Precision prompts turn ChatGPT into reliable EV concierge for sustainable mobility

Consumer EV choice is a decision filled with uncertainty, range anxiety, charging access, total cost of ownership, and model proliferation - factors that slow adoption despite policy momentum under the EU Green Deal and Fit for 55 package. An approachable adviser that converses in everyday language could lower these frictions and steer buyers toward vehicles that match real-world use.

ChatGPT-based “expert system” can guide consumers toward more suitable electric vehicles (EVs), but accuracy depends heavily on how precisely the tool questions users about needs, constraints, and driving patterns, reveals a new study published in Sustainability. In head-to-head tests against a human expert, performance was strong for knowledgeable buyers and improved markedly for everyone once the system asked more targeted follow-ups.

The article, “Using the Large Language Model ChatGPT to Support Decisions in Sustainable Transport,” evaluates whether a large language model can operate as a domain adviser for EV selection and how prompt design influences recommendation quality.

What problem is the study trying to solve?

Consumer EV choice is a decision filled with uncertainty, range anxiety, charging access, total cost of ownership, and model proliferation - factors that slow adoption despite policy momentum under the EU Green Deal and Fit for 55 package. An approachable adviser that converses in everyday language could lower these frictions and steer buyers toward vehicles that match real-world use.

To test that proposition, the study builds a conversational recommender that replaces technical forms with natural-language questions and answers. Instead of asking for exact model names or detailed specs, the system elicits budget bands, typical range needs, body type preferences, household capacity requirements, and access to private or public charging. ChatGPT then maps the answers to viable options.

The authors position this as a sustainability intervention: better-matched vehicles can nudge hesitant buyers toward electrification while avoiding over-sizing and mis-spending, two patterns seen among first-time EV shoppers.

How does the ChatGPT adviser work and what data does it use?

The prototype ingests Polish EV market information from public catalogues and specialist databases, augmented with live listings scraped from major portals; the data are cleaned and normalized into formats suitable for retrieval during dialogue. A representative set of 18 EV models, spanning city, mid-range, premium, and minivan classes, seeds the knowledge base used to ground responses.

Under the hood, the recommender blends familiar techniques, content-based filtering, collaborative and demographic approaches, matrix factorization, rule-based logic, and machine-learning heuristics, with the language model’s dialog abilities. During a session, ChatGPT interprets the user’s constraints, queries the curated data, and returns ranked suggestions with rationales framed in plain language.

For evaluation, 34 users were split into advanced, intermediate, and beginner EV knowledge groups. Each participant stated initial preferences, received system recommendations, and then had those recommendations assessed against the choices of a human domain expert to gauge match quality. The research team iteratively refined prompts and the decision flow to improve performance.

Does asking better questions change the outcome?

In a baseline test with five high-level questions, the system performed well for advanced users, averaging 9.8/10 on the study’s accuracy scale and diverging from the expert in only a minority of cases, yet it faltered with intermediate and beginner groups, often recommending vehicles that were too large or too expensive for stated needs. Inaccurate recommendations reached 92.31% for intermediate users and 81.82% for beginners under the baseline design.

The turning point came when the authors expanded the interview to 18 questions by adding 13 targeted follow-ups on range patterns, trip frequency, charging options at home and work, cabin and cargo needs, price ceilings, and winter operation. With this richer intake, accuracy rose across all groups and the share of system choices consistent with the human expert increased substantially. Inaccurate picks fell to 10% for advanced users, 53.85% for intermediate users, and 72.73% for beginners; average scores improved to 9.9, 8.85, and 8.64, respectively. The tendency to propose unnecessarily large body types effectively disappeared.

The authors’ process takeaway is direct: prompt design is the lever. When the dialogue elicits concrete, context-specific constraints, the language model behaves more like a domain expert and less like a generic search assistant. Conversely, shallow questioning leaves too much ambiguity, prompting plausible but ill-fitting recommendations for less-experienced buyers.

Why the findings matter and where the limits are?

Retailers and platforms can embed conversational advisers to reduce drop-offs among first-time EV shoppers, provided the flows press for the details that determine total cost and usability. Public agencies and NGOs promoting sustainable transport can adapt the approach to local markets by grounding the system in region-specific data sources and tailoring prompts for language and infrastructure realities.

The researchers also point to a broader lesson for decision support with large language models: domain grounding plus disciplined questioning matters more than novelty. By curating trusted data and forcing the conversation through critical constraints, range, charging, budget, size, systems can deliver advice that approaches expert quality for a wider slice of the public.

The study acknowledges some limitations too. The test cohort is small and drawn from one country; outcomes may shift in markets with different charging networks, incentives, or pricing structures. Language-model variability also introduces uncertainty that must be managed with prompt controls, retrieval grounding, and regular updates to the data pool. The authors call for larger trials and multilingual deployments to validate portability and fairness.