What every campaign, elected and advocacy organization needs to know about political discovery
As the 2026 midterm races take shape, AI answer engines are increasingly shaping how voters understand candidates, issues and contests. Orchestra partnered with AI search intelligence company Evertune on “Winning the AI vote,” a new report analyzing 63,000 prompts across six answer engines, including a close look at the Georgia Senate race between Jon Ossoff and Mike Collins.
The findings offer an early look at the political story AI is telling voters — and reveal which campaigns are beginning to break through ahead of November.
Google no longer simply returns a list of links — its AI tools give voters a ready-made account of a candidate's background and positions right on the results page. ChatGPT offers an even more direct route: ask a political question and get a narrative in seconds. Studies show people often trust and are influenced by these AI-generated answers.
What ChatGPT, Google's AI tools and other platforms say about a candidate can shape a race before a voter sees an ad, watches a TikTok or gets a knock on the door.
The work of making a campaign's positions and supporting evidence easy for answer engines to understand and present clearly.
Evertune ran 63,000 prompts across six answer engines, sampling each prompt multiple times to support statistically significant findings — nationally, and in the Georgia Senate race.
News stories account for the majority of citations in every answer engine tested. Across both news and institutional sources, such as government, university and research organization sites, many of the leading URLs are explainers or other pages that remain useful over time. Some were published in 2024 and still appeared across multiple answer engines in the 2026 sample.
For questions that require interpretation or judgment, answer engines look to independent reporting and analysis.
For factual questions about what a party supports, engines lean more on organizational writing and social posts alongside the reporting.
What connects the strongest URLs is clear explanation and structure: descriptive subheads and well-organized sections that make key information easy to identify.
A press release generates immediate attention. A clear explainer supplies the answer every time an engine meets the question again. The goal: a body of public information that makes the campaign's story easy to find, verify and reproduce.
Ask answer engines what the midterms are about and the answer is consistent: the economy and cost of living rank first, for voters overall and for each party. The interesting pattern emerges when models explain what each party would do about them.
The Republican economic story is strikingly consistent — energy, tax cuts, less spending. The Democratic story is more fragmented, shifting among tax credits, housing, drug costs, paid leave and utilities.
Neither party has a monopoly on message discipline. On immigration, Republican answers cluster around enforcement, deportation and border security. On healthcare, prescription-drug prices, the Affordable Care Act and Medicare negotiation reinforce the same Democratic affordability argument — across every model.
Begin with a simple question:
“What does AI think we stand for?”
If the answer changes across models, the campaign may have a messaging gap, a source gap — or too little consistent language online. A scattered answer is a diagnostic for deeper communications problems.
Voters enter AI-powered political research through two main doors: direct conversations with ChatGPT and answers generated inside Google searches. Each works differently — and success in one does not guarantee success in the other.
of observed midterm-election prompt volume happens on ChatGPT — making it the primary chatbot environment to monitor.
On ChatGPT, political research unfolds through conversation rather than a fixed set of keywords. Voters often start broad, add personal context and ask follow-up questions. A candidate’s story needs to hold together as the conversation gets more specific.
Google occupies a different place in the voter journey, because its AI tools appear within a search behavior that is already deeply established. In a 2024 study from the Bipartisan Policy Center and Morning Consult, a majority of voters said they used Google to find election information. Increasingly, AI Overviews and AI Mode deliver a generated answer directly within that familiar search experience.
Shapes direct, conversational research. Leans heavily on canonical party and platform documents. Use structured manual testing to see how the story changes as follow-ups get more specific.
Appears inside the search behavior voters already rely on, and pulls more visibly from polling, issue research, news and video. Track which searches generate AI answers and which sources Google elevates.
Combining both into a single “AI visibility” score obscures the differences. Measure performance model by model — and decide where to focus accordingly.
Jon Ossoff vs. Mike Collins — a window into how answer engines assemble a candidate narrative from individual pages. And in the battle to shape those answers, Ossoff is winning.
Both candidates have official congressional sites. Answer engines drew on Ossoff's nearly eight times as often across the sample.
Both sites contain real work. The difference is structure: one isolates concrete, retrievable answers; the other buries specifics inside feeds of dated updates.
Across models, Ossoff’s Senate site appears as a top source more consistently than Collins’s House site, while The Atlanta Journal-Constitution and other news sources also appear repeatedly. The pattern shows how first-party information and independent reporting both shape candidate answers.
ElectJon.com received more citation share than MikeCollinsGA.com on ChatGPT, Google AI Mode and Google AI Overview. Its strongest pages included an issue-specific contrast, a detailed biography and campaign-hosted news coverage.
On cost-of-living prompts, The Atlanta Journal-Constitution accounted for a quarter of all citations in the ChatGPT responses tested — more than any other domain.
Its most-cited pages are evergreen explanatory articles — candidate hubs, economic-agenda comparisons, health-cost explainers — not stories pinned to a point in time.
The most-cited YouTube channels are largely news organizations, not campaign or creator accounts. A local TV segment stays available to Google's answer tools — explaining a candidate to someone who never saw it air.
As more people begin their research with AI, the central question becomes:
Which version of the story
becomes the answer?