Help › How recommendations work

How recommendations work

ChurnReturn compares content, not click histories. Products and pages are represented as vectors, and similar things find each other, even on the first visit and without a profile.

Which page is the visitor viewing?

1. By address: If the page address matches a product's deep link, the product is recognized. Tracking parameters such as utm_* are ignored, and variants are recognized via a pattern in the address.
2. By page text: Otherwise the script takes the visible text of the page, which we can narrow down to a specific area such as the product description, and searches for the most similar products. This is how guides and category pages work.

The recognized product itself and its variants are not recommended; out-of-stock products can be excluded on request.

Fields for vectorization

For each product, a text is built from the selected fields and converted into a vector. Under Feeds, Settings, “Fields for vectorization” you decide:

• Which fields: By default title, brand, category and description. You can add your feed's attributes, such as color or material.
• Weight 1 to 3: A field with weight 2 counts twice. The title is set to 2 by default.
• Prefix: a word before the value, such as “Brand:”, which helps the model classify the value.
• Character limit: shortens long descriptions, 1000 characters by default.

Every change re-vectorizes the affected products immediately. For large feeds this takes a few minutes; recommendations keep running on the previous state in the meantime.

What makes good recommendations

• Meaningful titles and descriptions. “Item 4711” doesn't help anyone, “Ergonomic office chair with lumbar support” does.
• A well-maintained category, because it groups products by topic.
• One feed per language, so visitors get products in their language.
• Don't vectorize everything: prices and product IDs usually dilute the similarity.

Languages

The model understands more than 100 languages. The language of a page is detected from URL rules such as /en/, from the product language or from the text; the overlay texts come from the matching language template.

Next: Reading the statistics →