a cross-campaign research initiative at Tech for Campaigns exploring how ai chatbots shape public discovery of political candidates - and how strategic, structural decisions can improve clarity, accuracy, and representation within ai-mediated systems.
timeline
july 2025– jan 2026
organization
Tech for Campaigns
role
market research &design strategy
focus
ai systems, information design,
platform-mediated perception
↦ overview
As AI chatbots increasingly become the first place people go for information, political candidates are no longer discovered only through websites, news articles, or even social media. More and more often, the first interaction happens through a summary (generated by an AI system!) before a voter ever clicks a campaign-owned page.
Tools like ChatGPT, Gemini, Perplexity, and Claude now sit in the middle of the discovery process. They interpret questions, decide what information matters, and shape how public content is framed long before users reach the original source.
At Tech for Campaigns, I worked on a cross-campaign research initiative focused on understanding how political candidates are represented inside these AI systems — and how design and content decisions can meaningfully influence those representations.
Rather than asking “what performs best,” the research asked a different question: how do AI systems read, prioritize, and synthesize publicly available information in the first place?
To answer that, the team established a baseline view of how candidates appeared across multiple AI platforms, using a combination of human evaluation and system-level review. From there, we explored a set of strategic content and information architecture interventions - including updates to site structure and the use of publicly accessible content channels - to better understand which signals seemed to carry the most weight.
Rather than focusing on campaign outcomes, this work centered on system behavior.
After allowing time for AI models to re-crawl updated content, we planned follow-up reviews using the same evaluative lens to observe how representations shifted over time and across platforms. More broadly, the research explored how different structural signals shape AI interpretation in systems we don’t directly control.
Although formally categorized as market research, this project lived squarely within design strategy: mapping complex, opaque systems, identifying leverage points, and designing for legibility inside platforms we have to design with, even when we don’t design for them.
the goal
The goal of this work was to better understand how AI systems construct representations of political candidates - and where design and content decisions can meaningfully improve clarity and accuracy.
We focused on AI chatbots including ChatGPT, Gemini, and Claude, examining how each interpreted the same publicly available information in slightly (and sometimes surprisingly) different ways.
While this research also supported broader program goals around messaging, fundraising, and press readiness, the core focus stayed firmly at the system level: understanding how AI-mediated discovery works, rather than optimizing for any single outcome.
the hypothesis:
We began with a simple but important hypothesis:
Making strategic changes to a candidate’s website, online presence, and content strategy may influence how that candidate is represented in AI-generated responses.
Through the research, we explored how different types of publicly available signals - including content structure, clarity of policy information, and presence across open platforms - shape how AI systems construct candidate summaries.
Rather than treating optimization as a checklist, this work reframed content as infrastructure: something designed not only for human readers, but for AI systems tasked with interpreting, condensing, and presenting information at scale.
my role
I contributed to the design, execution, and synthesis of a multi-phase study evaluating how AI systems represent political candidates.
More specifically, my work focused on:
Designing and applying structured evaluation frameworks for AI-generated outputs
Participating in human grading across multiple AI platforms
Analyzing qualitative and quantitative patterns across candidates and systems
Supporting synthesis discussions that informed downstream strategy
I wasn’t designing screens or interfaces. I was designing strategy for systems that don’t have a UI, but still deeply shape user experience, perception, and trust.
↦ research as
design strategy
phase 1: baseline mapping & live review
(july → august)
The first phase was about getting grounded and understanding how candidates were actually being represented across AI platforms before any idealized fixes or assumptions got in the way.
We defined queries to mirror how real people search: questions about who a candidate is, where they’re running, what they believe, and how visible they seem overall. Baseline analysis combined machine-assisted review with human evaluation, with the goal of widening assessment coverage while avoiding unnecessary complexity.
Using a standardized evaluation framework, we reviewed AI-generated responses to consistent natural-language prompts. Once again, the emphasis wasn’t on performance scores or “winning” outputs. It was on interpretability, as in how clearly, consistently, and coherently candidates were being represented across systems.
Importantly, candidates were reviewed as they were, while content changes were happening in parallel. This wasn’t a clean lab experiment. It was a live, shifting information environment, and that messiness turned out to be a feature, not a flaw. It allowed us to observe how AI systems respond to evolving signals in real time, rather than to static, ideal conditions.
This phase functioned like a system map. It helped us pinpoint where AI responses broke down, where information gaps existed, and which structural signals seemed to carry disproportionate weight.
crawl pause: letting the system respond
(august → mid september)
After Phase One, we intentionally stopped. We paused evaluations to give AI systems time to re-crawl and reprocess updated content, which was a decision that ended up being just as important as the analysis itself.
