "What are the effects of right-to-work laws?" the AI synthesizes an answer from multiple sources: academic research, think tank analyses, news coverage, advocacy organization positions, and legislative records. The answer it produces tends to reflect the coherence of the available information. When framing on an issue is fragmented — different organizations using different language, citing different data, making slightly different claims about the same facts — AI systems are more likely to respond with balanced, hedged, "some argue / others argue" constructions. Whether this is happening systematically on progressive issues is an open empirical question; what is not in question is that coherence across sources is a property AI retrieval systems are designed to reward.
When framing on any issue is highly consistent across many sources — the same factual anchors, the same terminology, corroborated by multiple authoritative outlets — AI systems are more likely to produce confident, direct answers that reflect it. The direction of this effect is not partisan: it rewards whoever achieves consistency, regardless of ideology.
"AI systems appear to reward coherence. Organizations whose messaging has historically emphasized consistency — regardless of ideology — have a structural advantage in the AI information environment."
This is not an argument that progressive messaging should abandon nuance or intellectual honesty — it is an argument that progressive organizations need to coordinate on the factual record with the same discipline that conservative organizations apply to messaging. The challenge is making the factual record these organizations have built coherent, consistent, and machine-readable across the ecosystem.
| Framing Element | Traditional Practice | AI-Era Practice |
|---|---|---|
| Terminology | Each organization uses its own preferred language for the same policy | Coordinated vocabulary: consistent naming for bills, policies, and provisions across the ecosystem — not identical rhetoric, but consistent factual anchors |
| Statistics | Each organization cites different data points, sometimes inconsistently | Shared citation of primary sources, with consistent interpretation — AI systems verify statistics against original sources; inconsistent numbers create uncertainty |
| Candidate characterization | Multiple organizations describe the same legislator's record differently | Aligned factual record: same votes cited, same scores used, same legislative history referenced — the framing differs, the facts don't |
| Opposition characterization | Advocacy language about opponents that AI systems can't verify | Documented voting records, specific votes, named legislation — verifiable claims that AI systems can cite rather than dismiss as advocacy |
| Research publication | Reports published for press and member audiences in narrative format | Research structured for both human readers and machine parsing — key findings as explicit, citable claims with primary source links |
Organizations that serve adjacent issue areas — environmental, labor, education, reproductive rights — have a structural opportunity that individual campaigns do not: they can coordinate on the factual record across issue areas without running afoul of the legal restrictions that prevent campaign coordination. An LCV scorecard and a SEIU legislative rating published about the same senator, using consistent factual anchors about the same votes, create a more coherent picture than either organization creates alone. The AI sees multiple authoritative sources agreeing — and responds with confidence rather than hedging.
This is not messaging coordination. It is factual alignment. The distinction matters legally and strategically.
The communications operation of a progressive membership organization has historically been built around two audiences: members (who need to be informed, engaged, and mobilized) and press (who need compelling stories, credible data, and accessible spokespeople). The AI era adds a third audience — AI retrieval systems — that has fundamentally different needs and that currently receives almost no deliberate attention.
Adding this third audience does not require abandoning the first two. It requires understanding that some of the work your communications team already does — research publications, legislative analyses, scorecards, endorsement announcements — can be restructured to serve all three audiences simultaneously, with modest additional effort.
Know what AI systems say about your organization when a journalist, donor, or voter asks. Run systematic queries. Document the baseline. Identify gaps and errors. This is now a communications function, even though it looks like a technical one.
When a voter asks AI about climate change legislation, workers' rights, reproductive healthcare, or public education funding, what does the AI say? Is your organization's framing reflected? Are the facts your research has established the ones the AI cites? These are communications questions with technical answers.
When voters ask AI about candidates your organization has endorsed, is the endorsement reflected accurately, and is your rationale — the specific votes, the specific record — present in the answer? Or is the characterization coming from sources that never saw your research? Two things follow, and both stay on your own side of the line: measure what the engines say, and publish your own rationale, in full, on your own surfaces. Your endorsement and the reasoning behind it are your organization's speech. Shaping a candidate's own material, or acting at their request, is a different activity that belongs with counsel before it belongs in a workplan.9
Ensure that your C3's research, your C4's positions, your PAC's endorsements, and your social media all describe the same factual record consistently. Incoherence is invisible to human audiences who encounter one channel at a time. It is highly visible to AI systems that encounter all channels simultaneously.
The most important shift in communications discipline is the orientation from content (producing material for human audiences on a publishing schedule) to evidence record (maintaining a structured, consistently updated, machine-readable archive of your organization's factual claims and the records that support them).
Most progressive organization communications teams do not currently have anyone whose job includes AI information environment management. The emerging roles map roughly onto existing titles but with expanded scope:
Digital Director or Digital Manager — must now include AI output monitoring as a standard function, alongside social media management and email. Running systematic queries across ChatGPT, Gemini, Perplexity, and Google AI Overviews for the organization's key issues and endorsed candidates should be a monthly workflow, not an occasional curiosity.
