[00:00] Lexi: Here's a number. 65 percent of American shoppers say AI is coming into their Christmas shopping this year. And on the other side, just 8 percent of retailers say they feel confident using it. So I took that gap to the two people I know who actually build these systems, and I asked them what a retailer is supposed to do about it.
[00:18] Lexi: And between them they took away both of the answers. You can't out-build the thing your customer already has in their pocket, because that one knows them and your website doesn't. And you can't keep it off your site either, because a court spent 9 months on exactly that question this year, and the retailer lost. On a technical detail about where a screenshot gets processed.
[00:37] Lexi: One of them thinks this is fine. He thinks it's fine for everybody except Amazon.
[00:44] Lexi: Hi, and welcome back to the Aloudable HQ. This is Assistant to the CMO, I'm your host, Lexie Meskouris, and the idea here is pretty simple. Every week we take whatever just happened in AI and marketing, and we actually pull it apart.
[00:57] Lexi: Now, I don't build any of this myself. So each week I sit down with two people who do, Will Nash and Harjot Singh. Between them it's, honestly, over 20 years of building tech and brands from the ground up. And the best part is they've got no horse in the race. They don't sell the things we cover. They just see them differently.
[01:14] Lexi: So this week starts with a survey, and I want to be careful with it before we get going. Narvar, which is a post-purchase and returns platform, put out its 2026 Holiday Shopping Report on the 24th of August. They asked 1,348 consumers and 100 retail decision-makers. All of them American. There's no UK equivalent in it, so if you're listening here, the behavior probably travels but the number doesn't.
[01:38] Lexi: And the headline is that 65 percent of those shoppers say they'll use AI for at least one part of their holiday shopping. Break it down and it's 43 percent for gift discovery, 33 percent for comparing products and summarizing reviews. Then on the other side of the table, out of 100 retail decision-makers, 8 percent called themselves very confident using AI to improve the shopping experience. Will, signal or noise.
[02:03] Will: Signal. Um, definitely signal. And and I think the thing that makes it significant isn't really the size of any of these numbers, it's it's the direction of it. E-commerce has been at the frontier of internet technology for, you know, 25 years. And what you've got here, at least in the US, is a a situation where the consumer is actually ahead of the seller. Which for that industry is is quite unusual.
[02:25] Will: Now, the 65, I I'd be careful with. That's, that's obviously very woolly. It's it's a percentage of people who say they'll use AI for some part of it, which, I mean, that could be almost anything. Um. The number I'd actually look at is the 43. Gift discovery. Because that's a a genuine buying journey, and and that's going to be mostly cannibalised away from Google search. So if you're sat in a marketing team, that's the one that changes what your quarter looks like.
[02:50] Lexi: So the headline is doing the PR and the 43 is the one that costs me money.
[02:54] Will: Something like that, yeah. And the 8, look, I I'm not going to tell you what the 8 means yet. But it's it's worth watching, because you've got to consider where it's come from rather than where it is. The the headline reads small, obviously, in comparison to the other side. But that's, that's the number I'd be tracking, not the 65.
[03:10] Lexi: Tracking for what, though? Because 8 out of 100 is either a door standing wide open, or it's 100 people telling the truth about a genuinely hard thing.
[03:20] Will: Yeah. Um. I I'm not going to call that one, honestly. I I'd want to see the same question asked again next year before I said which of those it was.
[03:29] Lexi: Fair. And the flip is what gets me anyway, because it's normally the other way round. Retail had recommendation engines before most of us had a word for them. This time the customer walks in already holding the better tool. Harjot, same numbers. Where do you land?
[03:44] Harjot: Yeah, signal for sure. Um, and you know, with the mass spread of ChatGPT, of Claude, people are just leaning on these things more and more, right? Like people love that they can talk to this thing and it knows them, and it gives them back answers that feel considered, feel kind of bespoke to them. Um and the thing is, comparing stuff online is genuinely, genuinely hard, you know, you're often not that informed about the product in the first place. And that's actually the thing LLMs are really, really good at, right? You give it this kind of unstructured, messy request and it hands you back a structured comparison, um, in seconds. So yeah. Signal.
[04:13] Lexi: Right, and that's the easy half. What is a retailer actually supposed to do about it?
[04:18] Harjot: Um, and I think the honest answer, Lexie, is that the obvious move is the wrong one. Because the obvious move is you go, right, everyone's shopping with AI, um, we should put AI on our site. And every time a retailer tries to build AI on their own site, the AI doesn't know anything about the customer. It has to ask. It has to relearn stuff, right? Every single time. Um, and that's, you know, that's kind of the whole benefit of the thing people already have in their pocket. It's your ChatGPT. It knows you, and the retailer's website doesn't.
