What's needed for AI to solve the product discovery problem

Motivation

As we’ve discussed before, product discovery online is broken.  In fact it’s not broken, since it was never great in the first place.  Whilst there are great platforms to discover any music, films or videos that match your taste, in the world of taste-driven commerce there are few all-encompassing platforms and the largest (Google Shopping) is essentially an advertising and price comparison platform.

You’ll probably be surprised just how many different retailers online shoppers visit and how long the process of discovery can take.  We certainly were!  Moonsift shoppers used our universal wishlist to keep track of items across over 14 thousand retailers last month. Those that save 40 to 50 items a month, save from, on average, 15 different stores.  The time between saving an item and marking it as purchased averages around a month:  This isn’t a job for a few clicks on Amazon.

Why the copilot approach?

You may be imagining an AI that enables better product discovery as a single platform, perhaps a chat-bot, where you start feeding it with your desires and it keeps finding you things without the need to visit any more retailer websites. This may well be part of our eventual solution.  But we won’t start by attempting to launch a product like this straight away - a mistake we keep seeing in this industry!

When we started Moonsift, we knew we wanted to build the best AI-powered discovery experience. So we initially thought we could launch and AI-powered product aggregator where shoppers could start their discovery processes. With the help of feedback from shoppers and investors, we soon realised it wasn’t the best way to go for one key reason: Becoming the starting point for shopping discovery would require a large change in shopper behaviour. There were only two ways we were going to change the behaviour of online shoppers:

• A massive and unsustainable marketing budget to get shoppers to keep coming back to such a product.

• An AI tool that is from launch such a step change improvement over other product aggregators, it becomes viral.

We of course opted to take the AI route, but to create such a game-changing AI tool that is comprehensive across all retailers, two things are needed: Product data from all retailers - which is very expensive to acquire - and, more importantly, plenty of cross-retailer user shopping data to enable understanding of the discovery process and train machine learning models.

This is why the discovery problem lends itself to building a shopping copilot; a tool that increasingly improves the experience, augmented by reinforcement learning from real users.  It is a tool that users will not expect 100% accuracy.  Think Tesla’s driving copilot or Github’s coding copilot.  They enable us to tackle the very hard problems in AI in stages. A copilot also solves both of the above problems: It will not try to change shoppers’ behaviour of visiting many different retailers. It will also acquire the data required to build the full discovery experience.

Building blocks

State-of-the-art multimodal machine learning models have enabled a step change in generalising semantic understanding of product data across retailers.  At Moonsift we have already merged datasets from 10s of thousands of retailers and built tools to search them. Although this is a necessary first step, and albeit one that few companies are able to do on this scale, it’s not sufficient to enable an AI revolution in shopping.

The other piece in solving this puzzle is learning as much as we can about the discovery journey.  This journey happens haphazardly and in many different places online.  This is why we have built the best universal wishlist browser extension and app.

Hundreds of thousands of items are saved to the Moonsift browser extension and iOS app every month

Our diverse user-base love using Moonsift for saving items to one place to make considerate shopping decisions and sharing them with others.  In fact every month Moonsift shoppers save hundreds of thousands of items to wishlist-style collections.  Let me put that in perspective: this is the same order of magnitude as the number of sales Urban Outfitters, a multi-billion dollar US retailer, makes in a month!  And we are still a pretty small start-up.

This data also signifies strong user intent.  The key point here is that we at Moonsift are almost certainly going to know what someone is looking for before the retailer when a shopper lands on their site.  The retailer is limited to past sales data.  Past sales are helpful when you are selling someone fast-moving consumer goods like washing powder.  But it’s a lot less helpful when selling one-off taste-driven purchases.

Perhaps the retailer may have enough past data on a visitor to be able to “predict their style”, by feeding one of the many b2b retailer recommendation and search platforms that make up the billion-dollar personalisation industry.  But they will still have to guess not only the type of item someone is looking for on that visit but also a whole bunch of the visitor’s other requirements for it.  Even if a visitor has clicked through from an aggregator to view a specific product, they are still missing vital context from the rest of the discovery journey.

Therefore, on all stores shoppers visit they end up starting from scratch, having to dig out the appropriate category of the website and apply filters which are never quite the same on any two sites.  And this continues on on every site we visit until we are drowning in tabs and frequently either give up or end up buying something we don’t really want.  At least with our wishlist app you can close those tabs.  But still, discovery takes far too long.

