The e-commerce boom makes online environment more competitive. Internet retailers seek competitive advantages and a personalized experience for their clients is one of them. To achieve truly personalized experience, one has got to know individual users and learned their habits. With an increasing number of items, growing number of users and changing environment, it is not tractable to prepare personalized experience manually.
Recommender systems can be used to personalize the content of websites for each visitor individually. Other channels such as e-mail newsletters or mobile notifications can be personalized as well. User interactions from multiple channels feed a recommender system, increase the precision of recommendations and improve the personalized experience of users.
Most of the big companies understood the value of personalized recommendations and hired large dedicated data science team to develop and maintain their recommender engines. Some companies (Uber, AirBnB, Amazon, Spotify, Google, etc.) focus on their data so intensively that they can be treated as data science companies.
Before you decide to build and maintain your internal recommender engine, you should consider costs and benefits of such decision.
Costs of having a recommender system
Data are crucial for good recommendations. You have to start with profiles of your users and attributes of your products. You can start with anonymous visitors of your website and remember their interactions. Once you learn their identity, you should merge profiles over all your channels. The most valuable data for recommendation is the history of your users and their interactions with your items. For new items, you should provide at least some attributes (e.g. text description) to improve recommendation. Storing and maintaining such data is not cheap, but you can utilize it in many other ways (e.g. RFM analysis of your customers).
When you decide to develop your internal recommender, you need the team of 2 FTE data scientists to make prototypes, 1–2 programmers that make production ready code and at least two engineers preparing and maintaining the infrastructure. The development phase takes typically 2–3 years and costs around $1.5M depending on the cost of your staff and HW resources. The maintenance costs (infrastructure, HW costs) for a medium-sized company are around $20k/month, not taking into account data scientists that are not cheap, but you need them to maintain the quality of algorithms and improve the system. You can save most of these expenses by using external recommender system, such as the one provided by Recombee.
Benefits of recommender systems
Let us start with most general benefits. With a recommender system, you gain complex insights into your customer and product bases. Profiles of users maintained in the recommender are based on their past interactions with items and enable powerful analysis with business reports and dashboards generated on a regular basis. Such reports can predict possible problems so you can avoid them. Your business decisions influenced by analytics can save you money.
Even more powerful are “smart decisions” taken by the recommender engine every millisecond. Generated recommendations typically reduce the time required to find an item and significantly increase the probability of discovering other items of interest. The result is increased loyalty and satisfaction of your users with your web services. Typically, users also interact with more items and this behavior leads to increased consumption and higher profits. Also, newsletters, personalized promoted content and push notifications encourage users to return, increase the frequency of visits by regular users, reduce churn and increase their lifetime value.
We work with our clients and potential customers to evaluate and maximize the business value of personalized recommendations. This is however not easy because companies pursue different goals and objectives. Many companies define and evaluate their key performance indicators (KPIs) on a regular basis simplifying the exact measurement of recommender impact. Such measurement is then realized by AB test, where personalized recommendations are provided to users in a group A whereas group B gets standard recommendations or best-selling content.
The domain, where benefits can be evaluated seemingly easily is the e-commerce. Measuring income generated from personalized recommendation can be realized by a token or simply by computing price of purchased products that were recommended less than a few minutes ago. However, this number does not reflect the fact, that the customer might buy different product otherwise, or he can make the purchase even without personalized recommendation.
Personalized recommendations as a service
As indicated above, costs of a recommender system can be significantly lowered by outsourcing the recommendation service. Moreover, the quality of internal recommenders is often poor, because the team does not have sufficient capacity/knowledge to develop state of the art recommendation algorithms in scale. It is better to focus on your business, integrate personalized recommendations and customize it with the query language to get the full potential out of it. Building smart data products on top the recommender is your next goal. Having good performing and affordable recommender system becomes a must.
Some of our competitors charge small share of total sales volume generated from personalized recommendation. This model can, however, prevent clients to use recommendation in positions generating high volumes even without personalization.
Fair revenue share model involves running continual AB tests in all positions and channels where the personalized recommendation is deployed. This, however, decreases possible revenue and generates additional IT costs for clients.
Our preferred model that reflects costs of the recommendation service provider is the fixed fee or pay-for-recommendation model. The more recommendations you consume, the cheaper it gets. This model is profitable for the client even for high volume mailing campaigns, where the conversion rate is obviously low.
Learn more about the pricing at our website or contact us to discuss your business case.