[04:41] Lexi: Mm. So every conversation starts back at zero.
[04:45] Harjot: And then the second thing, and this is, um, this is from having worked inside one of these businesses, so this is me talking from experience rather than, you know, guessing at it. You don't even know what model they're using. So you can't really trust it, right? Like consumers now have a light understanding of what a good conversation feels like, um, and some of them are more sophisticated, they know, oh, this thing uses this model and that model. That's just not the case with a retailer's AI tool.
[05:07] Lexi: Pinned to what, though? Does it just sit there?
[05:10] Harjot: Yeah, basically. Um, the likely way these things get developed, uh, from what I've seen, is they build the tool, it gets pinned to some version of a model, and then it just, it doesn't really get updated until absolutely necessary.
[05:22] Harjot: But the sharpest one, to be honest with you, is the third one. The best use of these tools, the thing they're actually best at, is doing the cross comparison, right? And retailers really just don't want you doing that. Because the moment you're doing a proper comparison, that gives other retailers a chance to put their product in for the running. They only really want you comparing between their own products, within their own site. So the thing it's best at is the thing they least want it doing. Um, so it's, yeah, it's contentious for them as well.
[05:46] Lexi: Wait, that's a horrible position to be sitting in. Because a genuinely good version of that tool walks your customer out of the store.
[05:54] Harjot: Exactly, right? That's the whole tension of it.
[05:56] Lexi: Although I want to separate two things there, because you've described a thing nobody really wants to build, and you've also described a thing that gets built and then left alone. Those aren't the same problem and I don't think they've got the same fix.
[06:09] Harjot: Yeah, no, that's, um, that's a really good split actually. And the second one, the one where it gets built and then left alone, that's just, that's a structural thing about how these get treated internally, right? Retailers don't see this as an entire function that's maintained by a team. They see it as a feature. And every time something is a feature, it gets built, and then it only gets attention when something surfaces it, you know? Either through some monitoring, or through the product person seeing how it affects the bottom line, and then it gets a bit of attention.
[06:33] Lexi: So it's nobody's actual job.
[06:35] Harjot: No. Um, and it's not that anybody decided to let it rot. It's just what happens to a feature when nobody owns it.
[06:42] Harjot: And to be honest with you, it would have to be a really big effort from a retailer team to be genuinely monitoring how good that user experience is, and what that translates to, and being really serious about it, right? That's a very big ask for teams that are just not set up in that way. Um, and look, I should be fair here, because if a retailer had an AI team specifically focused on AI product features, um, I actually imagine that would be a really good case. Like, genuinely. But most retailers are just not set up for this at all.
[07:09] Lexi: Alright. Will, we have been here before and I'm going to say so up front. Back in episode 5 you walked me through how these models decide what they think about a brand, and it was training data on one side and live web search on the other. So I don't need that explained to me again. What I want to know is whether that's still the whole picture, or whether there's something in it you'd add now.
[07:30] Will: Yeah, no, you're you're right that we've done the two. Um. And and I'd stand by that, but there's, there is a third one, and and I think it's the one that matters most for what we're talking about today.
[07:41] Will: So the first is from memory. Which is, you know, the model already knows something about your product before it does anything at all. And the bit I'd add now, which which I don't think we got into last time, is the timing of it. Because typically for a frontier model the training data is prepared somewhere between, uh, 3 to 9 months prior to that model launching. That's, that's roughly my estimate of it. So whatever the model thinks about you, it it thinks that from a snapshot of the world taken the better part of a year ago. And and that is very, very powerful, it's heavily weighted, and it's hard to dig out and correct.
[08:10] Lexi: Hang on. So whatever it thinks of me, it had already decided that before the thing even went live.
[08:15] Will: Yeah. Yeah, that's it exactly.
[08:18] Will: Next there is live retrieval, which we've done, so I I won't go over it. It's it's the layer where the model goes off and actually looks. Fine.
[08:26] Will: The final one, and this is the one that's new, is a live product catalogue. So so what I mean by that is uh, the model isn't reading a blog post about your product and it isn't reading a listicle about your product. It's, it's reading your product.
[08:39] Lexi: Which is new, right? That isn't a thing you could have said to me a year ago.
[08:43] Will: No, no, it's, it's fairly recent. And there are basically two shapes that takes. One is the platform working directly with the retailer, so, so OpenAI have done this, um, there are arrangements where the catalogue is being fed in more or less directly. And the other is where they're able to scrape kind of directly from a merchant page. And and that's where things like schema come in, so that's, you know, the structured markup that says this is a price, this is availability, this is a returns window.
