Recommendations
In enKod, all recommendation algorithms can be divided into three groups: product-oriented, user-oriented, and self-driven.
For algorithms that focus on or depend on user behavior (Personal Recommendations, Cross-sale), we collect preference information on your site or app, namely:
- Product or product category views;
- Order content.
For product-oriented algorithms (Similar Products by Text, Similar Products by Image), we use your product catalogue for similarity in parameters between different products.
Self-driven algorithms (Popular Products, Popular Products in Category, New Products, Bestsellers) are calculated based on the history of orders and views on your site.
- History of product views (opens) for the last 60 days (for Popular)
- History of orders (purchases) for the last 60 days (for Bestsellers)
User-oriented algorithms analyze the purchase and viewing history of a particular user or all users on the site. Based on these attributes, products are selected that may also be of interest to a person.
Product-oriented algorithms analyse the interaction of all users with the products on the site and suggest the most viewed or most purchased ones.
You can learn about the details and results of implementing the enKod recommendations block in the cases of our clients
• how a large book publishing house doubled its click-through rate
• how a store of branded clothing and shoes increased profits from the site by 6.6%
Algorithms and display locations
In enKod you can use the following algorithms:
- Similar products by text
- Similar products by image
- Popular products
- Popular products in a category
- Personal recommendations
- Cross-sales
- New Products
- Bestsellers
Logic of algorithms
Similar products by text
The algorithm compares the names and descriptions of products from the feed and mathematically calculates similar products (similarity of which is more than ~60%). The better the description and title, the better the recommendations will be. Similar products in the text will always be from the same category as the main product, because the goods of the same category are always maximally similar to each other by the parameters taken into account. The block of recommendations displays the most similar products in random order.
Similar products by image
The algorithm compares vectors of images from the product feed and mathematically calculates similar products.
Popular products
The algorithm analyzes views of all products on the site for the last 2 months, selects the top, comparing it with the product feed (recommended products must be in stock). The selected product top is sorted by the value of openings (from higher to lower). The number of the most viewed products, which is set in the block, is output. When filters are used, they are applied to the top.
Popular products in categories
The algorithm analyzes views of all products on the site for the last 2 months, selects the top by comparing it with the product feed (recommended products must be in stock). Recommendations are issued one by one from each category selected when setting up the block, so that the selection includes products from all categories.
Personal recommendations
The algorithm analyzes all product views on the site for the last 2 months, selects the top by comparing it with the product feed (recommended products must be in stock). For each specific person, the history of his browsing on the site by category is analyzed, and a personal top is formed. Putting one thing with another, the most suitable products are displayed (2 products from each top category). When filters are used, they are applied to the algorithm output.
Cross-sale
The algorithm analyzes all orders containing more than one product. For each product or category, based on the history of all orders, a top is generated containing the products that are most often bought together. This top of matching products is sorted by the percentage of chance of appearing in an order with the product you are looking for (the chance must be more than 85%). Out-of-stock items are excluded from recommendations. This algorithm is highly dependent on order history and can produce unpredictable results if there is not enough data on your account.
New products
The algorithm gives out products that were added to the feed last. Recommended products must be in stock, otherwise they will be excluded.
Bestsellers
The algorithm analyzes purchases of all products on the site for the last 2 months, selects the top, comparing it with the product feed (recommended products must be in stock). The selected top products are sorted by the value of orders (from higher to lower). The number of the most purchased products, which is set in the block, is output. When filters are used, they are applied to the top.
The available display locations
To set up recommendations, the available display locations on the site are:
- Home page
- 404 page
- Account
- Cart
- Product card (selection of products of specific categories and subcategories is available)
- Category (selection of products of specific categories and subcategories is available)
If your site has pages that are not included in the list above, you can place recommendations on them as well, but the following rules must be observed: each algorithm belongs to one of three types - either tied to a product (similar products, cross-sale), or to user behaviour (personal recommendations), or not tied to any of the above (popular products, popular products in categories, new products). Each algorithm is allowed to be placed only in the place that corresponds to the logic of the output.
For example, you can't use the “Similar Products” algorithm on the 404 page, because there are no products on it that can be matched with similar ones. At the same time, it is possible to use “Personal recommendations” on a 404 page, because this algorithm is based on user behaviour and is not tied to a specific product.
Prioritisation
Within one block of recommendations you can use up to two algorithms, between which there will be prioritisation. This is necessary in order to fill the entire block with products, even though the first algorithm may not pick up the required number of recommendations. For example, you want to use the personal algorithm, but for some visitors to the site has not yet formed a sufficient number of recommended products, because the client is new and has not yet shown enough activity on the site. In this case, you can put the popular products as the second priority. In case of lack of personal recommendations, the recommendation block will be filled with the most viewed products. Not all algorithms need to be supplemented: popular, popular for category and new items are self-sufficient, as they do not rely on user behaviour and always have enough positions to be displayed.
In addition, after personal recommendations, similar and related products should not be used as a second priority, as this may cause inconsistencies in the name of the recommendation block on the site and in the output.
Limitations
When combining algorithms in a block, it is also necessary to comply with the conditions on the places of display, which were described above. I.e. if you want to place recommendations on the cart page, then both selected algorithms must be suitable for placement in this place. You don't have to worry about customisation, as the system will warn you if you try to combine mismatched algorithms and display locations.
Recommendations in email messages
Setting up recommendation blocks in email messages is almost no different from the site. The same combination restrictions apply to emails. The only difference is that you don't need to choose display locations, you only need to watch out for combinations of algorithms between each other: you can't use an additional algorithm for new and popular and similar and related algorithms in the second priority after personal ones.
Read more about dynamic content methods for substituting recommendations into emails in this section of the knowledge base.