This pause reinforced one of the project’s clearest insights: designing for AI systems requires patience, sequencing, and restraint. Unlike traditional UX iteration, feedback loops here are delayed, opaque, and non-linear. You don’t ship a change and immediately see the result. You wait. You observe. You resist the urge to overcorrect.
phase 2: re-review & targeted interventions
(late september)
Phase Two focused on revisiting candidates after the crawl window to understand how representations shifted across time, and across platforms.
Here, the work became less about what AI systems were saying and more about how and why those shifts occurred. The team examined where evaluative nuance was useful and where it simply added noise, helping us rethink how future research might balance depth with clarity.
This phase involved re-running core queries, analyzing post-update patterns, and synthesizing learnings with an eye toward broader implications for information design. Human testing questions were refined to reduce ambiguity, favoring clear yes/no evaluations when nuance wasn’t actually adding value.
Accuracy and completeness remained central. Tone-based analysis was reserved for questions where narrative framing or controversy could meaningfully shape perception, as in moments where how something was said mattered just as much as what was included.
Rather than cataloging every possible source, we shifted focus toward identifying unusual or unexpected sources surfaced by AI systems. Those moments felt more revealing and more actionable than exhaustive lists. (There’s always more research to do.)
phase 3: evaluation & learning
(completed late october)
In later phases, I completed additional evaluations using the same structured framework to maintain consistency across the study.
This period also included a rapid mini-sprint for a single candidate, where evaluation and updates happened within a tight window (a useful stress test for how quickly AI systems respond to new signals). Additionally, a publicly accessible reference source was updated as part of the broader research context, based on timing and observed platform behavior.
Across these later evaluations, we continued examining how timing and structural changes influenced AI interpretation. One thing became increasingly clear: responsiveness varies widely across platforms, and no two systems update or stabilize in quite the same way.
By this point, patterns were beginning to emerge - and not just about individual candidates, but about how different AI systems prioritize authority, structure, and contextual relevance.
↦ implications
answer engine optimization (AEO) as a design problem
As synthesis progressed, we began framing the work through the lens of Answer Engine Optimization (AEO) - a strategy focused on designing content so it can be surfaced directly as answers by AI-powered systems.
Unlike traditional SEO, which prioritizes rankings and traffic, AEO prioritizes clarity, authority, and structure. The goal isn’t clicks - it’s comprehension!
For political campaigns, that shift introduces both opportunity and risk: greater visibility in the places voters already ask questions, paired with less control over framing, attribution, and nuance.
From a design perspective, AEO isn’t a checklist. It’s an infrastructural problem. It requires disciplined messaging, clearly articulated issue positions, and ethical care in how information is structured and made available to systems that increasingly summarize reality on users’ behalf.
↦ analysis, synthesis & outcomes
early november synthesis
After the evaluation phases wrapped, the work shifted from measurement to meaning.
We moved into analysis and insight-sharing as a group - stepping back to talk through what we were actually seeing, not just what the data said on paper. Together, we surfaced patterns, tensions, and open questions, including:
What shifted in meaningful ways - and what stubbornly didn’t
Priority themes like AI accuracy and bias, website structure, challenges with location-based information, the role of media, and evolving information channels
Platform-specific quirks and behavioral differences
Which approaches felt genuinely design-leveraged versus simply content-heavy
Unexpected or counterintuitive findings that challenged our assumptions
The project concluded with collective reflection and knowledge-sharing, emphasizing organizational learning and collaboration. We also took a moment to celebrate the work (and the broader civic moment it aligned with) recognizing that this research sat at the intersection of technology, politics, and real people’s lives.
individual insights
As I mentioned earlier, this phase of the project was less about metrics and more about meaning. It was a chance to step back and reflect on what the research revealed - not just technically, but personally.
Here are a few of the insights I contributed to our synthesis:
local government is personal (and powerful)
While this research focused on AI systems, one of the most meaningful takeaways for me was how deeply local politics shape everyday life. Before this project, I mostly paid attention to national elections - presidential races, governors, big headline moments.
Working closely with local candidates changed that. Seeing how intimately they understood their districts - the cities, the issues, the people - really reframed local government as something personal rather than abstract. It made me more curious, more engaged, and more invested in my own community going forward.
ai chatbots are already decision-making tools
One of the most striking realizations from this work was how much people already rely on AI chatbots for deeply personal, decision-influencing questions.
