Meesho
Medium
How would you design the product ranking on Meesho?
This is a ranking design question. Not a recommendation buzzword question. On Meesho, the hard part is not relevance. Thousands of near-identical, unbranded products look the same, and some of them come back as returns.
Step 1. Clarifying Questions
- Which surface? Search results, category listing or home feed? Assume search and category listing.
- What is the goal? Orders, GMV or satisfied buyers? Assume orders that are delivered and kept.
- Who are the users? Assume value-seeking shoppers, many in smaller towns, many new to online shopping, often paying cash on delivery.
Summary: design how Meesho ranks products so buyers order things they keep, and good sellers get found.
Why it matters: Meesho's promise is low prices. A cheap product that is returned costs two shipping legs and the buyer's trust. A COD order refused at the door (RTO) hurts most.
Step 2. Stakeholders
- Buyers: the right product at the right price, quickly, and trust in it.
- Sellers: visibility, especially new and small sellers.
- Meesho: delivered orders, low returns, a healthy seller base.
Step 3. The pipeline
1. Recall: fetch a few thousand candidates. Text match plus image and text embeddings, because users search in Hinglish or with vague words like "party wear kurti".
2. Ranking: score each product on predicted purchase × predicted keep. Signals:
- Relevance to the query
- Price vs similar products
- Product quality: rating, return rate, return reasons (size, quality, not as shown)
- Seller quality: dispatch time, cancellation rate, RTO rate
- User fit: past price range, size, style
3. Re-ranking with rules:
- Collapse duplicates. The same photo uploaded by 20 sellers shows once, with the best seller.
- Reserve a few slots for new sellers (cold start).
- Keep a spread of price bands in the first 20 results.
Step 4. The trap
Rank on clicks or orders alone and the model learns to love flattering photos at the lowest price. Those products have the highest return rates. The objective must include the return.
Step 5. Metrics
- North Star: delivered, non-returned orders per search session
- Ranking quality: CTR, add-to-cart rate, conversion per search, position of first click
- Quality: return rate and RTO rate of top-10 ranked products
- Guardrails: share of orders going to the top 1% of sellers, new seller first-order rate, zero-result searches
Interview tip: on a value marketplace, put "predicted return" inside the ranking score. That one choice shows you understand Meesho's economics, not just ranking.
Community Answers (0)
No answers yet. Be the first to share your approach!