[09:07] Harjot: Mm, schema, yeah. Machine readable, sat right there on your own page.
[09:12] Will: And and I think what's interesting about that layer, Harjot, is that it's the only one of the three where the retailer is, is genuinely a participant. The memory happened without you. The retrieval is, is you hoping the right thing is out there. But the catalogue layer is, is a thing you actually maintain. Um, which, which is either an opportunity or, or a whole new operational burden depending on how you look at it, and and honestly I think it's both.
[09:34] Harjot: It's both, definitely, yeah. Um, and it's the only bit of the three you could actually put a person on.
[09:40] Lexi: Hold on, because there's something in that I've been carrying since episode 5 and it doesn't quite sit right. Last time, the live search was the good news. You told me it eroded the advantage the big incumbent brands had baked in, that the layer was newly contestable, and I walked away from that conversation feeling pretty good about it. Today you're telling me the baked-in part is heavily weighted and there's basically nothing I can do. Those are two different moods about the same machine. Which one am I supposed to leave with?
[10:05] Will: Yeah, that's a very good question. And and I don't think those two things are in conflict actually, but I I can see exactly why it lands like that. Um. So so the thing is they're answers to two different questions, and, and which one you're asking depends entirely on what the model already thinks of you.
[10:21] Will: So, so if the model has never heard of you, if it has no opinion of you at all, then, then yes, live retrieval is the contestable layer and that's genuinely good news, and and that's who we were talking about last time. That was a challenger. You're, you're filling a gap. There's nothing in the way.
[10:37] Lexi: And in that conversation I was the challenger.
[10:40] Will: Exactly, yeah. But if the model already thinks something about you and the thing it thinks is wrong, or, or it's just dated, um, that's a much harder problem. Because you're not filling a gap at that point, you're, you're arguing with something. And and the machine hasn't changed between those two conversations at all. It's, it's the same mechanism. It just, um, it depends entirely which of those two you happen to be.
[10:59] Lexi: So the good news was never general. It was for whoever the model had never heard of.
[11:04] Will: The irony, of course, is that at that stage there's, there's nothing they can do. About the memory part, I mean. That, that layer is closed. The only thing you can really do is be wary of it, which, which sounds like a cop-out but it isn't, because you can't be wary of something you haven't looked at. So so that's where an audit comes in. Going and asking the models what they think of you with the lookup switched off.
[11:23] Lexi: You've been pointing at the page all morning, Harjot, and Will's just told me the page arrives last and weighs least. And I'm not going to ask you how a model decides what to trust, because you answered that in episode 10 and I remember it. So give me the narrower version. What can it physically read?
[11:39] Harjot: Yeah, um, I'll tell you what it doesn't read. It can't watch your video. It can't look at your pictures. What it can do is read information, and ideally structured information, right? So your product specifications, your product reviews, all these things sitting on the page. That's what it's actually taking.
[11:54] Lexi: Wait, can't or doesn't? Because I've put a photograph into ChatGPT and it told me exactly what was in it.
[12:01] Harjot: No, that's fair, um, and I should be careful there, because the models themselves can obviously look at images. That's, you know, that's not the point I'm making. The point is in these cases these are like indexer pipelines. And for the most part they're not going to be indexing everything on your site. They'll be really just compacting down and vectorising and indexing the relevant parts of your site. Um, and mostly it's going to be coming from text.
[12:21] Harjot: Whether or not these indexes use images now, it's possible, but it's kind of unlikely, because the cost is just way bigger for something like that. Um, these things would have to classify them ahead of time, right? Which means interpreting all of those images ahead of time, across every site, and that's, yeah, that's just unlikely to be happening. It's possible that during your search, during a live search, the multimodal model might look at an image and so forth. But in reality, if it gets the answer from the text, it won't need to. So practically, um, they just won't really do that if they feel confident enough in what they already know.
[12:51] Lexi: So it only bothers with the picture if the words already failed.
[12:54] Harjot: Pretty much, yeah. And that's kind of why I land somewhere different to Will on this, um, not that he's wrong about the memory part. But I actually reckon you want your product page to be scraped, right? You want the AIs being able to understand and compare and read as much structural data off your page as possible. And you want it to be able to take a really strong holistic overview and not be guessing. Because the more confidence it has in that, the more you're going to appear in the discussions that it puts forward, and in the suggestions it puts forward. Um yeah.
[13:21] Lexi: So one of you is telling me to go and argue with something that got decided last spring, and the other one is telling me the thing that decides is reading my page right now. Those are two completely different quarters and I don't think either of you is going to resolve it for me, so let me at least find out what each one costs.
