Marketing Podcast for Shopify Merchants: The Ecommerce Coffee Break

How To Effectively Deliver Personalized Recommendations That Convert Sales | #185 Alexandre Robicquet

May 15, 2023 Claus Lauter: Ecommerce Podcast Host | Shopify Partner | Marketing Optimizer Season 4 Episode 48
Marketing Podcast for Shopify Merchants: The Ecommerce Coffee Break
How To Effectively Deliver Personalized Recommendations That Convert Sales | #185 Alexandre Robicquet
Show Notes Transcript

This episode of the Ecommerce Coffee Break Podcast features a conversation with Alexandre Robicquet, CEO and co-founder of Crossing Minds. We discuss how to effectively deliver personalized recommendations that convert sales.

On the Show Today You’ll Learn:

  • How to increase revenue with product recommendations, upsells, and smart bundles
  • The biggest challenges for business owners implementing AI
  • How to boost email CTR by recommending products
  • How many SKUs merchants need to have to benefit from personalized recommendations  
  • And more

Links & Resources

Shopify App Store:

Get access to more free resources by visiting the podcast episode page at

Episode Sponsor

For inquiries about becoming a guest or sponsoring the podcast, email

Subscribe & Listen Everywhere:

Listen On: ​ | Apple Podcasts/iTunes | Spotify | Amazon Music/Audible | Stitcher | Deezer | Google Podcast

  • Subscribe today so you don't miss an episode!
  • By rating and reviewing the show in the app that you are listening to, you will enable us to invite bigger and more impactful guests.
  • Tag the podcast on Instagram

💰 Supercharge Your Ecommerce Business with Remote Marketing Talent. Discover the secret to scaling your online marketing without the expense of hiring, at a fraction of the cost — enabling you to grow your business faster. Visit for more info.

Support the show

Get free marketing ideas to help grow your revenue on Shopify. In your inbox for free. Every Thursday. Consumed in 3 minutes or less. Join 4,000+ Ecommerce Merchants, Founders, and Marketers here:

Claus Lauter: Hello and welcome to another episode of the E-Commerce Coffee Break podcast. Today we want to talk about how you can deliver personalized recommendations that convert actually into sales. Now, that's the main target for most merchants, but personalized recommendations with all the AI and, , some are still doing it manually is not the easiest topic.

So therefore, I have Alexander with me on the show. He's the c e o of crossing He is a distinguished AI expert and entrepreneur with a background in AI and machine. Learning from Stanford University, Alexander holds three master degrees in mathematics, machine learning, artificial intelligence from E N s, Paris Kelley, and worked with Sebastian Trun at Stanford University to develop artificial intelligence research that provided the foundation for crossing minds.

So he's a true expert when it comes to ai and we wanna find out how that plays into e-commerce, into recommendations. So let's welcome Alexander to the show. Hi Alexander. How are you today? 

Alexandre Robicquet: I'm great. Thank you. Claus for having me. 

Claus Lauter: Just curious what got you into artificial intelligence very early into the game?

Alexandre Robicquet: , a long series of misunderstanding. I actually couldn't tell less for informatic when I was younger. I just love, physics and chemistry and math, and I remember. For a very long time, hesitating between those three topic I was trying to find the one that would be the most versatile.

And when I discuss with a few people, I always come back to math because someone says, well, math is gonna be for economics, it's gonna be for physics, it's gonna be for anything that you want in life. Even maybe researcher one day One step after another. Math doesn't happen on paper anymore. Math happened with a bunch of, algorithms and, I remember I met my co-founder was my professor in Paris Cycling, and my second cofounder was my professor at Stanford.

When I met them, for math internship became certainly a machine learning internship, who became an ai, , vocation. Just meeting the right people at the right time. With an open mind. 

Claus Lauter: Okay, life definitely took you in the right direction there. As we could see from last November when chat g p t Open AI came out is exploding in the marketing space in the e-commerce space everywhere.

Now we wanna talk a little bit more about how AI can help business owners in achieving recommendation in achieving more revenue in sales. So this is a very specific market within ai. , what are the biggest challenges that you see right now for business owners implementing 

Alexandre Robicquet: AI in the possible best way for their business?

