Olist Retail: Customer Experience, Retention & AI Sentiment
Retail (Olist, Brazil)
Who are the valuable customers, does a late delivery cost you a good review, and what does an AI sentiment layer on the review text add?
Olist is a Brazilian marketplace connecting small sellers to major online storefronts. This project analyses about 99,000 real orders placed between September 2016 and October 2018, answering questions a Retention Lead, an Operations Manager, and a Category Manager would each genuinely ask.
Findings
Repeat customers are rare but valuable, and revenue still depends on one-time buyers. Only 3.0% of customers ever place a second order. Repeat customers are worth almost twice as much individually (308 BRL average lifetime value versus 161 BRL for one-time buyers), but 94% of all revenue still comes from people who never come back.
A late delivery is the single biggest driver of a bad review. Orders delivered after the promised date average 2.57 stars, with 54% landing at 1 or 2 stars. On-time or early orders average 4.29 stars, with only 9% that low.
The AI sentiment layer on 2,500 real Portuguese-language reviews finds a real gap between text and star rating. Direction agreement is 67.8%, and 16.9% of 4 and 5 star reviews contain text the model reads as negative, a hidden-complaint signal the star rating alone misses.
Live dashboard
Full write-up
- Full report — methodology, limitations, and recommendations
- SQL and notebook on GitHub