Synthetic Users
for GEO
Transforming customer research to simulate car buying behaviour
An agentic research framework that builds brand-neutral synthetic personas and a large prompt catalogue.
LLMs are changing how people discover and evaluate brands.
People now ask large language models to research, compare and choose vehicles. In that shift, a brand that is not surfaced by AI is absent from the decision long before a dealer visit or a search result click.
Brands that are not referenced, cited, or selected by AI platforms effectively cease to exist in the eyes of users.
AI discovery is early-stage influence
Users are forming a shortlist inside conversational systems.
Visibility is conversational
Is the brand is cited and recommended in generated answers?
Measurement becomes behavioural
Evaluation requires archetypes rather than individual representations.
From Insights to Value
Convert fragmented qualitative research into a defensible behavioural framework capable of producing thousands of vehicle-purchase queries for large-scale GEO analysis across brands and customer journey stages.
Research was abundant but not operational for agentic frameworks.
Abstraction of first hand customer data to meet compliance constrains.
Users prompt differently in different phases of the customer journey.
Linguistic synthetic dataset capturing how different customer segments naturally ask questions.
An agentic synthetic user framework
The system translated qualitative research into synthetic users, generated prompts by journey phase, validated them, and handed the catalogue into GEO workflows for large-scale brand visibility measurement.
Research Analysis
Extract, normalize and abstract customer evidence from studies, interviews, workshops and repository material.
Synthetic Personas
Model behaviour-rich personas with enough context, language and motivation to simulate buyer decision-making.
Prompt Generation
Create a catalogue mapped to customer journey phases, emotional drivers and varying levels of purchase intent.
Human in the Loop
Experts review prompts for neutrality, duplication, coverage, language consistency and approving before scale.
Handover
Translate the outputs into a practical instrument for the GEO team.
From Single Agent
To Multi Agent Orchestration
Five architectural decisions transformed qualitative research into a scalable agentic workflow while preserving behavioural fidelity.
Modular prompts adapt to where the buyer is in the journey,
not just who they are
Knowledge level
Prompt depth adjusts to novice, intermediate or expert familiarity with vehicles and EV topics.
Journey phase
Discovery, consideration, decision and transaction each trigger very different prompting behaviour.
Emotional driver
Status, sustainability, reassurance or excitement shape the prompt wording as much as the topic itself.
Brand neutrality
Proprietary shorthand is removed so the catalogue can measure visibility fairly across competing brands.
A stronger foundation
GEO had a grounded starting point.
The team received a prompt catalogue rooted in first-hand OEM research rather than abstract AI assumptions.
Research became operational.
A fragmented repository became usable input for GenAI and agentic workflows, not just reference material for slide decks.
The project created internal momentum.
Leaders could see how research and marketing intersect in AI discovery and why investing in research infrastructure matters.
Time-to-market improved.
The method showed how AI research operations can accelerate delivery without abandoning rigour, compliance or human review.
The bigger outcome was cultural. This shifted the conversation from reporting insights to operationalizing research for AI systems.
A practical stack,
then five lessons
The stack was deliberately lightweight. The hard part was judgment: choosing data, structuring context, neutralizing bias and validating outputs before scale.
Behavioural evidence
Grounded the personas in observed customer behaviour.
YAML structures
Persona memory, taxonomies and prompt templates were stored in files that improved through iteration.
CrewAI + Azure OpenAI
Micro-agent workflows handled extraction, generation and validation while keeping the human reviewer in the loop.
Choose the data
Be critical with the material at hand and keep the signal tied to real behaviour, not just available documents.
Context data for LLMs
Highly descriptive, diverse material gives models enough context to infer intent with greater consistency.
Neutralize and anonymize
Plan time, compliance and brand fairness often require manual modeling work.
Use the right framework
Different agent architectures suit different phases of the research journey. Matching method to task lowers drift.
Validate qualitative output
Analytical checks, expert review and human judgment remain essential before any synthetic output can be trusted.
The innovation lies in bridging human insight with data-driven orchestration
Client identity and proprietary figures abstracted for confidentiality