There's a lot of hype and a lot of noise about like what you can do and what you should do, and there's a lot of people that are building AI on top of other tools, like the number I think, of services that were birthed by G P T. G P T four is a fantastic breakthrough. It's an incredible large language models.

It's been worked on for several years by open ai, , to get there. And there's an API and I'm sure a lot of services, and there's a ton of services that are built on top of that. that means like, it's just like a better wrapper around the same type of ingredients, but. It's very important when commerce, especially on the Shopify ecosystem, are looking at their investment sheet.

And what am I supposed , to focus on this year to make my business successful is to clearly, clearly, clearly identify what are the KPIs that they want to achieve per team. , not necessarily for the whole company, so, Marketing and sales might have a different focus than product that might have different focus than operation, have a different focus than even engineers at the end of the day.

So it's really a matter of like trying to identify what are those KPIs and then how the consumers can be impacted or should be impacted towards those KPIs. So in the AI set of things, g P T is a great tool for marketing, to some extent. If it's leveraged properly, it's something that help you with the copywriting.

It's something that help you. But GBT is nothing else but a subset of what we call generative ai. And generative AI is not whole full ecosystem. That also is based on the concept of generating content, from a specific type of input. It could be text to image, it could be text to text, it could be image to text, it could be, keywords to text, et cetera.

You have a lot of things that fit together. The investment and balance the prioritization across all those tools. in our case, our companies focusing principally on personalization and recommendation.

The reason why we decided to go that way in a specific other way is that. This kind of personalization of this like one-on-one relationship with the customers is something that if you are, or Shopify, or if you have any type of merchant anyway, that need to have a customer relationship, you need to make sure that you are understanding very, very early what the customer want with a very limited amount of feedback and.

If you become the best salesperson in the world and you have your own physical store, the name of the game is really to understand when someone enters your ecosystem, what are you here for and what can you get from you? What can I give you? It's not who you are, where do you live? It's that those are a question that may be for relevant before or after, but not now.

Now the question is like, what do you want and what do I have that you will need or that you will? Purchase. , and that is something that's super interesting because it's not just AI in terms of machine science, it's ai in terms of machine engineering. It's being able to play those implicit feedback that are really hard to grasp, right?

The click, the glance, the everything that you would do on the human to human level. And in less than 200 milliseconds, digest all this information, make sense out of it, and then send back specific listing of products for someone that would help you optimize the KPI p that you wanted at the specific time.

So on the home page, you could be, the click through rate on the p d p could be, , the add to cart. And on the add to cart, it could be the upside on the cross because you wanna show more product that you would add to the thing. So there's many ways to look at it. I've done like a quick overview of ai, , right now.

I'm just gonna pause here in case I'm going in the right direction. 

Claus Lauter: No, I love that. I think it was crossing minds to do things a little bit different than others. I think most versions are used to come with their profile, with their perfect audience profile. Trained by what they're doing for the last 5, 6, 7 years with Facebook ads where they know it's like, is this person doing this and this?

Now you basically coming. From the other side, you have a blank sheet of paper, somebody coming into her store and actually he's just browsing. He doesn't know what he wants. And that's how I understand it. And correct me if I'm wrong, that's where you come in. So you basically, you take all kind of indications that they do while coming to your store to find out what they are actually looking for or what might 

Alexandre Robicquet: be the right product.

Is that right? Yeah. So you're absolutely right we can play the game together. You are. Absolutely in love with music. And you have a store of music, Klaus that you cherish, and that's your little kingdom. Someone enter in your store and I appear as someone to help you.

And I tell you, Klauss, I can give you two things. I can tell you the social security number, where they leave their gender maybe their ethnicity, or I can tell you the three first album they're gonna look at in your store, or that they're gonna actually just touch you don't need to have an MBA from Harvard to know that you're gonna be the second option if you wanna read, to understand.

