Ep 596: 3 Shopify Hot-Fixes for AI Shopping: Kurt Elster on UCP, ChatGPT Referrals, and TikTok Shop

0:00If I had to pick like a single universal pain that everybody's feeling, it's customer acquisition cost. The problem is like getting started just feels like you're burning money. And then Andromeda, the algorithm update that Meta did that's very like LLM driven, is a different animal than people were previously used to working with. It seems like there is no limit to the signals the thing can draw from your website and your customer and your ads. So it's like you got to spend money just to figure that out. And then once you get ads that work, hope you could keep it working or scale it. It's tough.

0:31There's two things that I'm doing for optimizing for AI search. One is, you know, what we've discovered is...

0:45Kurt Elster:This episode is brought to you by Contentful. Marketers, you know that feeling when your creative clicks, when that social post sends engagement through the roof? When your outside-of-the-box campaign hits ROI positive, when a personalized homepage turns prospects into customers, it's utter marketing bliss. Contentful helps you create tailored omni-channel experiences without working overtime. No stress, no limits, only possibilities. Get the feels at contentful.com. Kurt, welcome back to the D2C podcast. I think you have one of the best vantage points in the space because you're dealing with merchants every week.

1:24Kurt Elster:And you guys approach things from an e-commerce first perspective quite often, which I really appreciate. We're always – I know you touch a lot of growth. You talk a lot about growth. But I'm always interested more about what's happening on the Shopify side. So what do you think is something that merchants are missing right now or they wish they would have paid more attention to right now, six months into the future? Okay. Well, number one, for the last year, Shopify added categories. like you're in your shot your product listing and there's this new field that pops up like right under the photo says category and it like auto suggests it and it's weird and then it like when you pick one you go okay that's fine whatever it'll be like a pair of old t-shirts like okay and then it'll be like well did you want to add these category meta fields and they're very prescriptive and we kind of were at first didn't know what to do with it we ignored it and then after a while we went all right well maybe we'll just set the category and see what it does and like it didn't seem to do anything but it freed us up to misuse the product type section for like like categories and a navigation.

2:20And it turned out, you know, this year when Shopify said, hey, we've worked, you know, with Google and other partners, and we've got a thing called the Universal Commerce Protocol, UCP, a standard for these marketplaces, for ChatGPT, for Google, to query a marketplace and then get back products. And well, it turns out that's what the category meta fields were for. It was a search engine can now go use that protocol, go to Shopify and say, hey, what products fit this? Then Shopify returns them. Turns out it uses those category meta fields. Like that was all for AI to understand a catalog. Because if you think of like, you know, Shopify as a platform, millions and millions of products, how do you search through that very quickly?

3:05And how do you search through it reliably? Well, that's why they, you know, had this very prescriptive taxonomy. And so once we figure that out, now we're like backfilling and rushing to go, okay, we got to fill all those in, you know, the more the merrier, that optimizes us for a potential, you know, AI shopping future.

3:22Kurt Elster:What is the AI shopping present? I think I understand that ChatGPT backed off a little bit, like making a native checkout or anything like that. I think they're still focused on partnering with through this protocol, as you mentioned. But where are we with like, how many referrals are you seeing on your stores from AI now? I'm happy to say we see them, period. Like it exists. It is a phenomenon that is occurring somewhat. ChatGPT, any link that it has in the chat, they're very reliable in adding their UTM query tag to it where, you know, it makes sure you know that refer source equals ChatGPT.

3:58And so if you are getting referrals from ChatGPT, you'll be able to figure it out. Like even in our Shopify apps, we could see the referral and installs. And now we're seeing ChatGPT pop up as a merchant will ask that, hey, I need an app for X and then it'll make a recommendation. And if we're lucky, it's us. Oh, wow. In our stores, we're starting to see that too, but not to like a huge volume. And that was happening before any of this other AI stuff. It was just like,

4:26Kurt Elster:you know, ChatGPT doing web searches for it. Now the idea is like it would skip that web search step, go straight to this commerce protocol, pull that stuff in and make the recommendation. Okay, that's cool. And then they were going to take it a step further with you then made the purchase inside ChatGPT. Now we're learning like, okay, well, maybe they're not going to do that. That's okay. You know, when we look at in the past, like other people have attempted it, Facebook was supposed to be able to make purchase in app, you know, in Instagram. Anytime people have done this, it doesn't really work, right?

4:57It goes away.

