If Your Products Could Talk, What Would They Say to Each Other?

In this article, we'll explore how Product2Vec transforms product data into numerical representations, what sets it apart from classical recommendation systems, which business problems it solves, and what results it delivers , backed by real-world e-commerce cases.

Product2Vec and the Next Generation of Recommendation Systems in E-Commerce

"Our customers already see recommendations , why do something different?" Every time I hear this in a meeting, my answer is the same: Because classical recommendation systems tell your customers "people who bought this also bought that." Product2Vec, on the other hand, can answer the question "why do these products make sense together?"

The difference is as significant as browsing a "bestsellers" shelf versus having a knowledgeable bookseller say, "If you liked that, you'll love this , but from a completely different genre."

The Hidden Problem with Classical Recommendation Systems

Imagine you're browsing a red dress on an e-commerce site. What does a classical Collaborative Filtering system recommend? Most likely, other red dresses. Maybe other items from the same brand.

But what if you knew that most customers looking at that red dress were preparing for a special occasion , and were also searching for heels, a clutch bag, and a statement necklace?

This is the biggest blind spot of classical systems: they recognize products through categories and tags, not through customer behavior.

What Is Product2Vec?

Google's famous Word2Vec algorithm learns the meaning of words using a simple principle: "You know a word by the company it keeps." If the word "Istanbul" frequently appears alongside "city," "Bosphorus," and "traffic," the algorithm understands what Istanbul is from that context.

Product2Vec applies the same logic to products: You understand a product by the other products viewed alongside it in the same session.

Say a user views the following products in a single session:

Session 1: [Running Shoes] → [Sports Leggings] → [Sports Bra] → [Water Bottle] Session 2: [Running Shoes] → [Running Socks] → [Sports Wristband] Session 3: [Hiking Boots] → [Outdoor Backpack] → [Water Bottle] → [Sunscreen]

Just as sentences in Word2Vec reveal relationships between words, sessions here reveal hidden relationships between products. The algorithm learns that:

Running Shoes and Sports Leggings live in the "same world"

A Water Bottle bridges both the running and outdoor worlds

Hiking Boots and Running Shoes are close, but not identical

Each product ends up represented as a "point" in a multi-dimensional space. Points that are close together = products that are similar or complementary.

What Changes on the Business Side?

The real power of Product2Vec lies in solving multiple business problems from a single embedding (numerical representation). From one model, you can derive:

1. Similar Product Recommendations The most fundamental use case: making the "You Might Also Like" carousel on a product page smarter. The difference from classical systems: it can recommend not just same-category items, but behaviorally similar products. (Direct increase in CTR and conversion rate.)

2. Cross-Category Cross-Sell Product2Vec is built to cross category boundaries. Someone browsing a coffee machine can be shown a grinder, coffee beans, and a milk frother , because these products are close in the embedding space. (Increase in average order value (AOV).)

3. Automated Product Taxonomy Expansion Product categorization is a nightmare in marketplace models. Using product embeddings, you can automatically detect miscategorized products or route new products to the correct category. (Improved search quality, consistent customer experience.)

4. Personalized Search Results A customer searches for "pants." A classical search engine returns every type of pants. But if that customer has been browsing running products in recent sessions, Product2Vec-based personalization shows running pants first. (Increased post-search conversion rate.)

5. Customer Segmentation By aggregating product embeddings at the customer level, you can map each customer's "interest space." Segmentation is then based on real behavior, not demographics. (More precise campaign targeting, better marketing budget efficiency.)

Real-World Numbers

Does this approach actually work in practice?

Pinterest reported meaningful improvements in add-to-cart and purchase metrics , especially for low-traffic products , using its embedding model called ItemSage. [1]

Companies like Amazon, Alibaba, Walmart, and Criteo actively use product embeddings across recommendations, pricing, and similarity use cases. [2]

Huckberry, an outdoor e-commerce brand, achieved a 9.4% revenue increase without increasing its ad budget after integrating an AI-powered personalization engine. [3]

According to data shared by McKinsey, 35% of purchases on Amazon are driven by recommendation engines. On Netflix, that figure rises to 75%. [4]

What Do You Need to Get Started?

