The Challenge with LLM-Driven Recommendations
Most AI assistants generate recommendations directly from large language models (LLMs). Because LLMs are probabilistic, the exact same prompt can produce entirely different outputs across multiple runs. While this variability is great for conversational fluency, it makes product recommendations notoriously difficult to control. For brands and retailers, recommendations must be consistent, explainable, and strictly aligned with their catalogue.
Aroura’s Approach: Separating Logic from Language
Aroura solves this by decoupling the decision logic from the AI conversation layer. Instead of relying on an LLM to guess or arbitrarily select products, Aroura uses a structured, deterministic scoring engine for product selection. The AI assistant then focuses purely on explaining those selections to the customer in natural language.
This ensures recommendations remain stable and predictable. If the same customer profile interacts with the same catalogue, the system will deliver the exact same recommendation every time.
Aroura’s 4-Step Recommendation Process
1. Understanding Customer Intent
The process begins at a kiosk or conversational interface, where customers describe their skin concerns, goals (e.g., hydration, acne care, a glass-skin glow), and lifestyle preferences (e.g., vegan, cruelty-free). Rather than treating this input as free-form text, Aroura converts the conversation into a structured profile to accurately guide the recommendation engine.
2. Searching the Product Catalogue
Next, the system searches the catalogue using a blend of semantic similarity and structured product attributes. Moving beyond basic keyword matching, it evaluates product descriptions, ingredients, benefits, and targeted concerns to narrow down the catalogue to a highly relevant subset of candidates.
3. Evaluating Products with a Scoring Engine
Candidate products are then evaluated using a deterministic scoring model across several specialised layers. A Dermatology Fit layer matches skin concerns with ingredient compatibility. A Cultural Beauty Trends layer accounts for regional routines. A Semantic Relevance layer measures alignment with the initial request, while Regional Signals detect emerging local demand. Together, these layers rank products to reflect both individual needs and broader market trends.
4. Building a Simple Routine
Instead of overwhelming the customer with a long list of options, Aroura curates three complementary products that form an easy-to-follow routine:
- Essential: Addresses the primary skin concern.
- Boost: Enhances the overall effectiveness of the routine.
- Treat: Provides deeper, targeted care.
Why This Approach Matters
By isolating the recommendation logic from the AI conversation layer, Aroura guarantees deterministic and transparent product selection.
The scoring engine decides what to recommend, and the AI assistant determines how to explain it naturally. This empowers brands and retailers to retain absolute control over their merchandising while still delivering a highly personalised, conversational shopping experience.