GPS research methods: how to check and compare AI answers
When reading GPS (GEO Positioning System) research, consider the questions, models, dates, regions, languages and counting rules together. This guide explains how to check AI answers and sources and compare records under consistent conditions.
What is GEO, and how does it relate to SEO?
GEO, or Generative Engine Optimization, concerns how content appears in generative AI answers. GPS starts by studying how AI describes a brand and which sources it cites. The team then decides what, if anything, needs to change.
Term and research background: GEO: Generative Engine Optimization — original research
SEO means search engine optimization. A clear site structure, accessible content and working links remain foundations; GEO does not replace them. Google says the same SEO fundamentals apply to its AI search features. Meeting the requirements does not guarantee indexing or citation.
How should PR teams choose GEO research and monitoring tools?
Start with the question your team needs to investigate and the decision the findings should support. REALI recommends these six checks to clarify what information you can obtain, what people need to do, and what the service covers.
Can you trace the test conditions?
Check whether questions, models, test dates, markets, languages and run counts are recorded. When conditions change, can you identify those differences before treating the results as a trend?
Can you inspect the original answers and citations?
Look for access to answer text and visible citation links so your team can check the source pages. Can you record missing citations, broken links or sources that do not support the answer?
Are review and handover responsibilities clear?
Who approves the questions, checks the findings and decides what to do next? Can another colleague follow the test conditions, reasoning and unresolved issues when taking over?
Can research inform what you monitor next?
Can your team use the findings to select questions worth tracking, then agree on retest frequency, ownership and review dates? Changes in test conditions should remain visible alongside the results.
Will the deliverables support your work?
Agree on the questions, test records, citation data, analysis or reports you will receive, including export options. Metrics should explain their calculation, denominator and exclusions so the team knows what the numbers mean.
Are data handling and service scope agreed?
Confirm available models, markets, languages, usage limits and fees. Also establish how data may be used, who can access it, how long it is retained, and whether interpretation and follow-up work are included.
Confirm the selected tool’s scope before deciding whether it fits your team. AI answers and citations can change; observed changes still need investigation to understand their causes and implications for your work.
Who decides the questions?
Your team decides the purpose, audience and scope. Question Skill helps explore directions and prepare question suggestions; the team selects, revises and approves the questions to use. The Skill does not decide what the brand should research.
Does a citation mean the answer is correct?
No. A citation points you to a source. You still need to check the original page.
- AI answerWhat did it say?
Retain the original wording. A summary is not the original answer.
- Visible citationWhich page did it list?
Record the link. If no citation is shown, record that too.
- Check the original pageDoes the source support this statement?
Check its date, context, scope and reliability.
Record the source and version of brand-provided material, AI answers and external pages separately. Having a copy does not mean its contents have been verified or that you have permission to publish it.
What should you consider when comparing results?
Look separately at which websites were cited, how often they were cited and what the answers actually said. If models, questions or test conditions differ, one figure is not enough for a direct comparison. A before-and-after change also needs further evidence before its cause can be identified.
Why distinguish mentions, citations and recommendations?
- Brand mention rate
- What share of the valid answers included in the analysis mentions the brand? First define what counts as a mention. Then divide the number of valid answers meeting that definition by all valid answers included. Count a brand at most once per answer.
- Share of citations to the official website
- First ask what is being counted: citation events, distinct pages or distinct domains? The denominator and rules for removing duplicates affect the official website’s citation share, so the report needs to state both.
- Recommendations or results
- A mention is not necessarily a recommendation. More citations do not, by themselves, establish business results or show that a particular update caused the change.
For each measure, check what is counted, which records are excluded, and the study period and scope.
How does a real test illustrate these principles?
Start with the design of the study asking whether AI keeps citing the same websites. It uses the same questions to compare each model’s source lists within a day and across dates. The guide explains why matching source lists can still produce different answers, and why reports with different denominators need care when compared.
See how shared websites differ from citation volume