Research or Policy Staff — must publish findings in formats that serve AI retrieval, not just human readers. This means explicit, citable claims as well as narrative analysis; structured data alongside prose; machine-readable formats in addition to PDF reports.
Web/Content Staff — must understand schema markup and structured data well enough to implement it on key content pages: the scorecard, the endorsement database, the issue position library. This is not deep technical work — it is learnable by committed non-engineers — but it requires deliberate attention that current workflows do not include.
Every communications director currently measures web traffic, email open rates, press placements, and social engagement. Almost none of them systematically measure AI representation quality — how accurately and completely AI systems characterize their organization, their issues, and their endorsed candidates when voters ask. This metric does not yet have standard tooling or industry benchmarks. But it is, increasingly, the metric that corresponds most directly to the communications goal these organizations actually care about: ensuring that voters who ask about their issues and candidates get accurate, complete, well-framed answers. The measurement gap will close. Organizations that start tracking it now will have a head start when it does.
The 2026 election cycle is the first one where AI visibility will be a meaningful factor in competitive races. For organizations with scorecards, endorsements, and research that could shape AI outputs on behalf of candidates and issues they care about, eighteen months is enough time to make a material difference — if the work starts now.
Query ChatGPT, Gemini, Perplexity, and Google's AI Overview for your organization by name, for your key issue areas, and for the candidates you have endorsed or rated. Document exactly what the AI says. Note what is accurate, what is incomplete, and what is wrong or adversarially framed. This is your starting point. You cannot measure progress without it, and you cannot prioritize interventions without knowing where the gaps are.
If your scorecard exists primarily as a PDF download, this is the single highest-value technical change you can make. Publish it as indexed, structured HTML with schema markup identifying legislators, votes, scores, and methodology. Maintain it as a live archive, not a one-time publication. The scorecard is your most AI-visible asset — it needs to be in a format AI systems can actually read.
Every endorsement your organization has made — going back as far as records allow — should be published in a structured, machine-readable format: candidate name, office, cycle, rationale, and links to supporting vote records. This archive is the kind of consistent, longitudinal data that AI systems tend to weight heavily. It also makes your organization's track record visible in a way that builds institutional credibility over time.
For every significant research report going forward, produce a structured summary alongside the narrative report: key findings as explicit, numbered, citable claims with primary source links. This gives a retrieval system a parseable version of your findings without compromising the depth of the full report, and it costs little. What it does to what AI systems actually say about your issue areas is not something we can predict — it is something an audit measures, before and after.3
Identify two or three organizations whose issue areas intersect with yours and open a conversation about factual alignment. Not message coordination — factual alignment. Are you citing the same primary sources? Using consistent terminology for shared legislation? Referencing the same vote records? A working group on shared factual infrastructure is a legitimate, legally clean activity that could substantially improve the coherence of the progressive information ecosystem on the issues that matter most to your portfolios.
Your board understands earned media. It understands digital advertising. It does not yet understand AI information environment management as a strategic function. Briefing them — with the live demonstration of what AI currently says about your key issues and endorsed candidates — is the single most effective way to generate the organizational commitment this work requires. Show them what ChatGPT says when asked whether your endorsed senator is good on climate. Then show them what it says about the opponent. That conversation tends to move quickly from "interesting" to "what do we do about this."
Every day a progressive organization's scorecard sits in a PDF, its endorsement history exists only in press release archives, and its research is published in formats AI systems cannot parse — is a day AI systems are answering voter questions about those issues and those candidates from whatever sources they can find. Sometimes those sources are accurate. Often they are incomplete. Sometimes they reflect adversarial framing that the organization's own research would contradict, if only that research were accessible to the AI.
The information environment is not neutral. It is being shaped continuously by what is published, how it is structured, and how consistently it is maintained. Organizations that have spent decades building credible, rigorous records of legislative performance and issue research have already done the hardest part. Making that work visible to AI systems is the next step. It is not optional. It is what the mission now requires.
lib/surfaces/registry.mjs), which is the collection code's single source of truth. Brave AI, named in earlier revisions of this paper, is listed there as a retired surface and does not collect; the claim has been removed. ↩getRemedyPricing() in lib/report-minisite.mjs requires an isOrg flag and throws when it is absent rather than defaulting, on the reasoning recorded in the code that "a defaulted noun mislabels half the subjects." Remediation is also priced separately for organizations. ↩schema/v1/) defines three separate type sets: candidate (policy platform, endorsements received, campaign background, candidate comparison), officeholder (voting record, legislative achievements, committee roles, constituent service, official statements) and advocacy (legislative scorecard, organizational endorsement, policy position, member constituency, funding relationship). How answer engines handle a custom JSON-LD namespace is not publicly documented; structured data is more parseable than prose, but the size of any improvement for political content is a hypothesis under test rather than a measured result. ↩