People ask tools like ChatGPT about health, finances, education - and increasingly, politics. While these systems shouldn’t be biased, they must be accurate. Misleading or incomplete information doesn’t just confuse users, it can genuinely influence how someone votes.
For much of the 2010s, digital strategy centered on websites and social media. This research made it clear that candidates now need to think one step further about how their digital presence is summarized, translated, and reshaped by AI systems they don’t control.
policy categorization is hard - for both humans and ai
During evaluations, “Key Policy Positions” queries consistently revealed a structural challenge:
Many candidate websites grouped policies under broad umbrellas like Economy or Supporting Families, with specific issues like healthcare or taxes nested underneath. Those issues were important, but not always surfaced as standalone priorities.
That ambiguity made categorization difficult - and not just for evaluators, but likely for voters too. It raised a design question that stuck with me: if policy priorities aren’t clearly articulated or structured, how can we expect AI systems (or people) to interpret them consistently?
structure matters more than ideology
Out of curiosity, I occasionally reviewed the websites of opposing candidates as well. While this wasn’t part of the formal study, a pattern emerged: some opposing sites appeared less structured and less cohesive overall.
It made me wonder how much AI visibility has less to do with ideology and more to do with information architecture. If a site lacks clarity, hierarchy, or structure, what does that mean for how AI systems extract and present information? This felt like a meaningful direction for future research.
synthesis extensions & emerging signals
Following the initial evaluation phases, the work extended into a synthesis and show-and-tell focused on pattern-level learning rather than campaign-specific outcomes. This phase was less about performance and more about behavior - how AI-mediated discovery actually operates over time, and what that means for designing content inside systems we don’t directly control.
Several high-level signals became clearer during synthesis.
First, AI-generated answers showed noticeable variability over time. Even when content remained largely stable, representations could shift in tone, emphasis, or detail. This reinforced an important constraint of designing for AI systems: outcomes aren’t static, and ongoing monitoring matters more than one-time optimization.
Second, human-generated content repeatedly appeared as an important ingredient in AI-generated responses. Sources like local news, long-form video, community forums, and FAQ-style content surfaced alongside more traditional web and press materials. This suggested that conversational, contextual, and question-driven content may be especially legible to AI systems - complementing, rather than replacing, polished campaign messaging.
We also observed that structural clarity mattered. Content organized around direct questions (for example, “What does this candidate believe about X?”) and clearly articulated issue stances appeared easier for AI systems to interpret than content framed broadly or abstractly. In several instances, introducing explicit FAQ-style structure aligned with clearer surfacing of campaign-owned information.
Finally, timing emerged as a design constraint.
Updates to public content were not always reflected immediately in AI outputs, reinforcing that feedback loops in AI-mediated systems are delayed, opaque, and non-linear. Designing for these environments requires patience, sequencing, and restraint - qualities more closely aligned with systems thinking than traditional UX iteration.
Taken together, these insights reframed the work from a study of AI outputs to a broader design question:
How do we create content that remains accurate, legible, and trustworthy when it’s interpreted and summarized by systems acting as intermediaries between people and information?
↦ when ai systems summarize reality on behalf of users, structure becomes ethics - and clarity becomes equity.
post-synthesis update
January 2026:
This case study reflects work completed through the initial research, evaluation, and synthesis phases. Subsequent discussions and show-and-tell sessions extended these learnings into broader strategic guidance around AI-mediated discovery and Answer Engine Optimization.
This page will continue to evolve as new insights emerge.
↦ reflection
personal takeaway
This project genuinely shifted how I think about elections, technology, and civic engagement. It reinforced that design decisions (even invisible ones!) carry real consequences.
Structure shapes understanding. Clarity shapes trust. And systems we don’t see still deeply influence how people form opinions and make decisions.
I’m deeply grateful I participated in this work. It wasn’t just intellectually engaging, it changed how I see my role as a designer working within complex, civic-facing systems.
More broadly, this project reshaped how I think about design’s role in AI-mediated environments. When systems summarize reality on behalf of users, design decisions happen long before an interface appears.
This work strengthened my interest in systems-level design, transparency, and dignity in information access and continues to shape how I approach design strategy in complex, high-stakes contexts.
Note: This case study reflects high-level process and personal learnings. All sensitive details, data, and materials have been anonymized or abstracted in accordance with confidentiality agreements.
a game discovery platform built for players who want to feel seen, safe, and supported.
