Why hasn't AI revolutionised how we shop yet?

We all know by now that AI is a brilliant tool for personalisation. When there are a lot of potential options to choose from (e.g. movies, music or media), AI is great at surfacing the most relevant results for each individual’s unique taste and preferences. However, despite the advances in AI, when it comes to certain online purchases, most people are still shopping much like they did a decade ago; manually searching, filtering and scrolling through thousands of products to find what's actually relevant to them. We refer to this as the ‘product discovery problem’.

Assisted by Machine Learning, the music discovery problem was largely solved years ago by Spotify et al. TikTok made it easy to find relevant entertainment and Netflix have by and large solved the movie/show discovery problem.

But while many have attempted to apply similar thinking to e-commerce, none have truly succeeded. So the question is why hasn’t it been solved? Why isn’t there a hyper-personalized algorithm that can sift out all the irrelevant products across the web and uncover the items that are of genuine interest to us?

My cofounder (a Computational Neuroscience PhD) and I have invested years of research into answering that very question. We've considered it from a data, technology, behavioural and business perspective. All play a key part in solving it, but here we will be focusing on data.

E-commerce is a different beast

Examining the problem from a data perspective, we believe ‘e-commerce’ has 4 characteristics that together make creating a personalized cross-retail experience far more difficult than solutions in other markets:

1. Ginormous potential pool of options: There are 1bn+ unique SKUs available to buy at any one time in the US alone, (compare this to movies for example, where only about 500k movies exist in the history of the world).

2. The pool of options is constantly changing: Not only are new products listed every day, but products only remain relevant for a short period, meaning data goes out of data very quickly. For example fashion often only lasts 1 season whereas music or film can remain relevant for decades. Furthermore, the key traits of products constantly change e.g. prices and stock change day to day.

3. Extreme fragmentation: Products are listed across hundreds of thousands of different retailers’ websites. Compare this to 3 major record labels for music.

4. Lack of uniformity: We’ve read the HTML of thousands of retailers, and while there are trends in how retailers capture SKU details, there is very little consistency. Almost every site independently decides how they will list a SKU and the different conventions they will use.

If we review some of the major names typically associated with delivering successful 'discovery experiences' against the points above, we quickly see why ecommerce is much tougher nut to crack (this is not to say these were not also hard solutions to build with their own unique challenges).

Personalisation in TV/Film: Netflix is not afflicted too much by any of the above problems. Roughly 500,000 movies exist in the world and about 817,000 TV shows to choose from. There are only a handful of major film studios.

Personalisation in Music: Spotify is largely only afflicted by the first problem; millions of options. There are now 100m tracks on Spotify, but Spotify was possible because these were aggregated between only three major labels. The so-called 'big three' major record labels being Universal Music Group, Sony Music Entertainment, and Warner Music Group.

Personalisation in Social: on TikTok, IG, YouTube and other social media the content is all created on the platform. This means there are no fragmentation or uniformity issues. You have 100m+ bits of content to analyse, but it is all in a uniform format which the platforms determine themselves. So they do not suffer from issues 3 and 4.

Who is positioned to solve the e-commerce discovery problem?

Amazon and other large retail platforms?

My cofounder and I previously worked on AI assistants and data science projects at large retailers and retail platforms. So having seen it from that side, we know that - as it was with TV/Film (Netflix) and music (Spotify) - personalisation in e-commerce cannot be solved by an incumbent. In film the incumbents were the film studios, in music it was the record labels. The e-commerce equivalents are the retail platforms, brands and marketplace - of which Amazon is currently the largest.

As a retailer you are intrinsically biased. You exist to promote and sell the products on your platform - even if better options exist elsewhere.

But even if we ignore the conflicting commercial objectives, retailers typically only see a small part of the path to purchase. While they have lots of interaction data in pure volume terms, they only see a fragment of each consumer’s shopping data. A small piece of the shopper’s path-to-purchase puzzle. This was the big problem we faced when working on AI solutions for retailers, we had no idea what customers were doing on other websites. They could have been shopping for weeks for a specific type of jacket, but when they arrived on our platform we would have no more information about what they were looking for than the first site they visited. And when they left, we had no idea what they went on to do after. 