[13:38] Lexi: Will, the audit, because I owe you an objection on it. And I want to name something first. In episode 10 you told people to go and find out what Reddit already says about them before they spent a penny on it. This is the same instinct, but it's pointed somewhere else. That one was a room full of people. This one is the machine's own opinion with the lookup switched off.
[13:56] Will: Yeah, that's, that's exactly the distinction and I I think it's worth holding onto.
[14:00] Lexi: So here's the objection. I run your audit. I find out the model thinks my brand is mid-range and slightly dated. You have just told me I can't change that. So the audit gave me a diagnosis with no treatment and I've spent budget to feel bad. What do I actually do on Monday?
[14:17] Will: Well, I I'd push back on the framing slightly, because the audit isn't the treatment and I don't think it's really the diagnosis either. It's, it's the thing that tells you which sentence to go and write. So so if you run it and what comes back is, you know, mid-range and slightly dated, that's, that's not a general problem, that's a specific one. And the treatment is that you make sure any live retrieval that happens speaks directly to that specific concern. So so whatever the model goes and finds when it does look something up, that objection has already been answered in the material it's reading.
[14:45] Will: And and that's a listicle, or, or it's a Reddit thread, we've, we've done that one -
[14:49] Lexi: We did a whole episode on that one, so I'm letting you off it. Though that does change what I'd use the audit for. I came in thinking it was a scorecard and you're describing a brief.
[14:59] Lexi: I want to come back to that 8 percent, Harjot, because Will wouldn't call it, and I should give you the rest of it because I only gave you the headline. Out of 100 retail decision-makers, 8 percent said they were very confident using AI to improve the shopping experience. 14 percent think greater use of AI shopping assistants will be the biggest behavioural change this season. And the coverage around the report says most retailers are still focused on shipping costs, discounts, operational execution. So is that a door standing open, or is that just 100 people being realistic about a hard thing in one Christmas?
[15:32] Harjot: Yeah, so, um, look, even outside of this report, yes, it is a competitive opening. And I say that having been at a luxury retailer and, you know, understanding a bit about how that kind of business is actually driven and how it's structured. Um, and I want to be fair about the risk first, because it's a real one. It is a big, big risk for these companies, right? These are businesses that are really optimising their bottom lines at this point. Nobody is sitting there with loads of appetite for a punt.
[15:56] Harjot: But, um, anyone who's able to pivot to almost an AI native approach, and has the confidence and the boldness to actually do that, and if they could get it right, you know, basically without all the downsides we've just spoken about, yeah, that would be a game changer. I think that personally. And I should scope that, because I don't think it's necessarily true of the smaller companies, the DTC brands, the non-large retailers. I'm talking about a retailer who's got somewhat of a real product function, right? Someone who could actually build the thing.
[16:19] Harjot: And it can't be another chat assistant. Everyone is sick of those. No one wants to talk to another chatbot on a website.
[16:25] Lexi: So what is it instead?
[16:27] Harjot: Um, it has to be a tailored shopping experience, and maybe that's a different modality, like voice. The form factor can't just be, you know, hey, tell me what you need, right? It needs to be much more guided, much more integrated.
[16:37] Lexi: Every retailer already thinks their site does that, though. What's the actual difference?
[16:42] Harjot: Um, a mixture of some of the flows they've already been building, these kind of very targeted flows that guide a user straight through to the best way to buy something, but then whilst they've got that, they can almost have a shopping experience similar to what you'd have in store. You know, you can ask questions while you try stuff on, see what pairs well together. So it becomes holistic and not boxed into a specific flow. Because that's one of the main weaknesses these retailers always had, right? You're locked into a very specific workflow that the retailer pushes you through, because that was only what was possible before.
[17:09] Lexi: That's the first thing anyone's described today that I'd actually want to use. Alright, let me get concrete before we move, because there's a budget sitting on someone's desk right now and it's already committed. Same question to both of you. It's Q4. What's the line item that just got less valuable?
[17:25] Will: Yeah, I mean, really the, the number one thing is if you're paying for any kind of, uh, complex copywriting. You know, florid text, florid prose. And I I want to be careful here because I'm not saying go and fire your writers, that's, that's not it at all. If you've got it, keep it.
[17:41] Lexi: Good, because I was about to get defensive on behalf of every copywriter listening.
[17:47] Will: No, no, that's not the point at all. But at this stage it's, it's fairly clear we're moving much more towards a technical style of writing. Really making the information clear. So so I suppose you may still pay for it, you'd just, you'd go to a different technical writer and give them a different brief. And, and that brief is really focus on the facts and figures of the product.