You don't need a multi-million dollar budget or a 50-person data science team to start a Product2Vec project. Three essentials:

1. Session-based view data A record of which products each user viewed, and in what order, within a single session. If you're using GA4, this data is already waiting for you in BigQuery.

2. Sufficient data volume Word2Vec-based models can't produce quality embeddings for rarely viewed products. The general rule: a product needs to have appeared in at least a few hundred distinct sessions.

3. A clear use case aligned with the business The technology is ready , the real question is "where will we use this?" Similar product recommendations? Cross-sell? Personalized search? Keeping the first project narrow and measurable increases the chances of success.

Limitations to Keep in Mind

Every powerful tool has its limits. Key considerations for Product2Vec:

Low-traffic products (cold items) produce poor-quality embeddings. New products require additional signals such as category, brand, and price range metadata. Approaches like Meta-Prod2Vec focus on addressing this.

Seasonality and trend shifts make the model stale. Embeddings need to be retrained at regular intervals.

It's not sufficient on its own. In practice, the best results come from combining embedding-based recommendations with rule-based filters (stock availability, margin targets, promotional rules).

Conclusion: Learn the Language of Your Products

The core idea behind Product2Vec is deeply human: context creates meaning.

We understand a word by its place in a sentence. We know a person by who they spend time with. In the same way, we can "understand" a product by looking at which other products it's viewed alongside.

This understanding doesn't just give you a better recommendation engine. It gives you a map of how your customers see your products. And that map can transform everything , from your marketing strategy to your product development decisions.

Your products are already talking. The question is: are you listening?

FAQ

How much technical infrastructure do I need to implement Product2Vec?

A basic Python environment and session-based view data are a solid starting point. However, if you want to deploy this in a scalable, sustainable way that's tied to real business outcomes, that's exactly where Analyticahouse comes in , we offer end-to-end partnership from data infrastructure and model development to recommendation system integration and performance measurement.

How many products and sessions do I need for meaningful results?

As a general rule, a product needs to appear in at least a few hundred distinct sessions to produce a quality embedding. On the total session count side, figures in the hundreds of thousands help the model learn meaningful relationships.

How often should the model be retrained?

Since seasonality and trend shifts cause embeddings to go stale, we recommend setting up a retraining cycle somewhere between weekly and monthly, depending on how dynamic your product catalog is. For fast-moving categories like fashion and electronics, shortening that cycle leads to more accurate results.

What should be done for newly added products?

New products suffer from the cold start problem since they haven't been viewed enough. In this case, approaches like Meta-Prod2Vec , which estimate embeddings using metadata such as category, brand, price range, and description , can be brought into play.

Is Product2Vec sufficient as a standalone recommendation engine?

No. The strongest results come from hybrid systems that combine embedding-based recommendations with business constraints like stock availability, margin targets, and promotional rules. Product2Vec should be positioned as a powerful signal source, not a decision engine.

References

[1] Pal, A. et al. (2022). ItemSage: Learning Product Embeddings for Shopping Recommendations at Pinterest. KDD '22, ACM SIGKDD Conference on Knowledge Discovery and Data Mining. cs.stanford.edu/people/jure/pubs/itemsage-kdd22.pdf

[2] Coveo (2022). Product Embeddings & Recommendations: The Ultimate Ecommerce Power Couple. coveo.com/blog/product-embeddings-conversion

[3] Core dna (2025). Top eCommerce Case Studies: Key Insights and Surprising Results. Huckberry, Algolia AI personalization engine integration case study. coredna.com/blogs/ecommerce-case-studies

[4] McKinsey & Company, cited in Coveo (2022). Amazon and Netflix recommendation engine impact data. coveo.com/blog/product-embeddings-conversion