Understanding the discovery journey

The discovery process has some clear stages at the start (inspiration) and the end (purchase), but in between these two is what Google likes to call the “messy middle”.  Messy because there are lots of paths a shopper can take to get to their chosen purchase, especially when personal taste is a strong component of the purchase.  We don’t like mess here at Moonsift, so instead we are pioneering what we call the concept of “Discovery Science”.

The discovery journey as Google sees it (Source: Google)

We are in the process of deconstructing the discovery journey into its various sub-stages.  

One example of a later stage in the journey we have been looking at is when someone has found an item but it’s not quite right.  They could search for similar items on the retailer or maybe even google shopping.  But there’s no way to tell these services why it’s not quite right.

An example could be it’s the wrong colour, or a more concrete example could be this bobble hat in the image below.  What if you love the hat but you want it a bit more brown and don’t want the pom-pom?  We have already created a tool that allows you to do just this on Moonsift.  Thus speeding up the discovery journey in a very useful way.

A vector-based search for a hat that is similar but is brown and does not have a pom-pom (Source: Moonsift Internal Tool)

You can find more examples of our discovery power-ups here.  In these examples we need to leverage our multimodal machine learning models to explore them.  These models allow us to work with product data in a way that is much more natural than before, since they have a much stronger semantic understanding.

Further, they will allow us to identify more points earlier in the discovery journey and suggest paths a shopper may want to explore with explainable results as to how they are relevant.

Finally, because Moonsift knows our users better than any retailer, we are also able to personalise our results based on an individual’s tastes.  This will require us to develop a way to combine multimodal models with recommender systems, but a discussion of this is for a future blog post.

Once again it’s important to remember that discovery is a process - it’s as much about giving options to learn about what someone is looking for as it is about the final choice.

Why Moonsift?

We are in a unique position to build the shopping copilot.

Firstly, we have data for the entire shopping journey, not just transactions, and we have it across retailers.  This allows us to train machine learning models for each part of the discovery process.

Secondly, we have a popular tool in which AI assistance can easily augment the existing discovery journey and, crucially, be incrementally improved over time.

Finally, we also have shoppers who are actively asking to use discovery AI and it’s these shoppers who will train our models to help improve the discovery journey.

As an analogy, consider the recent success of OpenAI’s ChatGPT.  A huge amount of data was required to train the foundation model, GPT.  But for the chat interface to work, this also  required years of training on conversations with actual users before it was good enough for a general launch.  Yet GPT was still able to provide some value, for example in the form of Github’s copilot during those years of training ChatGPT.

Likewise we believe we can provide value But we have the data we need and the people to train it so we can get there.

The “flip”

Giving the shopper their own discovery AI (Source: Moonsift)

We are entering an era where relatively cheap multimodal machine learning models enable a step change in processing product data and will enable product discovery to be solved.  But to do this you also need to have:

• Cross retail data

• A view of the whole discovery journey

• The ability to learn how to deliver your AI results in a useful way

At Moonsift we will be able to address this last point by developing our copilot as part of our current product, though this isn’t the case elsewhere.

Retailers who have tried to develop new discovery experiences have launched them in the past and they have fallen flat, as they rely on chatbots or recommender systems that start the user on a new shopping experience that hasn't been delivered.  A retailer’s incentive is rarely to offer shoppers the best discovery experience but instead to drive the shopper to a sale. This leads to miss the messy middle:

Commoditised recommendation engines are employed to show you very similar items (the near-end point of the discovery journey) as the model is trained on just converting those people who are almost ready to buy.  Elsewhere, AI is aimed at initiating discovery in the world of ads, bombarding you with ads for things that may be in your style if you are lucky, but rarely what you need and all too often are fast-fashion brands part of a race to the bottom in quality.

It is time that applied machine learning in online shopping delivered more than this and helps people with the entire shopping discovery journey.

In other words, the AI is currently on the side of the retailers, not the shopper.  After all retailers don’t want people to find everything in one place, they just want them to stay on their own website.  We will flip this situation on its head, and give the shopper an AI that will be on their side.  It’s an AI that shoppers want to invest time in training with their tastes, so it’s the best discovery AI that exists.