[18:07] Will: Um. And then there's a specific inside that which I think is, is the bit people miss. Because when you're comparing products, a lot of the time you're comparing the same product between different sellers. It's, it's the identical thing. So the price is level, or, or it's near enough level. And at that point, what's the agent actually deciding on?
[18:23] Lexi: Honestly? I don't know. Like, brand, maybe?
[18:28] Will: And and it's the stuff a regular human being would just skim straight past. Your returns policy, your, your shipping information. Um. Which, I mean, nobody really writes that properly, do they, it's, it's usually a template somebody pasted in years ago and never looked at again. But when an agent is comparing 3 or 4 different places to buy from and the price is a tie, that, that is the sort of information that will be picked up by those bots. So, so that's certainly worth considering.
[18:53] Harjot: Yeah, mine's the one we've already done on this show, so I'll just say it. The load has shifted away from paid search, right, away from it being the thing consumers directly use to find things out, and towards the things models read. So paid search is a lot less relevant. Keep the approach up, but you want to be shifting more into AEO. That's it, that's all I've got on that one.
[19:10] Harjot: Um, but you know what's funny, thinking about all this. I did think, wouldn't it be interesting if these retailers started building MCPs? And, um, actually, maybe that is the right approach. Because, you know, as we've been saying, people are starting to shop with their ChatGPT, right, that's where they already are. So maybe you go a step further. You build an MCP that's access to your product catalogue, or an ordering experience, or something like that, that the customer can just use.
[19:32] Will: Which is the catalogue layer again, isn't it. Just, just with the retailer opening the door themselves rather than waiting to be scraped.
[19:38] Harjot: Yeah. Yeah, and, um, huh, that's actually kind of the reverse of everything I've been saying, isn't it. Because everything today has been the retailer trying to keep up with the shopper's agent and losing, right? And this is just letting it in through the front door instead. You stop trying to drag them onto your site to talk to your thing, and you just go to where they already are. Um yeah. I think there's something in that, actually. That'd be super, super interesting to watch.
[19:59] Lexi: And I want to sit on that for a second, because it's a good idea and it's about to get complicated. You've just described a retailer deciding to let the agent in. The whole second half of what I've got today is a company that spent 9 months and a great deal of money trying to keep one out.
[20:13] Lexi: So. Amazon sued Perplexity in November of last year over its Comet browser agent. Breach of terms, degraded shopping experience, privacy risk. And in March a district court gave Amazon a preliminary injunction, which blocked Comet from shopping on Amazon. Then on the 4th of August a Ninth Circuit panel vacated it, finding Amazon unlikely to succeed on its Computer Fraud and Abuse Act claims, and the California equivalent.
[20:39] Lexi: And the reasoning is the whole story, so stay with me, because it's architectural. Comet takes screenshots on the user's own machine. It sends those screenshots to Perplexity's servers to be processed. It gets navigation instructions back. And the court found that Perplexity itself never directly communicates with Amazon's servers. So it was the user who accessed Amazon's computers, not Perplexity.
[21:03] Lexi: Two caveats before anybody gets excited. It's narrow, it's only those two statutes, and the breach of terms and contract claims are expressly still open. It's preliminary, not a final judgment. And the panel went out of its way to note that an agent with more autonomy, or one talking server to server, could come out differently. Harjot, this is the under-the-hood one, so give me the gut reaction.
[21:27] Harjot: Noise.
[21:28] Lexi: Noise? Amazon just lost a ruling it spent 9 months chasing.
[21:33] Harjot: Um, and I don't mean it doesn't matter, right, I mean it can't work. Like, we're going to see more and more of these sorts of claims opening up, this whole kind of defence against robots using sites, and ultimately it's all futile. There's no real way of stopping any of this stuff happening.
[21:47] Harjot: And to be fair to them, it does hurt Perplexity, right? It does slow them down. It does get in their way, and it does give Amazon a bit of time to get their own act together. Um, so it's not that it achieves nothing. But really it's just a time-buying exercise more than it is anything else. I mean, they must know they can't win this.
[22:01] Lexi: I want to test whether futile is analysis or whether it's a shrug. Because this is Amazon. That's the deepest legal bench in retail sitting next to the deepest engineering bench in retail. What could they actually do?
[22:14] Harjot: Yeah, I mean technically speaking they could, um, they could do user agent sniffing. They could do some sort of fingerprinting to determine whether the browser is acting like a human or not, right? The attributes it sends across when it makes requests, how it makes those requests, the timing of it. There's a bunch of stuff they can do there.