4:58Kurt Elster:TikTok's the only one that's threatening, right? TikTok and Amazon. Well, Amazon, like they're very much built for that. Like that we accept. A marketplace we'll buy from, a brand, a manufacturer we'll buy direct from. But it turns out like other surfaces where we're very aware that it's the man in the middle, we don't do it. I don't know if it's like context switching or what. It's just a difficult thing to crack. So I was in no way surprised what you had to achieve. You said, hey, we're not doing that. But okay, where we're at now in theory is I can use shop campaigns, which like Shopify has its own advertising network built into Shopify called shop and shop campaigns.

5:37You can really easy to set up, doesn't have like the crazy granularity that, you know, like meta ads do. And then they've got just a checkbox. It's like, hey, did you want third party placements? Yes, I do. And those third party placements, one of them is chat GPT ads. And so the idea is like, I use ChatGPT, I'm on a free plan or like the lowest tier plan, similar to like Netflix, they've got ad supported plans. ChatGPT has an ad supported plan. And if I ask that, you know, the right query, hey, I want, you know, new socks, then maybe someone who sells socks and has shop campaigns enabled and has those category meta fields filled out suddenly appears as an ad in ChatGPT directly.

6:16Now, here's the problem. I have yet to be able to trigger this. I have tried. I have not been I have yet to see a chat GPT ad myself and I test it just like I open an incognito window visit chat GPT directly and then you know start asking it leading questions like I need new running shoes do you have any thoughts there you know and I just haven't I haven't gotten it to show me an ad I've seen screenshots of ads from it but yeah so I think I think that's a thing that's around the corner uh but just not quite here yet and then how much tinkering on the UCP is that uh

6:49Kurt Elster:universal commerce protocol, are you just sort of letting it set itself up? Are you going back in and having to readjust it a lot? No, what's crazy about this is, one, you are opted into it by default in Shopify. It's in, if you go in sales channels, in your settings, you could find this. And I think it's at the bottom. And you could go in there and like check it status, see if you're opted in. Like that's part of it. And all this gets rolled into like the Shopify product network. So it gets a little confusing in like naming and quite what you're looking for. because they also do a thing now as part of shop and shop by product network where like it'll place my products into other people's stores if they've opted into it where like on the thank you page it'll show other people's products and then if the person adds that to their order the store it appeared on gets a commission which they're like that's pretty cool that's brand new just rolled out similar to this like the shop campaign ads and these chat gpt ads but you know different surface different thing but it needs like you need that But I talked about the category meta fields and like those extra taxonomies.

7:51That same, you know, those same fields, that same protocol within Shopify itself is still what's surfacing those products. So benefit all around to filling those things out.

8:01Kurt Elster:I like your note here about Shopify product network being able to allow vendors to test into like new categories, maybe. So like if you don't have a category, but you're thinking of expanding into it, you could always show maybe a competitor for that category and just sort of test the demand of your audience for that category. Yeah, well, it's kind of cool. The idea is like, because I got to talk to the director of product at Shopify, who was responsible for product network. What she was saying was like, that's the idea is, you know, you sell in one category, person lands on a thank you page, Shopify through AI, which when you think about what they did is really impressive considering the volume of catalog and then like the wild variation in customer that they're dealing with.

8:41it'll then surface the products that it thinks the person is likely to buy you as the store owner get a commission on that so great like a little bit of extra profit why not you know i'll take it but more importantly to your point you can then you can it'll show you what's sold and so you could start to figure out like oh you know maybe those are items i should list in my store and then it like and you know exactly what's sold and so you could reach out to that merchant go hey you know do you do wholesale? And they probably do. Can I list this stuff in my store as well? Right. And then Shopify gives you, makes that very easy with Shopify collective, where it essentially is like a dropshipping tool for two Shopify stores.

9:21It'll just automatically, you know, I can list products from one store. If those get sold, the order gets sent back to the originating store, you know, and fills it out. It works like dropshipping, but, you know, entirely housed within Shopify and fairly automatic to set up.

9:34Kurt Elster:Yeah, world of possibilities there. Back to AI a little bit. Is there anything else vendors need to be thinking about for optimizing their PDP or for their, you know, their product descriptions? I think you have a note here about AI are pulling roughly the first 6 ,000 characters from product descriptions. Is that causing you to do any edits, like push the meat up front? I'm sure we're doing that anyway. Yeah, well, there's two things that I'm doing for optimizing for AI search. And one is what we've discovered is structured data. So JSON, J-S-O-N, it's a markup language that has a standard. You can add this into a web page.

