Time and time again, we've shown that product-specific messaging (messaging that mentions a specific product) outperforms generic messaging that doesn't.

Just look at this graph from our case study with Zalora, Southeast Asia's leading online fashion retailer:

This graph shows the purchase probabilities for different types of messages over time.

This graph shows how likely a user is to purchase a product from an abandoned cart notification.

See those blue lines that are outperforming all of the other lines? Those are messages that included specific product names (even if they weren't the exact products that were abandoned).

So, if we know that mentioning a specific product will improve conversions, why don't more eCommerce apps do it?

Because most apps have a CMS (Content Management System) that's full of product names and descriptions that were built for longer formats (i.e. to optimize for catalog inventory/categorization or website SEO):

The same content just doesn't work in short-form messages. It doesn't fit. Also, besides length, product (SKUs) names just tend to be really messy.

Here's an example of an actual product name for an app we were working with:

5-Piece Age Defying Get Started Kit Apricot Probiotic Cleansing Milk - 0.8 fl oz, Blossom + Leaf Toning Refresher - 1.0 fl oz, BioActive Berry Fruit Enzyme Mask - 0.5 fl oz, Goji Peptide Perfecting Cream - 0.4 fl oz, Resveratrol Q10 Night Repair Cream - 0.4 fl oz.

Total word salad, right?

...and this is far from a rarity. Messy product data is the norm for ecommerce apps:

Messy product names/descriptions are the norm.

So how can we make these long product names and descriptions more useful for short form messaging?

In the past we've just truncated long names—and truncation works fine (we've shown that here)—but it's not a solution that feels good for most ecommerce teams we meet.

So, a few weeks ago we tried using an LLM to clean/transform product names for a unicorn ecommerce app with a product catalog containing over 4 million different items, and it worked amazingly well:

Here are a couple of examples of how the LLM cleaned up these product descriptions:

✔️ An item name in the catalog started with "Roll over image to Zoom in", and then had 20 spaces before the actual item name. The transformation correctly understood what was going on, removed the leading text and spaces, and then cleaned the product name.

✔️ Another item was called "Pharaons.com - Paperback Belin Edition" followed by all of the product info like ISBN and language. The transformation understood that this was the French edition and summarized it as "Pharaons Paperback French".

✔️ In several cases, these descriptions had nonsense symbols between each word. The transformation dropped the nonsense and just kept the words.

...but we can't stop there.

Just having clean product names/descriptions doesn't solve the larger problem—You need to send the right product description to the right user.

So, how do we know which product to send to each user (Especially if we're dealing with a cold start issue, and/or we don't have the luxury of an abandoned cart event telling us which product(s) our user is interested in)?

Logistically, we solve this problem by integrating with your CMS, building your CMS content into your outgoing messaging, and then pairing this with our recommender system.

After doing the physical CMS integration, you:

Type a message just as you normally would:

Click on the shopping cart icon to add a placeholder for your desired CMS field:

Complete your message including a Greeting, CTA, Incentive, and other best practices:

Once your message template is completed, we add these messages to your content library which feeds our conversion and consumption-based recommender system.

Typical recommender systems operate more like "popularity contests." Our recommender system focuses on the likelihood that the item will lead to conversion (given the user's individual interests, preferences, and behaviors) as well as overall consumption patterns, which allows our system to be flexible and operate with a variety of different product catalogs of various types and sizes.

Views (vertical axis) don't always correlate with conversion probabilities (horizontal axis). In fact, most of the time they don't).

So, if you want to increase your relevancy and conversions with product-specific user messaging, reach out!

We'd be happy to show you how we've paired GenAI/LLM output with our AI-reinforcement learning-based messaging system to deliver clean and compelling product-specific messaging that get users to convert!