No Public Numbers for AI Political Ads in Local Races

As of September 1, 2026—Q3 2026—none of the supplied sources provides a named, dated campaign-level record connecting AI political advertising in a US local race to spend, CPA, ROAS, or vote outcomes. That is a finding about the reviewed evidence packet, not proof that no relevant public record exists anywhere. No targeted FEC filing or platform Ad Library search was performed, so a claim of universal nonexistence would go beyond the available research.

For media buyers evaluating AI political campaign advertising strategies for local elections, the practical answer is therefore narrower than either the sales pitch or the backlash: there is enough public context to justify controlled testing, but no public benchmark in this packet that can responsibly set a budget or forecast performance.

The evidence audit

The missing connection matters more than the volume of material surrounding the subject. A buyer needs to know which campaign used which AI method, over what dates, with how much money, against what baseline, and with what result. Without that chain, generated creative volume, election wins and broad claims about efficiency remain unevaluable.

Why the visible record looks more conclusive than it is

One study examined 3,333 news articles about AI in US elections. It found that 63.58% addressed deceptive uses of AI, while 8.58% covered AI improving campaign operations or outreach.[1] Those figures describe the composition of news coverage. They do not show that deceptive applications occur at the same rate, that operational tools rarely work, or that any particular advertising strategy changes votes.

The study instead helps explain what a buyer encounters during ordinary research. Deception stories are much easier to find than dated accounts of workflow, spend and conversion outcomes. That coverage skew can create two opposite errors: treating abundant warnings as evidence that AI has no operational value, or treating the scarcity of performance reporting as an open field in which vendor projections can stand in for benchmarks. The data supports neither move.

Brookings narrows the performance question from another direction, reporting “little evidence” in the academic literature that generative-AI political advertisements are persuasive.[2] Persuasion is important, but the scope must stay intact. This finding does not cover every operational outcome a campaign might pursue, such as faster production, lower editing workload, volunteer recruitment or donation acquisition. Nor does it rule out an undocumented result from an individual race. It says the persuasiveness literature is limited.

The reason campaigns remain receptive is not difficult to understand. In a Pipeline Fund survey of roughly 1,000 local Democratic candidates, 49% said they received only some, very little, or none of the campaign help they needed.[3] This is evidence of a support gap, not evidence that AI closes it. A time-starved local campaign may reasonably value assistance with drafts, variants or routine production, but need should not be converted into an expected CPA or vote effect.

An election win is not an attribution model

A statement such as “all clients elected” may sound decisive while leaving every buying question unanswered. It does not reveal who the clients were, when they ran, whether the races were competitive, how much was spent, which parts of the program used AI, what comparison was used, or whether advertising contributed to the result. Winning candidates also have fundraising, field operations, endorsements, incumbency, earned media and local conditions acting on the same outcome.

The useful comparison is not between an AI claim and no information. It is between that claim and an auditable record. The site’s FEC-verifiable political-spend analysis shows the basic standard: identify the spender, source the record and attach amounts to a defined period. Conventional spending records do not prove advertising effectiveness, much less AI effectiveness. They do demonstrate what verifiability looks like before anyone begins arguing about impact.

Build the record the public evidence lacks

A local campaign does not need to wait for a national dataset before testing a promising tool. It does need to decide in advance what evidence would justify continuing, expanding or stopping the spend. Recording the plan after results arrive makes it too easy to promote whichever number happens to look favorable.

The outcome should match the job assigned to the test. A donation campaign might select completed donations and cost per completed donation. A volunteer campaign might use qualified sign-ups rather than form starts. A production test may focus on approved assets delivered per staff hour, provided quality requirements remain fixed. ROAS is meaningful only when the return and attribution method are explicitly defined; it should not become a decorative label for engagement.

Vote totals require particular restraint. They are observed at the election level, while advertising is one input among many. A campaign can record vote outcomes, but attaching the entire result to AI-generated ads without an appropriate design does not produce an AI performance benchmark.

Quality control belongs in the performance file

The Brennan Center’s testing makes the operational risk concrete. Generated political ad copy included vague slogans such as “a brighter future,” invented candidate commitments, and fabricated URLs or social handles that could send traffic to the wrong destination.[4] The testing was connected to the 2024 cycle rather than a live 2025–2026 local campaign, so it is not a failure-rate benchmark. It is a documented list of errors that a review process should be designed to catch.

That review is part of cost, not an administrative footnote. If staff must verify every promise, repair destination URLs and recheck regenerated variants, those hours belong beside vendor and production costs. If an invented commitment is published, the consequence is not offset by the number of inexpensive versions the system produced.

• Archive the prompt or input, raw output, approved version and identity of the reviewer.
• Test every destination URL and social handle from the final rendered ad rather than from the working document.
• Check candidate commitments and biographical assertions against an approved source.
• Record rejected assets and correction time so efficiency calculations include the full workflow.
• Keep disclosure and authorization approval with the creative version that actually ran.

Platform rules constrain the test; they do not benchmark it

Google’s support material describes a creative-labeling rule taking effect in July 2026, although the supplied policy is a general advertising rule and is not explicitly limited to political ads.[5] A third-party account also reports that Google’s political-ad restrictions prohibit first-party voter-file targeting, affinity and in-market audiences, lookalikes, and behavioral retargeting; the exact underlying policy language was not independently verified in the reviewed packet.[6] Buyers should confirm current requirements directly and use the site’s AI-generated political-ad disclosure tracker for the broader policy backdrop.

Compliance can determine whether a tactic is available and how creative must be handled. It cannot supply the missing campaign result. As of Q3 2026, the reviewed public context supports a disciplined test, not a performance promise. Only a campaign’s own dated record of method, timeframe, spend, verification work and selected outcome can justify what happens to the next dollar.

References

1. Artificial Intelligence in Election Campaigns: Perceptions, Penalties, and Implications, Taylor & Francis Online.
2. A Policy Framework to Govern the Use of Generative AI in Political Ads, Brookings Institution.
3. How AI Could Reshape Down-Ballot Campaigns, Campaigns & Elections.
4. Generative AI in Political Advertising, Brennan Center for Justice.
5. Google Ads Policy: AI Labeling for Ad Creatives, Google Ads, July 2026.
6. Winning Social Media Political Advertising Under 2026 Ad Restrictions, Strike Social.