Best Practices for Generative Engine Optimization (GEO/AEO)

This post outlines Upswing's recommendations for generative engine optimization (GEO/AEO). Large language models (LLMs)—the systems behind tools like ChatGPT, Claude, Gemini, and Google's AI Overviews—increasingly sit between people and the information they're looking for. GEO/AEO is the practice of influencing how those systems source and present information to users.***

How to Make Sense of GEO/AEO Offerings

Any person or organization evaluating GEO/AEO services or assessing firms to conduct it should first look at the methods of such firms and ask the following questions.

• On which models is the analysis based? Different tools provide different answers—and even within tools, the answers from paid and free versions can vary. 

• Does the analysis use an API feed to pull answers from the LLM or the consumer user experience? An API draws data straight from the model, while a consumer version shows what an actual user might see when they use the tool. 

• Are the LLMs provided context on the user or tested as blank slates? Blank-slate prompts rarely reflect actual use, since most users have histories the models draw on. Our research shows this pre-context can have massive implications for outputs.

• Does the analysis have statistical rigor? Answers can vary from day to day or even between minor word choice shifts, so a handful of data points is difficult to build a coherent strategy around. 

• Which search queries are the LLMs using to answer users? Most LLMs are conducting their own web searches behind the scenes; those searches determine the output. 

• What is the process for measurement, evaluation, and strategic retooling? Because this field is so new, trial and error is essential to getting things right. An answer you shape today could shift over time, which means your GEO/AEO strategy must be iterative. 

Upswing’s GEO/AEO Best Practices

Upswing Research & Strategy’s GEO/AEO best practices framework is based on original analysis of LLM responses across thousands of prompts spanning candidate identity, policy issues, voting records, and electoral decisions. Below are our high-level recommendations, grounded in those findings and supplemented by emerging field best practices.

• For political campaigns, aggregator profiles are often more important to AI than your own website. Ballotpedia, OpenSecrets, VoteSmart, GovTrack, and Wikipedia together account for a large share of LLM sourcing, despite virtually no social media presence. AI models reproduce structured candidate data from these platforms near-verbatim when answering questions about voting records, ideology, campaign finance, and qualifications. 

  • Profile completeness, accuracy, and regular updates are key to shaping how these aggregators might influence voters. Ensuring issue positions, biographical data, and donor information are current and structured is a low-cost starting point. 

  • Where possible, using similar copy and key phrases across sources can strengthen the likelihood of LLMs pulling from your preferred content. 

  • Government websites also provide an opportunity for incumbents. Our research finds that ChatGPT will often pull citations from sources that contain a .gov domain. This means that incumbent legislators, their official bios, committee pages and press releases are a controlled ecosystem for defining themselves and issues.

• Earned media—including local media—matters. In our research, local media alone can account for more than 10% of citations on state-level political queries. The greater proximity of an outlet to a story, the more likely the outlet’s perspective will be considered authoritative on a topic, particularly if the reporting is cited by state-level or national outlets through syndication. This suggests that earned media relationships are primary mechanisms for shaping what AI systems surface as information to users.

• Content gaps get filled one way or another. Every topic you do not invest in shaping independently is an opening for adversarial framing that becomes embedded in AI responses. When no structured, factual content is available for model training, less credible sources fill the void. 

• Know the difference between GEO/AEO-first and social-first domains. Some outlets are heavily cited by LLMs but nearly absent from social listening data. Others have strong social traction but limited or inconsistent LLM citation. These require different strategies. GEO/AEO-first domains reward structured, linkable, policy-dense content, regardless of digital engagement. Social-first domains reward engagement-driven content that may never reach LLM training pipelines.  

  • Reddit was formerly a prominent source for LLM citations (though that appears to have changed, and can always change back), while YouTube is a currently popular citation source. So, social media does also present a unique opportunity to establish LLM visibility, though in our view is not the top priority for AEO/GEO.  

• Even where outputs may align, each model draws from different input pools. This means content that reaches one model may not reach another. In our research, ChatGPT drew about twice as many of its citations from government sources as other LLMs, while Google AI Mode followed structured HTML content most closely among the remaining models. Our data also reveals a sharp divide in how models treat social content: Claude and ChatGPT drew less than 2% of their combined citations from social networks, while Grok, Google AI Mode, and Google AI Overviews each sourced more from social platforms. This necessitates a “yes, and” approach, press releases, structured FAQ pages, earned media, etc. Similarly, social media community engagement should be treated as GEO inputs, not just social outputs. 

• Audit AI crawlers’ access to your information and messaging. None of these recommendations matter if the models cannot read your site. Many publishers and orgs block AI crawlers by default. If you are focused on eyeballs and are not reliant on traffic for revenue, you want as many crawlers on your site as possible. Separately, heavily JavaScript-rendered sites can appear nearly empty to crawlers that do not execute JS, so key issue pages should be server-side rendered. This is a cheap, one-time diagnostic. 

• Structure content so AI can cite it, not just find it. There is a meaningful difference between content that exists online and content that AI systems parse and cite. Structured formats, clear headers, FAQ-style layouts, clean issue-specific URLs (e.g. /issues or /economy rather than /news?id=1234) consistently outperformed unstructured content in LLM citation rates in our research. Google AI Mode, for example, follows structured HTML closely and is among the most citation-dense models tested. 

• Measure the field for timing. How fast can you change an answer? Organizations should be measuring how fast rapid response on an issue can modulate answers.

** GEO/AEO is a nascent but rapidly growing field built on the idea that the same discipline applied to SEO over the years can similarly shape large language models as these tools become synonymous with traditional search. These tools and models change rapidly and often without any notice. What a model cites today may differ significantly tomorrow, particularly as many of these platforms use live web retrieval and train on consistent user inputs.

Because developers do not disclose details about what drives LLM outputs, every strategy in the field of GEO/AEO—including Upswing’s—is based on present observations and does not reflect any definitive claims about backend mechanisms. By comparison, SEO, while similarly complicated and opaque, developed over decades of practitioner testing and some degree of documentation from search engine developers. GEO has neither.

The recommendations here represent Upswing’s best current read of the evidence, grounded in original data analysis and consistent with emerging field practice. They are also subject to change as the technology, the platforms, and the competitive landscape evolve. Practitioners should consult their own technical and legal advisors, as needed.