[22:30] Will: Sorry, can I just, um. Fingerprinting, is that, is that the same family of thing as a captcha, or is that separate?
[22:38] Harjot: No, it overlaps very, very heavily, right? What captcha does, what general bot blocking behaviour does, it's all the same toolbox. Like this is not new territory for them at all.
[22:47] Harjot: However, um, this then enters the arms race, so to speak, of human versus bot prevention. And now, you know, with the advent of LLMs, they've been able to come along and actually offer the possibility of bypassing some of those things. So it's just a matter of time until someone trains a model that learns how to pass their specific tests, right? So again, it's, yeah, it's a pointless game. They could try and spend a bunch of money on this, but it's just a waste at the end of it.
[23:11] Lexi: Then why file? Because if they know, this is an expensive way to lose in public.
[23:16] Harjot: Um, I suspect it's two things. Well, three. The first one is it buys them time to figure out how to approach the strategy, right, they haven't got one yet. The second one is just to test the waters, and see exactly how, from a legislation perspective, how things are playing out. What's in the air so far. You file and you find out.
[23:33] Lexi: Which is a very expensive way to read a room.
[23:36] Harjot: It is, yeah. And then, um, from a petty point of view, to be honest with you, it just makes Comet a bit useless. You know? And that might eat into their base a bit, and allow Amazon to get people to jump off Comet onto whatever the Amazon native approach is, which is presumably Rufus. So yeah, it's, it's not nothing for them.
[23:52] Lexi: Right, so nobody was ever really trying to win it. They were buying 9 months and making the competitor annoying to use in the meantime. Which I can respect, in a bleak way.
[24:02] Lexi: But here's the thing I genuinely want to understand, because the entire case turned on it. Comet takes a screenshot on your own machine, ships it off somewhere else to be processed, gets instructions back. That is not the cheapest way to build an agent. So did somebody's lawyer design that? Is that architecture a hedge?
[24:20] Harjot: Um, no, I mean, I can't tell you what was in someone's head, right, and it may well suit them legally, that's fine. But that's not why it's built that way. It's more a consequence of making this accessible to everybody.
[24:29] Harjot: So, like, at some point the LLM is going to have to interpret that screenshot. Um, and this is how computer use generally works with other models, right? The other big assistants do the same thing, the coding agents that drive a browser do the same thing. They have to interpret the image. Which means their model has to read it. Which means it can't be done on device, unless they have an on-device model, which they don't have.
[24:47] Will: So so the picture has to leave the machine. There's, there's no version of this where it stays on the laptop.
[24:53] Harjot: No, and even if they did have something on device, um, they're going to be so hardware constrained, because they can't control the hardware their customers are on, right? Like they've got no idea what you're sitting in front of. And most people just, they won't have a GPU attached that's anywhere near powerful enough for this kind of inference. So the processing has to leave the machine. There isn't another shape for it.
[25:10] Harjot: So, um, yeah, and that's kind of, huh, that's actually the funny part of the whole thing when you think about it. The entire case turned on where that screenshot gets processed, and where it gets processed is decided by, you know, what GPUs cost and who's got one. Nobody sat in a room and designed that for a court. It's just, that's the only way you can ship this to normal people. Which is insane, right? That's, yeah, that's the bit I'd not have predicted.
[25:31] Lexi: So Amazon lost on an accident of GPU economics. That's what I'm going to be chewing on all week, because it means the law ended up on the shopper's side of the door for a reason nobody in that courtroom chose. Will, you've been letting this one run, and I think it lands right on the layer you were describing earlier.
[25:47] Will: Well, first thing I should say is I I haven't got a view on the ruling itself. Um. That's, that's genuinely not my area and I wasn't across this one, so, so I'm not going to pretend I know how it lands. But the mechanism you've just described does, does bear directly on the thing I was talking about earlier, so, so let me go at it from there.
[26:04] Will: Because if the agent's way in is a screenshot of a rendered page, that's, that's the agent going through the front door like a customer. Squinting at a picture of your website, effectively. And and the layer I described earlier, the live product catalogue, that was the one where the retailer is actually a participant. That's, that's the one you maintain. So so if everybody's route in is a screenshot, then, then I I don't really know what that layer is for. And that's, that's genuinely a question rather than a point.
[26:25] Will: Now, the reason I'm sceptical about the tidy version of this, and, and this is just as, as someone watching it get built rather than a prediction, is that we've, we've already had a go at the tidy version. So OpenAI launched Instant Checkout inside ChatGPT, that was, uh, September 25 I think, with Shopify and Etsy merchants. Direct plumbing. And, and it was pulled about 6 months later, so, so roughly March. And and the reasons are quite instructive actually. Almost no merchant adoption. Buyers preferring to finish on the retailer's own site where their payment details already sat. And, and real-time catalogue sync across millions of merchants just, just not scaling.