And that's what's happening with most of the business. And this is one of those things that the Shopify can also help you out loud because you can actually understand those interaction while respecting the privacy to some extent of the those consumer. The other thing that's interesting is that in the e-commerce ecosystem, So you have a lot of cookies, depreciation, you had a lot of things happening with Apple emails and all this stuff. Of course. And if you're in Europe, like me as a French and German, that the G P R is no joke there. , the amount of users that are anonymized or the amount of users that are completely.

New to a store and commerce stores is roughly between 70 to 80% because people are also logged in. Usually when they check out. It's really rare for you to get someone. So that means like any other system of recommendation that is not taking care of those real time type of feedback and are supposed to give you recommendation based on the history of someone are living out.

80% of your user base that would receive just the most standard recommendation that would are the most popular without looking at who they are and what they're doing. And this is the thing that would really e-commerce merchant, especially on Shopify need to start cherishing the most is this raw first party data that everybody has.

If they have Google Analytics or I think there's an integration now Shopify, Google Analytics. Like this first raw part about what are the user doing and not who they are, what are they doing is fundamental for them to really expand and capitalizing on the moments on the session and on the follow up with the marketing and the sales because it's much more interesting to do cohorts on the behaviors and cohorts on the locations or other type of providing characteristics that are competitive suite nowadays.

Claus Lauter: Yeah, I think that takes a lot of, what we have learned in, have passed away so that people need to be locked in their Gmail account, need to be locked in and all this falls away. So basically every visitor coming to the website gets, , hits his own recommendation. Now, I can imagine that takes a lot of, calculation power on your side to figure out is like, what is the person doing and then coming up with the right solution.

From a technical background, and you don't need to gain away your secrets, but how does it work? 

Alexandre Robicquet: We have to have our own environmental server, , where we have bunch of GPUs in there to train models for each of the merchant. And what's super important is that each merchant needs to have their own fine tune models.

That's why I'm not a huge fan of , the solution that are completely relevant, by the way, so much better than nothing but the one that just. Do a one of, here's your model, here's your model, here's your model, here's your model, here's your recommendation, and then cheers. , that's great because that's still gonna increase your cell.

And encourage everybody that is listening to try a recommendation system, , just to see how that thing could improve their cells. And they might see a leaf of five, 10% in cells, which is quite consequence. , But if you wanna really dive into personalizing for your store with your user behaviors and patterns and who they are, you need to be able to model per store.

You can actually sometimes believe in more models because some models can be fine tuned for very specific KPIs recommendation system that's here to anchor a clicks, sir, certainly is not the same that the one that would increase the purchase or the long-term retention, like a click bait or something that's here to convert on the long term or 

completely. it's different, , but thanks for Shopify and all the integration that you can have with the APIs in maybe a few milliseconds, 10 milliseconds or something like that. You send the interaction to the api. The api, is on the Google Cloud, , system, and it's a model that has been trained on all those servers, and it took like a few hours, a lot of energy.

Those models are trained and retrained every day. , the full models, I'm talking about how the items are represented, how the user exist, the correlation between those twos. That's uploaded to a cloud Shopify and the clouds are communicated to one another, and every time an action happen, you need to retrain or reevaluate the user and set of taste or embeddings and then send back the recommendations, all that in 200 milliseconds.

So you won't retrain the whole model, but you just would retrain what's the dna, the taste, DNA of the user in real time. To make sure that if you're capable to grasp the context because someone that goes on your stores on a Thursday afternoon and the one that goes on a Sunday, when is relaxed and stuff, or two different users, even though they have the same email address, they might be a two different context.

Like it would be a different one and more stress. We can almost up and more, more relaxed on Sunday. And maybe it's important for us to adjust the models on the same person just because the mindset is really to bit switched. , But all that needs to happen extremely fast if you wanna be relevant, because it has to be 220 milliseconds for that to be relevant.

Claus Lauter: Okay. I think 200 milliseconds should work. I usually say everything that's above three seconds is bad. The four loading is so 

Alexandre Robicquet: 200 millisecond. That's just the API they need to think about. Like, okay, how does Shopify is gonna display that recombination? How does the character gonna load? I do have a very theme that's gonna make everything complicated.

What I know is like on my side, I can't, go, , too slow either, because so many things others need to happen at the same time. Yeah. 

