At Hello Retail, we have been working on a fundamental change to how our solutions operate, and we are finally ready to share more about it. If you haven’t already, I recommend you read the introductory post of our CEO about the new direction. First, a little background information.
Changes to the market
We have all seen the increasing change in the market where users are becoming more privacy-aware. The need for privacy can stem from many reasons, one probably being the dominant social media players who have historically abused data collection. In addition to the shift in consumer sentiment, lawmakers took a privacy stance by introducing GDPR.
Browser vendors accommodated the change by removing support for 3rd party cookies. In turn, it became harder to track users across domains and the decreased cookies’ lifetime affected how long you could identify each user.
There is no reason for the user sentiment, lawmakers, or browser vendors to take a less firm stance on user privacy – it is here to stay.
While users do not want to get tracked, many still expect personalized recommendations that deliver relevant product suggestions.
With these changes affecting the market, there are two ways you could go about it: play by the rules, or play cat and mouse with users’ requests for privacy. The latter could be through fingerprinting, sometimes referred to as cookieless tracking. Remember that fingerprinting or using cookieless tracking would not get you around the Cookie law. Customers still need to accept your cookies for you to track them. The Cookie law is not about the technical measure to add a cookie but about tracking individuals, even if done by other means than utilizing cookies.
At Hello Retail, we have decided to go for the former option and respect customers’ wishes and concerns for privacy. We are building upon our vision to provide a completely anonymous solution that continues to personalize the eCommerce experience.
Personalization
Solutions Today
Hello Retail
Historically, eCommerce personalization relied on behavioural tracking, which works wonders when customers allow for it – but we found a better way. We believe true magic happens when you understand how products correlate and enrich this knowledge with user intelligence. This means we have a fully fledged privacy-aware solution, where, when allowed to, we can sprinkle with a layer of behavioural data.
Product Intelligence Engine & vector databases
The Product Intelligence Engine, is one of our underlying technologies that helps understand products and how they correlate using advanced algorithms supported by an advanced vector database.
There are many types of product relations, the most common being:
- Products related to each other
- Products that go well together (are being bought together)
- Products’ purchase order (what will people buy next)
Understanding the different product relations lays the foundation for providing a personalized experience based on product knowledge instead of behavioural data only. More about this in just a bit.
As previously mentioned, an enormous vector database is at the heart of the Product Intelligence Engine. The database excels at finding similarities across data points in a multidimensional space.
When loading new products into our system, we first transform them all into vector embeddings using multiple AI models. Once represented as embedded vectors, we look at the product vectors in many dimensions to identify similarities and map their purchase relations.
Purchase relations
A simplified example is presented in the adjacent image. Related products are products that have a short distance to the product you are looking at. In the example image, you can see the the different shoes that are circling the primary shoe (the large one in the image). The shoes circling the main product could be variations or similar products.
It is important to note that when building the embedded vectors, you cannot just compare SKUs when trying to understand similar products due to two main reasons:
- Not all product feeds contain SKUs/EAN numbers
- It will only show products that are exactly alike
- Click image for details
The second point is not ideal, as products don’t have to be identical to share characteristics. An example is if you buy an iPhone 13 with 512 GB of storage, it will have one SKU, but another iPhone 13 with 256 GB of storage might have another. For most use cases, these should identify as similar products. There are two key benefits to doing that:
- These products become alternatives to each other, as their attributes are similar.
- We can utilize the bought-together data better from both to recommend better upsell products. The more data you have in the purchase relations, the better they perform. In this case, the cover doesn’t care how much storage the iPhone has, and by using similarity to combine the purchase patterns, we can support this with more accurate upsell suggestions.
Making the Product Intelligence Engine Global
If we just looked at the product data on each shop, many would not have sufficient data points to provide insights for the AI to get smart enough to capture the correlations optimally. Luckily by understanding related products globally, we can also find purchase relations on a higher level and utilize this intelligence across all our customers for their collective benefit. It also means that as more customers join the platform, the solution becomes better and more intelligent for everyone.
