AI is on everyone's lips and has been a recurring topic in recent editions of this newsletter. Last week, I was invited to a webinar to discuss the question: "What is a trend, and what is just hype?" Naturally, 45 minutes was far too short to cover it. Among other things, we discussed whether we will soon be doing all our online shopping via chatbot. I decided to put it to the test—you can read the results here. For over 15 years, Comsysto Reply has been implementing complex e-commerce solutions for clients across all industries. In parallel, we track and analyze relevant trends in our E-Commerce Lab to ensure we are ready for the challenges of tomorrow, today. We share our insights here—every Tuesday!

The great fear of being left behind

Both Dana Nedamaldeen and I see a lot of FOMO—the "fear of missing out"—among many companies in the market. The AI train is moving, and no one wants to be left behind. Often, the goal isn't to solve a tangible problem, but simply to get "something AI-related" into the tech stack. From a customer's perspective, this seems quite absurd, though it can still be a useful way to build internal expertise. One thing is certain: those who ignore the topic entirely will face problems sooner or later. Either way, there is a significant risk of betting on the wrong horse.

Specialized or broad AI models?

A key question is, for example: Yes, AI will definitely find a place in most e-commerce landscapes, but what kind of AI should it be? Despite the current hype, the topic isn't new; AI models for product recommendations, for instance, have been around for a long time, albeit with mixed success compared to very simple heuristics. What is relatively new is GenAI, which has become widely known primarily through ChatGPT. It is now possible to talk to a computer as if it were a customer advisor and receive quite solid answers.

This is where we hit a sticking point during the webinar. The two most extreme viewpoints would be these:

• I train a completely custom model for my online shop that is specifically tailored to my products, my customers, my current marketing campaigns, and my brand's tone of voice.
• I use a completely generic LLM that is no more optimized for e-commerce or my industry than it is for politics or the weather forecast.

Providers like Neocom.ai base their products on the premise that neither of these two options leads to an effective result. A completely in-house development is far too complex and risks becoming technically obsolete in a short time. And the results of a completely generic model are simply not good enough. The solution is a middle ground: individual data from the respective retailer is used to refine a model that is already designed for the e-commerce use case and also features various interfaces to other systems (keyword: RAG - Explanation from Fraunhofer).

Ultimately, this is a bet that very generic models will not be good enough in the long run. The question of quality has two dimensions. First, how good are the content-based answers and recommendations? This is largely about data quality. Second, how natural, precise, and pleasant is the communication style? Anyone who cannot keep up with the best in the market on the second point—for example, by constantly misinterpreting questions—will fail to leave a positive impression on users and, consequently, will not build trust in the content, no matter how factually accurate it may be.

How good are broad AI models currently in e-commerce?

Back to the question of whether generic models might be good enough after all. I decided to try it out and asked ChatGPT, in a completely new and context-free session, to advise me on my new headphones. The initial follow-up questions seemed plausible and linguistically flawless.

After this introduction, the chat continued with concrete product suggestions, their pros and cons, and descriptions that were fundamentally correct but somehow meaningless. For example, the sound of one model is "detailed and balanced," another is "neutral and clear," and a third is "clear and balanced." I don't know exactly how that would sound in my ear, but I would probably have to test them in a physical store anyway. Ultimately, our conversation led me to the Apple AirPods, and I wanted to know where the best place to buy them would be.

It lists a few retailers I would have searched manually anyway. On top of that, I get this warning: “Be careful with offers that are significantly below the market price, especially on platforms like eBay or classifieds sites. There are reports of well-made counterfeits of the AirPods Pro 2 that are visually almost indistinguishable from the original but perform significantly worse in terms of sound quality and functionality.” This might actually be relevant for some people.

I decide to buy from MediaMarkt, and lo and behold: the price is wrong! Saving €40 is good news at first, but I wouldn't even have noticed if I had chosen Apple. I would have been just as surprised there to suddenly have to pay €279. So, ChatGPT isn't really up to speed on price comparisons yet.

I confront the AI with its mistake. It immediately asks for feedback on what kind of response I would have liked in this case. This is how ChatGPT learns from me, and communication improves through such user feedback. Unfortunately, this doesn't automatically apply to the content: the prices mentioned are different now, but still wrong. Well, I'm not entirely convinced.

Apparently, OpenAI hasn't implemented live function calling for price comparisons. Or perhaps it simply lacks access to the right data? For Google—think Google Shopping—this definitely shouldn't be a problem, but even here, Gemini doesn't (yet?) provide me with live data. So there might be something to the hypothesis that generic models simply aren't good enough as shopping interfaces yet. At least for now.

What will online shops generally look like in a few years?

Nevertheless, the question arises whether online shops in their current form will even have a reason to exist in a few years. What if ChatGPT, Siri, and their peers soon become so good at guiding us through the entire exploration, search, and ordering process without any media disruption that we no longer have any reason to use classic storefronts? I asked the LinkedIn community last week whether that will happen—you can still participate.

There are certainly signs that OpenAI can very well imagine such a future without its own storefront. Or at least an additional important channel, similar to Google Shopping. Hanns Kronenberg reported just last week that OpenAI is apparently working on a closer integration with Shopifyto be able to initiate orders directly from the chat—that would be a real game changer!

And what should we tackle today?

Currently, this is still a thing of the future, but it shouldn't be ignored today. At least from a risk management perspective at a strategic level! If it really happens, it would have dramatic implications:

• SEO will practically no longer play a role; only discoverability by the major chatbots will matter.
• There will be hardly any way to sharpen your own brand on your own platform because users won't be there. This makes social commerce all the more important.
• If you can no longer differentiate yourself through the customer experience in the online world, the offline world will count for all the more. Excellent logistics will therefore become even more important.

No matter what the future brings: thinking about and addressing these topics today is definitely not a mistake. And in parallel, no one should forget to implement the small and simple things that all users will soon expect everywhere. For example, good and valuable summaries of all customer reviews. Or a good search function that truly deserves the name. And, of course, prices should also be accurate and competitive. The current priorities of online shoppers in Germany give an idea of what is already a must-have today and where there is still more room for experimentation.

That’s it for today – and here’s what’s next

One final recommendation for today: No one should underestimate the power of analytics data from their online shop’s search bar. During the heyday of Google, users learned how to search in a way that was as “machine-compatible” as possible. Today, they are learning that they can type more or less whatever they are thinking into a search field and get a helpful answer back. User inputs are therefore much closer to the user’s actual intent than they were a few years ago. A real goldmine for understanding them better!

Next week, we will take a closer look at the topic of logistics. Whether you have your own e-commerce frontend or not – the goods have to get to the customer…

See you next week!

Christian & the Comsysto Reply e-Commerce Lab

‍