The "You might also like" block in most emails is built the same way every time: take the category of the last purchase and drop in products from that same category. It's better than nothing, but it's not personalization — it's simplified logic that's easy to mistake for real recommendations.
Neural networks change how precise the selection is. Instead of a single category, a model factors in dozens of signals at once: purchase history, browsing, time on site, and the behavior of similar users. The result is more precise than the standard rule of "bought a coffee machine, show them coffee."
Why standard recommendation blocks perform poorly
The "same category" rule ignores context. A person might have bought a coffee machine as a gift and never be interested in coffee again. Or they might buy a winter coat and immediately want gloves, not a second coat.
Simple rules give the same result to every buyer in a category. A neural network takes into account a specific person's individual behavior and compares it to similar users' behavior, so the selection ends up different even for buyers of the exact same product.
How a neural network builds a recommendation
There are two main approaches, and they're usually combined.
- Collaborative filtering. The model finds users with similar behavior and suggests what they bought or viewed. The logic is "people like you also liked this"
- Content-based filtering. The model analyzes the product's own characteristics (color, material, price, brand) and selects similar items based on those parameters
Modern recommendation services usually use a hybrid: collaborative filtering delivers precision on a large user base, while the content-based approach works even for new users with no purchase history.
What data you need for accurate recommendations
The quality of recommendations depends not on the model itself but on the data it's trained on. The more behavioral signals the algorithm receives, the more accurately it predicts interest in products.
Signal | What the model uses it for |
|---|---|
Purchase history | Which products a user buys regularly, which categories they prefer |
Product views | Which products and categories caught attention but didn't lead to a purchase |
Abandoned cart | Products with high purchase intent |
Seasonality and purchase timing | When a user typically makes purchases and which products are relevant during the current period |
Email engagement behavior | Which emails, categories, and offers spark interest |
If there isn't enough data, personalization becomes a formality. For example, without purchase and browsing history, the algorithm can't determine a user's real preferences and tends to suggest popular products instead of relevant ones.
Where to place personal recommendations in emails
- Post-purchase emails. 7–14 days after a purchase, based on what the person already bought
- Abandoned cart emails. The product from the cart plus one or two similar items
- Reactivation emails. A selection based on old purchase history that's relevant right now
- Regular newsletters. One block with a personal selection inside a regular email, without a dedicated campaign built around it
Canva recommends templates based on a user's past actions. Emails like this continue the product usage scenario and suggest the next logical step.
The recommendation is based on the user's behavior, not on popular offers. This approach increases the likelihood of a response, because the email continues a scenario that's already underway.
Tools to connect
- Ready-made recommendation services (for example, built on your own CDP or analytics platform) that integrate with your ESP via API
- A custom model, if you have a large volume of data and the resources to build one
- Automatic product card parsing in Letteros, which pulls current product data (photo, price, availability) into the email without manually coding each card
More on automatic card parsing
Common mistakes
- Showing a product the person already bought. A basic check that often doesn't get set up
- Too many cards in one block. Three or four products work better than ten
- A selection with no explanation. Users trust a recommendation more when they can see the reasoning behind it
- Ignoring seasonality. A winter coat in a July selection undermines trust in the entire block
- Launching without testing. Compare an email with and without the personal block on the same audience before changing the entire template
What to check before launching
- Purchase and browsing data is connected and updating
- Already-purchased products are excluded from recommendations
- The block has no more than 3–4 items
- Cards are up to date (price and availability match the website)
- There's an A/B test against a regular block with no personalization
The bottom line
Personal recommendations work when the model sees a specific person's actual behavior, not just the category of their last purchase. Collaborative filtering and the content-based approach together deliver a level of precision the "same category" rule never can. Without clean data and a test against a regular block, the effect of personalization stays a guess.
Read also: