Examples of Personalized Recommendations in eCommerce
People often use the words ‘recommendation’ and ‘personalization’ interchangeably.
To start with the barest distinctions in the e-commerce related definitions of these two terms, ‘personalization’ is an extended category within e-commerce website optimization and it is also applied in the field of product recommendations too. Product Recommendations can be personalized (with user behavior profiling) or non-personalized (using the data mass of item attributes and other purchases).
Sometimes the personalized, sometimes the non-personalized recommendations entail more conversions, therefore professional recommendation engines have to make thousands of decisions in every second: ‘does this customer have enough history here to get personalized offers that might imply higher probability for conversion or shall we ignore the user’s profile and apply general item-to-item recommendations?’- This decision type (fallback scenario) is the most often used by recommendation engines. If the engine works in a very data-rich environment (enterprise-sized e-stores with ten thousand of buyers) in a given second plenty of personalized, non-personalized and hybrid scenarios compete each other. These can be based on the given user’s history, the similar users’ histories, product sales histories etc., using the results of AB tests and the calculations of data scientists.
Standard logics: the buyer has a thin profile or has no profile at all. Common recommendation logics can be authored, (most trending, freshly arrived, recently viewed etc.) these recommendations are often made manually.
Advanced logics: the buyer has a meaty profile with sound history, the product recommendation she is given based on meticulous calculations. Or the product pool is very rich, traffic is heavy, there are many buyers, item-to-item logics can be set up and tried. To launch these recommendations, you need a recommendation engine provider.
Recommendations containing personalization are more likely to belong to the advanced recommendation logics. Let’s have a look on the personalized recommendations with a walkthrough on 4 prominent locations: Onsite widgets, Popup, Messaging and Offline-to-Online.
Onsite Recommendation Widgets
Main page
The visitors of our main page should have already made some footprints in the e-commerce store history, if we want to display personalized recommendations for them.
First-time visitors are not assumed to look for something specific, the purpose of main page recommendations for them serve the goal of informing the customers and getting them engaged. For first time visitors the recommendation engines usually show item-based offers (latest deals, most trending, freshly discounted, featured etc.). After the visitor made several movements (clicking on an item or a category like ‘teakwood tables’) and after returning to the main page the recommendation engine is able to present the first ‘recommended for you’ widgets that can contain similar or matching products or item-hierarchy based accessories (I clicked on a rowboat the engine offers a pair of oars).
Returning visitors: after the user gave consent that cookies can deliver information to the store and the recommendation engine provider — the cookies will store plentiful user history, that can used to trigger personalized offers across all pages.
Displaying new products for a profiled visitor: featuring your e-store’s new products and collections in a general way for a newcomer customer is obvious, but a returning user needs fine tuning. For example, if your customer has bought a long-term product (a road bike), she’s not likely to buy that category again in the next year (drop the road bikes from her recommendations) but all the new products can be mixed with the accessories that might match to the already bought product (offer seats matching her taste).
Seasonal offers, Daily deals, Rating-based recommendations for profiled visitors: these seasonal and impulsive categories also can be beefed up with recommendations based on our customer’s previous year’s activity.
Daily and weekly deals generally build trust in the visitor (representing a well-groomed and maintained e-store) especially when algorithms filter the daily deals to display the user’s preferred items.
Product page
Product information page is the place where visitors find detailed info packages about the product. A significant number of visitors arrive here from non-direct traffic (PPC campaigns, price aggregators, ad displays on affiliate sites). The objective here is to keep the visitors in the sales funnel, and display the most relevant items to them, stretching their purchasing willingness (with recommending accessories and launching upselling efforts).
You can try many different combinations here, apply easily detectable widgets, mix their content from item-based and user-based recommendations.
Conduct experiences, Run ‘ABC multivariate’ tests
You can test different recommendation algorithms against each other on your product pages. Either do this by adding more placements or split your traffic and recommend using different logics in one widget.
You can try different approaches on your product pages like:
1. ‘Customer who bought/viewed this…’ — item-to-item collaborative filtering (This logic is based on the relation of items by looking at how many times they’re present together in histories of other users)
2. Recommendations based on the user’s preferences, without collaborative filtering.
3. Hybrid: ‘Customers who bought/viewed this…’ collaborative filtering mixed with the user’s preferred brands.
To launch these experiments, you have to possess large amount of user data. If you don’t have enough, set up your test with a simplified version with 2 components. Depending on your product catalog and category structure, for that you may reach out to a recommendation-as-a-service provider that offers customization like these to see how it works for you.
Cart page / Checkout page
Checkout / Cart page recommendations find the buyers in a very favorable, elevated emotional state, where they are about to complete the purchase, therefore the likelihood of them saying yes to a 5–10% higher priced product is more. Recommending similar products or accessories here can result with higher average order values.
