The Citizen Can't Tell When It's Paid For

A call for disclosure standards on AI-mediated political perception work

Orchestra — the strategic communications holding company behind BerlinRosen, Derris, and several other agencies — recently commissioned research from Evertune analyzing 63,000 prompts across six AI answer engines, with a close look at the Georgia Senate race between Jon Ossoff and Mike Collins. The findings are genuinely useful: AI answer engines reward clear, structured, retrievable content. Ossoff's campaign organized its record into standing issue pages with named legislation and funding by county; Collins's site remained a feed of dated press releases. The gap in citation share that followed wasn't manipulation — it was one campaign writing better web pages than the other.

That distinction is worth holding onto, because it marks the boundary of a much more important question the report doesn't address: what happens when the content shaping an AI system's political characterizations isn't a campaign's own honest account of its record, but a paid intervention designed to shift how a citizen understands a candidate, a PAC, or an advocacy organization — without the citizen ever knowing an intervention occurred?

Two activities that are technically identical and democratically different

The mechanism behind AI-answer-engine optimization is the same whether the subject is a shampoo brand or a Senate candidate: seed clear, well-structured evidence into the sources a model draws from, and the model's characterization shifts accordingly. For a consumer brand, this is unremarkable. A shopper doesn't need to know a brand invested in structured product content to be getting accurate information about a real product.

Commercial markets presume persuasion. Democratic institutions presume transparency. A consumer expects companies to compete for attention; a voter expects to know when organized political influence is being purchased. The same technical intervention therefore carries different legitimacy depending on the institution it serves.

For a political actor, that difference is the whole issue. A citizen asking an AI assistant to explain a candidate's record, or to characterize a PAC's influence, has no way of knowing whether the answer reflects the balance of public information or the success of a paid campaign to shape that balance — even when every fact seeded into the system is true. The commercial work is disclosed nowhere the citizen can see it. That is a structural property of the activity, not a question of whether any particular firm executes it well or in good faith.

This is not a hypothetical concern about some future bad actor. Firms across the emerging GEO/AEO industry have begun offering these services to political organizations, and interest from campaigns and advocacy groups appears to be growing alongside the underlying measurement technology.

Why an outright ban is the wrong tool

It would be tempting to call for Congress to prohibit "partisan optimization" of AI answer engines outright. We don't think that's right, for two reasons.

First, it isn't definable in a way that survives contact with what campaigns actually do. A candidate organizing their record into clear, structured issue pages — precisely what the Orchestra/Evertune research found Ossoff's team did well — is indistinguishable, as an activity, from what a paid remediation vendor does on behalf of a client. Any statute broad enough to reach the latter would also reach ordinary campaign communications, chilling exactly the kind of clear public communication that helps citizens.

Second, restricting how a candidate, party, or advocacy organization structures and publishes its own positions is core political speech. A ban aimed at "optimization" runs into the same wall that has stalled far narrower efforts — synthetic media disclosure, robocall restriction — and would likely fail faster, on firmer First Amendment ground, than either.

What actually tracks the harm

The harm isn't optimization. The harm is undisclosed optimization.

Disclosure — not prohibition — tracks that harm precisely, and it needs a concrete home, not just a principle. The most natural fit is to extend existing paid-political-communication disclosure regimes to cover this activity directly: a political actor, PAC, or vendor engaged in a paid campaign to shape how AI systems characterize a candidate, ballot measure, or advocacy organization would report that engagement the way independent expenditures and paid digital ads are already reported today — through campaign finance filings where a campaign or PAC is the purchaser, and through a comparable disclosure obligation on the vendor side where the buyer is a third party. A visible disclaimer on the AI-facing content itself, analogous to "paid for by," is a natural complement, not a substitute. None of this requires solving every implementation detail in advance; it requires establishing that the obligation exists. This asks nothing of a campaign that simply writes a good website. It asks everything of an undisclosed influence operation wearing the same technical clothing.

Where AIVO stands

Our own view has been to separate measurement from intervention. AIRS, our political AI-representation methodology, measures how AI systems characterize candidates and organizations across the political spectrum. It does not attempt to change those characterizations, even though the same underlying technology could. We've held that line even where the measurement capability would support crossing it. We think that boundary is right for us. We also think the industry needs a rule that doesn't depend on every vendor choosing it voluntarily.

The question the Orchestra/Evertune report leaves open isn't whether campaigns should communicate clearly — they should, and AI answer engines reward it. It's whether the public should be able to tell the difference between a campaign that communicated clearly and one that paid to be perceived that way. Democracies don't prohibit persuasion. They require that citizens know when persuasion has been paid for. AI-mediated political representation should be no different.