[26:58] Lexi: Wait, they pulled it? Like, I don't think I ever saw that reported anywhere.
[27:02] Will: But but here's the part I keep coming back to. They kept the discovery half. So so ChatGPT still does the product discovery and it routes the purchase intent back to the retailer's own checkout. The, the direct plumbing is the bit that got withdrawn, and, and the layer that survived is exactly the one we've been talking about all episode. So, Harjot, that's my question really. Is there a way for a retailer who genuinely wants this traffic to invite the agent in properly? Or, or are we all just stuck with agents reading pictures of web pages for the foreseeable?
[27:31] Harjot: Yeah, Will, so, um, I think the industry is going to eventually move to direct server to server. Like, everyone can get around it with the screenshotting approach, fine, that works today. But at some point this is going to hurt people on token costs, on efficiency. You're paying to have a model look at a picture of a page to extract information that could have just been handed to it. Um, it just seems very wasteful from an industry point of view, from an economic point of view. And from the customer's point of view it's lose-lose all around, right?
[27:57] Lexi: Although that's an economic argument, not a legal one. Nothing in that ruling actually pushes anybody there.
[28:03] Harjot: No, exactly, and that's why it just doesn't seem likely or viable that screenshotting is going to be the standardised, unified path that everyone agrees on. Um, and what we're really waiting for is more and more of the other big tech companies who are working on this stuff to push out more on the standards side. You know, collaborating with other people in industry, these groups coming together and pushing out more aggressively these standards that will become accepted, and, um, will maybe become somewhat even legislated for. That's the path.
[28:29] Harjot: And ultimately, you know, how do you even enforce this stuff? Like you can try putting in little blocks and whatnot, but agents basically understand context and what's going on, right? They're not static. So they can adapt, and very quickly get past stuff. And take even a much weaker model, you know, something a couple of tiers down, not a frontier model at all, and it would know how to adapt. Let alone the good stuff.
[28:48] Will: Can I, um, can I ask the thing I actually want to know here. If the agent can now walk into Amazon and compare, what's it comparing on? Because, because if the price is level everywhere, we're, we're straight back to the returns and the shipping, aren't we. And and I genuinely don't know who that favours.
[29:02] Harjot: Whoever bothered to write it down, to be honest with you. Um, that's, that's basically it.
[29:07] Lexi: So who does this actually hurt? Because Amazon made 19.8 billion dollars in advertising in the second quarter alone. And if a model picks the product, nobody is looking at a sponsored placement.
[29:19] Harjot: Yeah, I mean, look, this sucks for Amazon ultimately. It sucks for people whose primary marketplace is ads, um, genuinely, I'm not going to pretend otherwise. But this is the reality that we're moving to, right? Which is that the services people are going to be using are fundamentally changing. And, you know, this is my bet, not a number I've got from anywhere, but maybe people are going to be 90 percent in ChatGPT, in Claude, in the next 5 years, rather than browsing websites. And agents browse the websites on their behalf. So this is happening one way or another.
[29:47] Harjot: And ultimately this is great for everybody else, right? Because for everybody else it means more exposure, it means more avenues, a new channel to be able to get their products out on. You know, to some extent Amazon has monopolised marketplaces in a way that's very advantageous for them, um, but not really so much for others. So yeah. It's, it's bad for Amazon. I'm not at all sure it's bad for retail, you know, those aren't the same thing.
[30:06] Lexi: Last thing on this, and it's the one I'd want answered if somebody were selling to me. Because there is now a whole category of vendor out there selling agent-ready commerce. Schema, agent-friendly checkout, the pitch. What would they never admit?
[30:21] Harjot: Ha, um, it's funny, isn't it, that someone can sell readiness for agentic commerce. Because the thing itself is so early, right, it's so early days that it itself is not really ready for public consumption. It's really experimental. It's really just yet to be seen how this fully emerges and plays out. Um, so I think, yeah, what they're selling you is maybe the ability to enter into today's experiment. But they're not selling you the future, because they also don't have the future.
[30:44] Lexi: Monday morning. And the two of you did not give me the same advice this week, which I'm actually glad about, because the difference is where the useful part is. Harjot, I'll take yours first, because it's the one somebody could do this afternoon. Christmas is coming, they're not re-platforming, and we've just established they can't buy the future off anybody. One change to their site this quarter. What is it?
[31:04] Harjot: Yeah, structured information all over the place, um, even if it's hidden to the customer. Structured information, add some FAQs, add structured information. Especially if your site is highly visual.