Claus Lauter: Talk about the other things. I know you integrate not only into Shopify, it's , the whole technology stack, , that you work with. What other kinds of tools or solutions are, , you're working with Swiss Crossing lines?

Alexandre Robicquet: you have an intelligence that is capable to know what user wants, you want to capitalize on that on as many touchpoint with the user as you can. That means it could be. A survey that you send them for brand users and that they fill up, and how do you leverage the survey to personalize right away, even before them interacting with your brand?

It could be integration with the emails, , for all the campaigns for the follow up, but with the right timing for the flows that when they abandon the card for, I have a new product. Who am I seeing that new product too. It's like the reverse of recommendations, like recommendation of user four product.

You have integration with s m s, , because you might want to ping them about, Hey, I created a new page personalized for you, or here's a selection of, item for the summer for you, or anything like that. You could have to some extent, but that's for bigger customers, even integration with their physical store.

Think of Sephora, , as one of the potential massive store. They have a ton of product. This is something that says a bit complicated for them to evaluate. , and you have. Every summer, , a bunch of teenagers, 17, 18, 19, 20, or maybe more that comes to an internship or a salesperson, this is impossible for them to know exactly all the products say for has, there's like 3000 products.

, and you have someone that comes and say, oh, I'm a little bit oily today. What cream should I put on my skin? And. How do you expect that person, with all the goodwill and intentionally in the world that started two months ago, that would find Oh yeah, definitely. I'm gonna ping you this one.

No. So then certainly realize like, wait recommendation is not just for the end user. It's also could be for the customer's report to make sure that they can recommend you even if they're not expert. What would be the best thing for you to recommend to that person? It goes to everywhere that you think I need to have a personal touch.

Then this is where that thing should be plugged. And this is what you should expect from any recommendation partner that you have. Because it's the best way to create like a common intelligence because if they're dismantled, you could have recommended six times the same item at C different touch one because you never remember that you already presented that.

And the user would be like, I don't want that shoe. No, I don't want that peruse. Thank you. , so you need to have this kind of like memory or so cross all those search 

Claus Lauter: points. I like the example that you can basically upscale your own employees with that it might even be your support staff that people call in.

Exactly. That's a great example there. Now, in regards to implementation into a Shopify store, what is needed? What are the steps to get it up and running? 

Alexandre Robicquet: For any recommendation system, that you just installed the app. After that, if it's us, you contact us and you install the app and we do a kickoff call and we.

Especially on our case, like if anyone wants to try something, we have a three months, free P ooc. And the funny fact is like 95% of the people that did the POC with us converted, we make sure that we do a good job at deploying those recommendations on emails and also, but with the ease of integration and the flexibility of system is just a matter of installing the app and having the right customer support.

Claus Lauter: What kind of results do you see for new merchants that are using your system? 

Alexandre Robicquet: The lowest result we ever had on the increase in sales, was around 50%. , and after that it's really a matter of like, how do you combine the different algorithms from the homepage, the email list, we work with companies like Bright , and Chanel , on much bigger scale, and we reach results, like 270% increase in some of the clicks or some of the emails but that's a different story. For Shopify, it's also very good, but the amount of data that you have is a little bit less because you have less user coming there, and so algorithm takes a little bit longer to warm up, but it's at least 50%.

Claus Lauter: Okay, I chose you the power of recommendation when it comes to data. We were talking about data. What's the learning time before it really kicks in, or what's the minimum traffic that you need to have as a merchant before it really makes 

Alexandre Robicquet: sense? For a merchant, it's also a matter of Tiering or like budget investment and what you get out of it, right?

There's like always an ROI that you need to keep in mind. How much am I putting in those recommendation and how much am I getting out of it? This is why, , it's always better to work with companies that can do either price on volume base or price on attribution revenue. , but something , that's fair, right?

If I win and if I don't win, you're not winning anything. to go back on this, why I'm saying that is like, because some models might take much longer to get there, and it's unfair to pay when to all the time to get there, versus some models can work right away. In the case of Shopify, you can without any data, just by looking at all product catalog.

