Promoting products globally means being prepared not only for various sales and communication models, but also for various language issues. This is especially true if the campaigns, products and their searches are available in different languages. With Synerise this task is much simpler, thanks to our functionalities that enable users to search for products in different languages.
Product recommendation is a filtering algorithm that tries to predict and show the best possible products that specific users would like to purchase. We have a various types of AI recommendations, including similar and visually similar products, cross sell, basket recommendation and of course a personalized recommendation.
In this article we will focus on search examples, which will not be the easiest case because it is connected with search in different languages and regions in the same e-commerce platform.
Implementing a recommendation engine for a brand with multiple languages, regions and currencies
In this implementation, we have one e-commerce platform with different variants that depend on region chosen by the customers, currencies available on the chosen region and website language. Additionally, each market has its own product stock.
So, in fact we had several e-commerce operations under one domain.
Our customer wanted to have all data in one place and execute all campaigns from the one place as well.
In Synerise the place where you collect events, have your customer database and run the campaigns is called the business profile. And in such a case, customers can collect data from each market in a separate business profile, but it is possible to also collect all information from different markets in one business profile.
What obstacles did we encounter and how did we handle them? In the case of the customer database, we collected attributes - what language and currency the customer chose when he was on the website, to be able to effectively segment those customers later in purpose of any marketing campaign. But in case of AI recommendations, we had to have separate product feeds for each market.
A product feed is a source of product information for AI models. So, in fact, for each market, later separate campaigns were prepared which were built on a product feed dedicated to a specific region. This is especially important in case there are different product stocks for different markets.
How to do it - Setting up AI search
Step 1: Prepare product feeds for multiple regions with different product stocks
For any AI campaign the source of products from the website is the product feed, which contains all products and all information about them, based on which AI models and search are trained. As in this case all regions have different product stocks, so there is a separated product feed for each region, and in fact separate search and AI campaigns as well - thanks to that we can choose, based on which product feed we want to build our models.
Step 2. Prepare feed to handle 2 languages on 1 domain
In this case the customer can decide if he wants to browse an English or Arabic website, so for 1 region we had a separate product feed for those two languages. But our client had an additional requirement – the possibility to search an Arabic query on English version of the website, and vice versa. That’s why one product feed had to contain 2 languages - every attribute like description, title, type, category were added in both English and Arabic.
Step 3. Create search configuration and AI recommendation campaign
At the beginning you have to define what type of recommendation you want to have. In this case we used personalized recommendations on the landing page and similar recommendations on the product page without any additional filters added. A customer just relies on a what AI models recommend to the customer but of course, you can add a little bit of a human touch to these recommendations, and choose any additional filters you want, like brand, category, price filter, and so on.
In case of a search engine, there is more set up to do and the most important one is to prepare a search ranking. We indicate which attributes around the product have to be searchable and how important they are.
What is important, it can be changed anytime a customer would like, he can add new attributes, he can change them places, remove whatever he wants.
Additional options which make your ranking match your needs
At the top of this section we have more things that you can add to make your search more configurable.
- Ranking formula: lets you customize your ranking order of search results based on past page visits, past transactions or personalization.
- Preview: lets you test your search before activating it.
This is a crucial part of integration and has to be done with the help of the customer because he has the best business knowledge, and he can advise us on how what set up should look like. Read more about this case and how to implement it step by step >>
Step 4. Implement search via API
Our customer decided to have a backend integration with our AI engine, which means that using appropriate API methods and forms, developers responsible for e-commerce get results from Synerise and they decide where and how it will be shown to the customers.
Such a solution allows recommendations to be implemented in the most optimal way. Product frames can be shown faster, but if API integration cannot be prepared by customer, there is of course an option to prepare front end integration. We can inject the recommendation to any place on the website using our dynamic content module.
How exactly did it look in this case? As we mentioned before, each region has his own campaign and search functions, which have to be prepared in Synerise.
We built the list of AI campaigns and AI search indexes in Synerise Platform and based on this, the client’s developers set up the API configuration in a way that ensures that the appropriate campaign is shown on the appropriate region. They also decided in which places on the website the recommendation will be shown so they have implemented similar recommendation on the product page and personalized recommendation on the landing page as well as a 404-page, empty basket recommendation and so on.
Step 5. Enhancing search with additional features
When the search was implemented, the client could see what his users were looking for and which phrases they insert in the search. Based on this he could recognize opportunities to enhance search results using additional features.
Additionally, clients could use synonyms and query rules. Synonyms are an easy tool that you can use If somebody misspells a word, which can be automatically changed to the right word.
This is a really simple way to influence the search results, but you can do more using a query rules engine.
Query rules are something that allow you to influence the query that customers put in the search in more advanced way.
Its settings consist of three steps:
At the beginning we have to choose the conditions - what the customer has to put in the search. For example; the exact query or something that contains some specific words or starts and ends with something.
Later we have to define what has to happen if the query appears. We can change it for something else, replace the query, replace these specific words, remove them from the customer query and so on.
At the end we can schedule when this search has to work.
So for example, this client wanted to replace a query for sweatpants to pants and sweatpants together, because he does not have enough sweatpants in his offer and sometimes there is lack of them in the stock.
It is worth checking the most popular queries in our system and queries with zero results to avoid them and replace them with similar categories to be sure that your customers can always find what they want and do not see blank pages. This option also helps you in finding popular misspellings among your customers. It lets you prepare appropriate synonyms and query rules to avoid situations where a customer does not get any results.
At the end we would like to show you a comparison of two sets – one generated before using Synerise, and one done with us, to clearly see the effects.
Example results before and after Synerise:
The fact that AI models generally work better than people’s ideas does not mean that we do not take those opinions into consideration. Search configuration is a process which has to be done together with our customer because he has the most knowledge about his business. He knows his expectations, his needs and thanks to this we can achieve the best possible result. Of course, sometimes it really depends on the business and there will be different search setups in two separate businesses, because it all depends on the specific needs involved.