Checkout page is an excellent place to launch ‘also bought’ or accessories-type recommendations. Once the users get to the Checkout Page, we already have plenty of information about them — as they have at least a 3–4 steps long customer journey — on the checkout pages the recommendation types must be formulated by their preference.
Examples of information already gained that can be used by recommendation algorithms that can be used for personalization:
- The products they viewed — Recommend items that the previously viewed and left, and other similar items
- The products they (previously) purchased or added to cart — Offer products that other shoppers also purchased or that are connected to their previous purchases
- Cart value (and the difference if you offer free shipping from a certain basket size) — Remind them and recommend product to qualify for free shipping
- Their preference for product categories, price and different product attributes (like color, brand) — Show them other similar items that are popular/on sale from their preferred categories and product attributes
Category page
How to launch product recommendations on the category page? After you have some information about the users you can start experiencing with same sets of recommendations as we suggested in the product page section.
The dilemma of “Exploration or Exploitation”
On the category page we often face two conflicting tactics, what we can both focus on is: shall we utilize the exploration or the exploitation? Shall we help the customer to maximize the capacities of the category container, and tuck all the relevant products in there? Or help the customer to explore other products that do not belong to the category but have strong relations to it? (example: ‘Customers who viewed this category viewed these related products too’).
Onsite pop-up
The same rule applies at the recommendation widget pop-up windows as the pop-up windows generally: it must not irritate the customer and it must not increase the bounce rate. Before applying this, ask ourselves: Can’t we substitute it with proper recommendation widgets? Should we block the natural customer journey here with a pop-up? Run AB tests, get the plain recommendation widgets and the pop-ups competed, and examine the buying process in its whole span: some pop-ups may increase the bounce rate on the given page, but in the end the conversion could be higher as well.
Messaging
Once we have gathered behavior data on the course of visits and purchases online, the email campaigns should be based on those to boost our revenue. Email can be considered as an extension of your online store — you can recommend items almost the same way as you can onsite. Personalized e-mail recommendation strategies vary considerably from user to user: for example, one of your visitors might have been interested in a product very much — you can author personalized, recommendation-based email campaigns, the data is easily available about:
- Each of your customers.
- Each of your visitors — non-purchasing visitors should be logged in to become trackable by their email addresses. So, you have to grab them onsite first.
Facebook Messenger
For e-commerce stores interconnected with Messenger 2017 was the year of embracing the quicker response policy: the consumers appeared to show less patience to wait on the phone or await email responses.
Messenger-displayed product recommendations can be considered as a vital and more and more important category of personalized recommendations if the user data is collected on a passive way (recording user behavior onsite and importing the onsite-recorded user behavior into the messenger app, where, after recognizing the user, launching recommendations on him). This depth of data migration and identification is a ponderous work, and the users are more and more protected by the laws of data processing. The results (relevant recommendations in our messenger flow) still look strange and evoke uncomfortable distrust in the user: ‘are they spying on me in the e-store and sell my data to the Facebook?’ The sponsored offers occasionally popping in our messenger stream still show way lower relevance, which makes us think that this data stream is still very feebly applied in the world of e-commerce. The prominent user preference mapping still comes from the active channel: the bot asks the user, and recommends the solution, which is somewhere halfway between the ‘search’ and the ‘product recommendation’ function.
With chatbots we can redirect the customers to the appropriate pages where the possibility of sale is the highest, or the recommendations can be displayed right in the messenger window, however the scarcity of space allows only one or two products plus initiated conversations. Recommending 4+ products in one block in a chat window can be a bit a bit too much and become counterproductive.
When you get chatbots to recommend products, ‘inchat recommendation’ or product link are the options, or you can experiment with sending filtered search links too.
Offline-to-Online
Connecting online and offline snippets of identities is a challenge, a jigsaw puzzle work, where Yusp pioneered last year with a well-benchmarkable solution. Using loyalty card has long been in Cora Romania’s retention marketing arsenal: in early 2017 the retail giant looked for means to utilize and leverage the 6 years of customer data accumulated in their Loyalty Program. The RFM-segmentation Yusp provided with its personalization engine generated unique and tailored promotions for each individual loyalty card holders. The blended technology Yusp implemented for Cora is generating 7500 additional store visits / week with an average 24 USD basket value.
How to choose recommendation logics?
Although every user and eCommerce store is different, there are recommendation algorithms that are already well tested by most recommendation engine providers and come ‘built-in’ for you. Testing the above mentioned standard and advanced methods for yourself, mixing with your unique approach and insight can help you understand more the user behavior and ultimately increase your sales.
Originally published at www.yuspify.com on August 23, 2018.