We mentioned earlier that Moonsift users have saved items from over 30,000 sites. Very few sites have more than 1% share of this. The largest is Amazon with just over 8% of total saved items.

Amazon dominates in low cost, commodity and functional products, but this cannot be said for higher cost, more taste driven categories that typically require a lot more research and ‘shopping around’. Amazon have almost no market share in the $100bn+ luxury fashion market for example - which is spread over thousands of more boutique retailers and platforms. In Europe Amazon are not even the leading retailer within the wider ‘apparel’ category. Given Amazon exists at the commodity end of the spectrum where ‘taste’ and ‘style’ have little impact on choices, search is a reasonable tool and personalisation is actually less of an issue. 

We have no doubt that individual retail platforms will see some improvement in the quality of their search results by leveraging advances in multimodal AI. But they are certainly not the ones to solve the cross-retail discovery problem. This needs to be solved by an entity that is agnostic and therefore not tied to selling its ‘own’ stock. 

Google shopping and other large aggregated product datasets?

Google shopping has billions of products and while they have increasing amounts of ads, they are otherwise unbiased. However, the problem when you ingest large amounts of raw product data (scraped or uploaded by a brand), there is no context on who has interacted with the product and therefore no data on who it is for.

When you only have raw data like this it is very hard to figure out which products are relevant for which shoppers. So if a shopper searches for a ‘red dress’ you will likely have at least 50,000 options - but how do you choose which 10, 20 or 100 to show them?

This is a “cold start problem” and one we experienced in early experiments we ran at Moonsift. The early experiences for shoppers on these platforms are very poor and ‘unpersonalised’. It’s statistically very unlikely you will serve the shopper any items that are ‘relevant’ to them and their unique taste/preferences. Which means they are much better off going to a more ‘curated’ store or platform that is targeted towards their ‘market segment’ e.g. Anthropologie, Urban Outfitters, Farfetch, John Lewis etc.

The result is that search platforms with large product datasets typically only become useful to shoppers if they can articulate very specifically what they want using keywords the engine understands. This point is further confirmed by what we see when analysing data from the thousands of shoppers that use Moonsift: only 0.22% of items saved to Moonsift are from Google shopping.

The same cold start problem has afflicted many startups that attempted to solve the product discovery problem in the 2010-2017 period when Machine Learning was having it's first major impact on e-commerce (commonly referred to as the ‘Tinder for shopping apps’).

Platforms that simply aggregate large datasets of raw product data (whether horizontally or within a specific category) and hope to create a 'personalised' shopping experience on that shopping site or app have thus far proven unable to compete with curation in the context of shopping discovery. There is no reason that advances in AI alone will change this.

Cross platform transaction and discount code tools?

There are several browser plugins and apps that serve shoppers across the different sites they shop on. Could these be better positioned to solve the discovery problem?

Companies such as Honey, Klarna and Fetch have been built to serve the final transaction through discount codes, delayed payment and rewards. This makes sense, because this is the easiest place to earn a fee. But while these companies have useful data on what some shoppers end up buying, they are missing the whole path to purchase. They are missing the messy middle part of the journey that is product discovery: the thousands of products browsed, clicked and considered before coming to a decision. This means they are not learning and collecting the data needed to solve the discovery problem.

Fear not, the future may be bright

Personalisation and AI-enabled product discovery in e-commerce hit a ceiling many years ago due to a combination of the reasons highlighted above and a lack of advances in AI. But this is all about to change. The multi-modal image-text models that have been trained for ‘Generative AI’ purposes, such as CLIP by Open AI, will soon have a significant impact on how we find products online.

Moonsift has been experimenting with these models and recently wrote an article exploring the new possibilities with semantic retrieval.

But this is only the beginning. Both the models and the infrastructure that enables us to apply these models is coming on in leaps and bounds. Critically the speed is reaching a rate that will soon allow large vision language models to be applied to vast product datasets in a shopping context.

We hope to share this fascinating journey with you. Check out www.moonsift.com/ai to learn more.