Disclaimer:
As a Hello Retail customer, you own all your data, and you can, at any time, request your data to be deleted. We guarantee that there is no personally identifiable information, or pricing information shared across clients. Products in our system are stored as coordinates in a vector database. The vectors are mapped with other vectors to find, e.g., what has been bought together. All vector embeddings are hashed, which means that even the products are anonymous, as you cannot recover a product from its vector representation.
- Click image for details
With Global Product Intelligence Engine, it also means that you will see much better results from your language-specific sites than if you were using a traditional personalization service that only relies on behavioral tracking. The reason is that the conventional vendors will not have enough data on your smaller language sites to present satisfactory results. In addition, our Product Intelligence Engine identifies the different products across your shops, which means that your primary shop will help provide good purchase relations to e.g. your language domains.
Product Intelligence can transform the prediction of user affinities
The Product Intelligence Engine was not only built to empower our recommendation solutions- we also use it to predict user behavior and utilize that throughout all solutions we provide.
When Hello Retail came to life, our tracking solutions accomplished their purpose with 3rd party cookies. Cookies helped create a user profile for each visitor across sites to present personalization to said user on their first visit. As part of this, we took all product properties and defined a user affinity profile.
An affinity profile is a user profile showing the top interests and preferences of the user – e.g. see the image below.
- BRAND: Nike, Adidas, Champion
- Category: Shoes, Shirts
- Sizes: MEDIUM
- GENDER: WOMAN
With browsers blocking 3rd party cookies, user bias profiles are created for the individual site instead of across sites. To achieve decent personalization results, you need more data to produce these profiles and make good predictions.
With the Product Intelligence Engine, we can now build this user affinity profile without knowing who you are by relying on a single product you have seen. It means that while adhering to customers’ requirements for privacy, we can still serve visitors with a top-notch personalization experience by using our understanding of product correlations.
How the new user affinity is used for category & brand pages
If you look at a product, say Nike shoes, we can use our Product Intelligence Engine to see that it goes well with Adidas shirts & Under Armour shorts. By looking at that single product, we know that you might have preferences for.
If the said user then goes to a category page for shoes, we will, by just knowing the single product a user has looked at, be able to prioritize the following brands: 1. Nike, 2. Adidas, 3. Under Armour
- BRAND: Nike
- Category: Shoes
- BRAND: Nike, Adidas, Under Armour
- Category: Shoes, Shirts, Shorts
The described interaction improves the personalization experience of, in this case, our Pages product when very little knowledge is available for the visitor. The users’ affinity personalization does, of course, work well with our boosts.
NOTE: This example is, of course, simplified. The user affinity is a lot more complex, and the priority of the different attributes and values is not equally important.
Show me the stats
Before we launched our new algorithm for bought-together, based upon our global perspective in the vector space, we A/B split-tested it on an upsell recommendations to ensure the results were on par or better than the older algorithm. The results were stunning – our customers typically saw:
Conversion Rate
It’s a clear signal that looking at the holistic knowledge obtained from thousands of shops, and using that intelligence in the specific webshop instance, drastically improves the performance.
Having achieved such results, we rolled out the algorithm improvement to all customers.
AI + Human
While AI can make sense of big data, it will not be perfect in all cases. Those who tell you otherwise will be lying. We want to ensure that our customers continue to have control over their results. For recommendations, we have added Alternatives & Bought-together as two foundational algorithms that can be tweaked to ones goals. One use case could be, e.g., prioritising bought-together products from your white-label brands before other brands. In short, while our AI continuously improves, you still have control if you want to tweak the results.
This is just the beginning
Understanding the collective product catalog across all our +2000 clients means we can derive insights from verticals, which can be used collectively. We can see which groups of products and brands might spike in interest. More insights will follow as well as more blog posts describing how this technology is already used in our solutions such as Search, Recommendations & Pages, as well as the impact it has had.
Join us
We believe the future of eCommerce lies in understanding product correlations and respecting users’ privacy to deliver intelligent, personalized, and compliant solutions. We hope you will join us on this exciting journey!