[31:14] Lexi: Wait, hidden how? Because hiding text on a page to feed a machine something different from what the human sees is about the oldest penalty in the book. And these things are reading rendered pages now. So do you mean actually hidden, or do you mean something else?
[31:29] Harjot: Ah, no, yeah, that's, um, that's a fair pull-up, let me correct that. Visually de-emphasised. That's what I mean. Collapsed FAQs are fine, spec tables are fine, don't just hide stuff. Um, and again, it's the same as accessibility, right? You can't just make it pass the standard as a checkbox exercise, because it's really just bad for everyone. Bad for the people who actually rely on the accessibility primitives, and bad for your customers too. You never end up with a good site that really truly incorporates the principles, um, the spirit of it. So it's the same thing here.
[31:55] Lexi: And I have to be straight with everybody here, because we have told people to put a question and answer block on their page before. That was episode 5, and it was Will who said it. So I don't want the thing to add. I want to know how I'd tell whether it worked.
[32:07] Harjot: Yeah, and that's, um, that's the better question anyway, to be honest with you, because the FAQ isn't really the interesting part. The test is. You have to remember how a machine would read and parse your website. And it's kind of like accessibility, right? If someone was using a screen reader, you'd want to look at some of these same things. Site structure, heading structure, the clarity of those headings, so that the page makes total sense without the visuals or the layout being relevant at all.
[32:28] Harjot: So switch the pictures off in your head and read it. Um, and if the page stops answering the question once the pictures are gone, that's your answer. That's where the machine stopped reading too.
[32:37] Lexi: Will, you've got a different one, and I don't want us to blur them together, because I genuinely don't think these are the same morning.
[32:44] Will: So mine's, um, mine's really two things and the first one's very small. Put your returns and shipping terms on the product page. As words. As text on the page, not, not behind an icon or a tab or a link off to a policy page somewhere. That, that's the thing that gets read when the price is a tie, and, and most people have never written it out properly.
[33:02] Harjot: Yeah, and the returns one is, um, that's just a paragraph. Anybody can do that this week.
[33:07] Will: Um. And then the second one's different, because it's, it's not really a fix, it's more of a diagnostic. Go and ask the models about your own brand with the search switched off. So so that means actually controlling the tools the model has access to, turning the lookup off, and, and doing it in incognito as well so you're not getting something shaped by everything you've ever asked it.
[33:25] Lexi: Which I'd have got wrong, honestly. I'd have just asked it on my own account and believed whatever came back.
[33:31] Will: And and that's the common thing, yeah. Whereas this way you're seeing what it believes rather than what it just went and found. Um. Because that tells you what you're genuinely dealing with, and, and honestly it tells you whether the page was ever the problem in the first place.
[33:44] Will: And, and look, I I don't think Harjot's page work and my audit are in competition, they're, they're just answering different questions. One of them's, uh, it's about what the machine can read when it comes to you. The other's about what it already thought before it turned up. So.
[33:57] Lexi: And what you do about what you find, that's episode 10, and I'm not making either of you say it twice.
[34:03] Lexi: What I came in with this week was a gap. 65 percent of shoppers on one side, 8 percent of retailers on the other, and I assumed the story was that the retailers needed to hurry up and catch up. And what the two of you did was take away both of the ways of catching up. You can't out-build it, because the thing already in your customer's pocket knows who they are and yours starts from nothing every single time, and nobody inside the building owns it anyway. And you can't shut it out, because a court has now decided that the person holding the phone is the one who walked in, and it decided that on the basis of where a screenshot gets processed, which comes down to who owns a graphics card. So what's left is the least glamorous text on your entire website. Your returns terms, written out in actual words, so that a machine which cannot see a single one of your photographs can still work out why it should pick you. Will, Harjot, thank you both, genuinely.
[34:52] Lexi: And if you're listening, thank you so much for being here. We'll be back every week with something new.
[34:57] Lexi: Before we go, I have to tell you something. Every voice you've heard today, all 3 of us, me included, it's all AI. But the opinions are absolutely ours, and the judgment behind every word is human. That's the part you can't replace. And that's what Aloudable does. We take what you'd write, turn it into a show in your own voice, and we keep the taste where it belongs, with a person.
[35:18] Lexi: And this episode is about to become the exact thing we've spent an hour describing, a piece of text that a machine will read long before a person does. So if you want to hear what that would sound like for your brand, there's a link in the show notes.
[35:31] Lexi: And that's the tea. I'm Lexie Meskouris, this has been Assistant to the CMO, and we will see you next week.