And extracting the, , images, , the text with embeddings like g p T could do for the text. But like there's many other models that could do that, like the tags, the materials, the pricing and all the stuff. You can already start doing what we call unsupervised content-based models that do recommendation in less than five minutes just based on the content of those things.

Now it's like how crafty that thing can and you'll be shocked. , to see how goodo recommendation can be because if you go on G pt, for instance, charge G P T and you say, I love those movies, can you gimme more movies like that, he would give you actually pretty decent recommendation, but he doesn't know anything about you except the fact that he said, I'm gonna find out movies that are similar to the three that you gave me.

And that's pure content base recommendation. So People tend to look down on content based, but if it's done really well, it can actually be. a great first day. So content-based is a five minutes. Then you think of a collaborative filtering, which is mixing, , all the patterns of the other users like, and it's the collaborative filtering is more of like other people like you did that.

That's what you see on Amazon, like other people per so also purchased. That's very much based on trying to map two different behaviors. And if something is missing from one person, they just see what the other person has done. And then after that you get into more crafty model, which are more hybrids and that mix of collaborative and the content base and, all those stuff.

And those ones can take a few weeks. But You can start recommendation on a good level with nothing. And if you wanna start improving and improving and improving, then you have to make sure that you are in a pricing model that worth to improve or in a pricing model that's fair for you to improve.

Because if it's a flat fee and you just let person will tell you, yeah, you have to pay $2,000 a month, which is absurd. Without seeing any results, that's, no, you shouldn't do that. 

Claus Lauter: Okay. Makes total sense in regards of store size, how many skews does a, , merchant need to have before it really makes 

Alexandre Robicquet: sense?

If you have only five products, you probably won't 

Claus Lauter: need it, but Exactly. 

Alexandre Robicquet: It's all a matter of like, how many product can I put in the page for my user to not be bored and scroll to the bottom? usually 50, 40 poor products, you don't really need reation.

If you have uh, customer journey and path and there's something that are logged with the emails and all this but it doesn't have to be super 

Claus Lauter: complex there. Okay. You mentioned a couple of your brands that you're working with. Are there any specific industries or verticals that are using it already, in their business or using it more than others?

Alexandre Robicquet: Recommendation system. Or VP of engineering was the former head of recommendation in Spotify. , so entertainment since the down of time, hammered down recommendation, , and to the point that they built always the home team now, , Entertainment was always two or three steps ahead , on personalization because that's the only thing that matters.

Retaining people, , showing them great content in my massive library. Then after that, you go to the, , e-commerce ecosystem because recommendation is, , Very interesting on the sense that compared to entertainment with like retention and click, it's hard to measure, , e-commerce. Super simple.

It's like , sales and clicks. And I can measure that today if they once had deployed the thing. So people were actually much more likely to adopt that kind of technology because the result were so concrete. And with that in mind, it was mostly the mid-market that was very eager to try because they were young enough to be hungry and have to take S to do that, and still flexible to just try things.

, But not too young to not have the mean to, or for it to be the end of the world if they fail. , so mid market was always very, very excitable and exciting. , when it comes to thinking of using this technology in this application. Now it becomes a little bit more and more democratized. So small business and mid market can also do that.

Of course. , enterprise is always, , A bit more painful, as I'm sure when everybody knows. 

Claus Lauter: Alexandra, as we're coming to the end of our coffee break today, where can people find out more about Cross Sigma? 

Alexandre Robicquet: LinkedIn, Twitter, , the website, crossing , please feel free, shoot us an email. We happy to chat. , happy to brainstorm. We happy to help. We have a great partner network.

Also, if you guys want to. Discuss about that ecosystem in general. We have also like mailing list with a lot of eBooks and cool little, , infos and numbers and stuff like that's about recommendation and all the stuff that I always encourage people to look into.

So the more people we can talk to, the happier we are. I will 

Claus Lauter: put the links in the show notes as always, and just one click away. Alexandra, I think you gave a very good , intro into what artificial intelligence actually is and can do.

And I hope that a lot of people will try out your system. , the numbers speak for themself you only can win, , with recommendations that are really made for this specific customer. Thanks so much for your time. 

Alexandre Robicquet: Thank you. Close. Have a great day.