10:10And essentially, it's like a very a programmatic way to describe the content of the page in like a very strict way. You know, these AI tools, they have to spend a lot of resource on figuring out what the heck the meatbags are talking about, right? You know, like the start and end of a sentence is obvious to you, and LLM has to think through that. And so when you can structure that data where the LLM no longer has to guess as to like where the sentence starts and ends and what that means, that really helps its understanding. and then when you think about like power consumption and the volume of data that has to get ingested by a search engine, you can see why they will take the easy path out every time, right?

10:52Like generating an image I saw, it's like running a microwave for 30 minutes or something crazy. Yeah, they've got like this crazy energy requirement and so by putting the data in a format it likes, well, that's like the new SEO. And so every Shopify theme, every Shopify store by default has structured data set up, but you can go create, like one, verify it. Google gives you a free tool to do it. And then two, once you verified it works, like you can go crazy with it. You can make listicles and then like make a listicle article that's like the top seven e-commerce podcasts. And immediately you're like, boom, number one, Eric, D2C.

11:27Boom, number two, Kurt, unofficial Shopify podcast. And it's just going to like, and then you wrap that in JSON and then Google's going to quote that back to the person when they ask what are the top podcasts. So like a lot of black hat SEO stuff or like gray hat SEO stuff that worked 15 years ago is back, you know, for these answer engines.

11:44Kurt Elster:En masse. Yes. Super cool. And then like the AI makes it very easy to do. Like you could literally download this podcast, give the transcript to Claude, then tell Claude, hey, what's he talking about? Okay, this is my site. How do we implement this? It'll just give you the plan. Like that's crazy. But you got to put in the effort. It is crazy. And you got to know where to start. So that's like traditional SEO. Now it's going to be called answer engine optimization AEO. All right, so there's that. And then for like optimizing for universal commerce protocol, this is a little funny. Shopify tells you like, well, these are the, this is exactly the fields in your, your product listing that we look at for this.

12:21And this is what we don't. And it turns out like product meta fields where we would normally like custom fields where we'd normally store, you know, extra bits of info. They don't look at it. And for the product description, they look at the first 6 ,000 characters. and you know i think they go so far as like oh look at like uh the image and the colors in the image to try and get everything to match i know product network does that but then like beyond that they're really looking at that category and then the associated meta fields and so that's where you know if you've gone through the effort and you know stuck with like even though it seemed prescriptive use those category meta fields and filled them out i would imagine very few people did.

13:01And so the people, especially if you're in a competitive space like apparel, the people who filled those out immediately have an advantage in getting their products surfaced more often.

13:10Kurt Elster:Do it right now. You mentioned Claude. Just yesterday, we redesigned the D2C website, figured out a plan to port and remake all the content, built my rowboat website the other day. This thousand half-finished projects have been, a million half-finished projects are launched every day now on Claude, but you find yourself in the middle of what, you know, the SaaSpocalypse. And what's your take on Vibe Coding, you know, as a developer? Like you released a tool this year, right? Yeah, well, we launched a Shopify app and we did, Claude helped code a lot of it, you know, plugged it into our IDE, you know, developer environment, and used it.

13:50It just like, Claude is so good at coding, generative AI, like when it comes to coding, that it is really good at it. And part of it's just like the nature of coding is very structured. It's logical. It follows a clear syntax. And more importantly, there were all these open source projects and all this documentation. And then these LLMs went and they ate every gross thing they could find on the internet. And that includes these projects and this code and the documentation. And so at this point, you know, I could load up Claude Code, drop a spreadsheet into it and be like, hey, we got to, you know, rearrange this into this format for Shopify.

14:25you know and you and like this app i didn't even have to point into the documentation anymore

14:30Kurt Elster:half the time the darn thing already knows and it knows what you don't know it knows the questions you don't even know to ask like if you're building a little piece of software this or that like it knows because of i guess all the other git repositories of people that built similar things like you're gonna want something like this right yeah no 100 and so like the way i try it is i have uh i got a uh an arcade machine in my house that's just a computer i got a home a media server in my house that we use and I have a home automation server in my house. And those were all like varying levels of working fine because I had set them up myself.

15:03I'm not ready to deploy Claude in a production store. I don't have a Shopify store. I'm managing client stores. I'm not going to do this to them, you know? So I deployed it on those home projects. And well, okay, the first thing it did was break my arcade machine. But after that, it was a learning experience and it set everything up. And it like, it was really cool and phenomenal. And what I learned with it was garbage in, garbage out. You know, it's very dependent on spec. It works better if you work in stages and milestones and be like, all right, here's all the context. Now let's come up with a roadmap and, you know, tell me how you're going to troubleshoot this problem.

