AI-Powered Retail Recommendation System with Personalized Product Discovery

In retail, maintaining visibility and brand recognition requires constant audience research, market analysis, and a willingness to explore beyond conventional methods. For Verde Local, the decision wasn’t whether to implement an AI-based recommendation system, but how to do so in a way that added value for customers while maximizing results.

The goals for Verde Local’s recommendation system were clear:

• Introduce customers to new products they hadn’t previously purchased, enhancing their shopping experience and exposing them to new brands;
• Suggest products based on past purchases and the preferences of other shoppers, helping customers discover items they may find relevant and interesting.

At the outset of the project, the team conducted a comprehensive analysis of customer interaction data collected across various channels. Specialists compiled a unified dataset, harmonizing inputs from both online and offline sales, and delving into the digital profiles of repeat customers and loyalty program members.



This meticulous groundwork enabled the identification of top-performing products for recommendation, while approximately 50% of low-demand items were excluded following an in-depth analysis. As part of the system’s development, Plavno experts experimented with multiple recommendation engine architectures. Chief among them were autoencoders and various collaborative filtering techniques, including neural network-based approaches. One of the key challenges was the high sparsity and inconsistency within the dataset, which required advanced solutions to ensure accurate and effective recommendations.

Plavno’s development team proceeded by segmenting users into clusters, identifying the most popular products within each group. They experimented with varying coefficient influences on the recommendation score — ranging from 30% to 80% — ultimately finding that a 50% influence yielded the best performance.



In the initial phase of real-world testing, users were offered recommended products at a 25% discount. The most effective recommendation models demonstrated a conversion rate within a 95% confidence interval of 9.5% to 10.4%. For context, random product suggestions generated a conversion rate of just 1.5% to 2.4%, while even the most popular products achieved only 3.3% to 4.1%.



When tested on historical data, the newly developed models outperformed standard “top popular” algorithms by 50% to 70%. In live user testing, the conversion gap widened even further, with the new system delivering three times the effectiveness. According to the team, this advantage stems from the fact that many users are unaware of certain products altogether — while popular items tend to dominate top positions in app listings by default.