15:38And then, you know, tell me when I'm going to check in, like add checkpoints to your work, give it really clear spec. That's when it works well. And then the other problem I found with it is, and maybe you've noticed this, especially with like web development, you finally give it a really clear spec and it builds what you want and then you revise it like sweet it's working and then by the third revision suddenly it like has suffered a head injury and can't achieve the most basic things well it turns out that this is um uh context rot where like for coding it's got to load all this info and it burns through the context window very quickly and so what they do is they say we're compacting our conversation they're not compacting anything it's just summarizing what it was doing stuff's gonna get lost when you do that especially when you do it successively.

16:22And so that's why like working in phases and having, you know, a mass, like a text file that it can reference of like, this is where you were, buddy. That's why it works so well. It's like, okay, you finish phase one. Now you clear the context and start phase two. That was my most recent unlock with development with Claude. That's like, oh, you know, that's how people are building these really complex projects is you, you got to break it down smaller.

16:45Kurt Elster:It's just wild. Like, do you think the rumors of the SaaSpocalypse are overblown? I have a friend who is a video company and they just put their IPO on pause because of, you know, all the swirling around here. I think like Shopify, you know, as a business model, no one's going to try to – it's going to be a while before anyone's trying to replicate that code base, I'd imagine, for how it works. But what about the app ecosystem? Like what is AI doing? Like you said, it helped you guys build your code better. Yeah, the problem is it helped everybody else build apps too. Yeah, exactly. Yeah. No, currently the, I mean, you could see it on Twitter, the wait time to get an app reviewed when you submit it to Shopify.

17:21And I bet this is true of like any app store with a review process has ballooned. It has really skyrocketed because so many people are just vibe coding an app and then submitting it. A year ago, it was really easy to weed those out because the apps would show up and they just didn't work right. You know, like you could tell a vibe coded app because the darn thing didn't work. Today, that's not the case. You know, now like the giveaways with the vibe-coded apps are like just the way they look.

17:48Kurt Elster:Because by default... M-dashes. Yeah. Well, it's more like by default, what I've discovered, the LLMs are not great at web development. Just like the visual portion of it. So especially CSS. And so they lean on frameworks. And so they often have like, unless you've specified otherwise, they get a real similar look. And it's because like they're all using Tailwind CSS, you know, or something similar. Claw just seems to default and enjoy Tailwind. but yeah now they all get this like that similar similar look it's like big headline over gradient background but yeah like you know you see i see posts from merchants uh who previously were not developers did not think of themselves as developers and they're like you know suddenly all these simple apps and tools that they were using you know and maybe paying a little bit of money for they're taking the time to try and figure out if they could vibe code it you know and make their own or find alternatives or make like theme customizations that'll do it and you you know what, they're getting there and it's working.

18:43Now what they don't realize is like there are limitations to it. And in some cases there are security vulnerabilities that you're creating unknowingly. And then the larger problem is like, okay, well now you're responsible for it. Now you have to maintain it. And it turned out like a lot of that monthly fee was peace of mind. But yeah, now it is a weird time on the internet. So like twice a day, I think to myself, I don't even know what's real anymore.

19:07Kurt Elster:Most of the scripts that you see people, So like any of the, you know, matter how deep and weird the video is on whatever, you can tell that they're like, this isn't just a rupture. This is a so-and-so or whatever. It's not X. It's Y. It's not X. It's Y. It's like all of our thoughts are being filtered through these machines now in some ways. And yeah, it's a wild time to be alive. I got to say everything hits at once. That's for sure. The problem is like when you're really used to the AI tools, you recognize the output immediately. And I don't know why. It's like it's so annoying. I know. Like NBC News did a survey.

19:4151 % of Americans say they use AI tools daily. All right, so just over half, a majority. Now, okay, it's self-reported. So what are they really doing? Like, oh, I used Siri today. Does that count? But all right, take that with a grain of salt. But the damning thing was 25%, one in four approved of AI. Okay. The majority of us are using it, but the majority of us don't like it?

20:09Kurt Elster:My 12-year-old hates it. She's like, she's very skeptical of it right now, which is interesting because she's going to be coming up into it fully. Oh, yeah. No, my kids, they're like, they swear they're not using it for their homework. I don't think they are. But they also are like, think AI is, like their view is just AI is bad. I don't know how that happened. I know. And I was very much like, you know, play, like these, you're not going to avoid these tools. Generative AI is here. Please. Like I encourage you play with it and understand its limitations and like where it's dumb, you know, and like I, they all did it and they played with it.

20:42And then we're like, it was early with chat GPT. I think they kind of decided it was a novelty and moved on from it. And now they're just like, all their friends are like, hey, I'm dumb and bad. Like, I don't know if it's like the edgy thing to say they're teenagers. I think so.