Podcast: Re-evaluating agentic commerce (with Andrew Lipsman)

On this week’s episode of the podcast, I am joined by Andrew Lipsman to discuss the evolving landscape of agentic commerce and retail media. Andrew is an independent analyst and consultant who runs the Media, Ads + Commerce blog, which covers the retail media landscape. Prior to this, Andrew held roles at eMarketer and comScore. This is Andrew’s third appearance on the podcast.

In our conversation, Andrew and I take stock of the evolution of the concept of agentic commerce since we last spoke six months ago, moving from the initial hype of autonomous AI agents to a more grounded reality of AI-assisted shopping experiences. Our conversation examines the failure of independent instant checkout experiments, the strength of established retail ecosystems like Amazon and Walmart, and the emerging opportunities in performance television and in-store digital advertising. Among other things, we discuss:

• Whether the failure of instant checkout experiments signals a permanent preference for direct retailer relationships over AI intermediaries
• How the paradox of choice and the need for basket building hinder the efficiency of single-option agentic transactions
• If the success of Amazon’s Rufus proves that AI utility belongs on retail platforms rather than independent LLM interfaces
• Why the western market’s fragmented ecosystem makes the development of a Chinese-style shopping super app highly improbable
• What the massive investment in AI infrastructure means for the competitive landscape against established giants like Amazon and Google
• When performance TV and in-store retail media will finally become a core priority for brand-focused chief marketing officers
• How Shopify’s role as an audience network might evolve as direct-to-consumer brands seek diversification beyond Meta and Google

Thanks to the sponsors of this week’s episode of the Mobile Dev Memo podcast:

• ⁠⁠⁠⁠INCRMNTAL⁠⁠⁠. True attribution measures incrementality, always on.
• Xsolla⁠. With the Xsolla Web Shop, you can create a direct storefront, cut fees down to as low as 5%, and keep players engaged with bundles, rewards, and analytics.
• ⁠Branch⁠. Branch is an AI-powered MMP, connecting every paid, owned, and organic touchpoint so growth teams can see exactly where to put their dollars to bring users in the door and keep them coming back

Interested in sponsoring the Mobile Dev Memo podcast? Contact Mobile Dev Memo advertising.

The Mobile Dev Memo podcast is available on:

• YouTube
• Apple Podcasts
• Spotify

Transcript

Eric Seufert: Welcome to the Mobile Dev Memo podcast. I am your host, Eric Seufert, and I am joined today by Andrew Lipsman, who is rejoining the podcast for the third time. Andrew, welcome.

Andrew Lipsman: Thanks, Eric. Great to be back.

ES: It is good to have you back. We spoke a little more than six months ago, and I wanted to use this conversation today to take stock of what has happened in the media and advertising landscape since then. Before we jump into that, please reintroduce yourself to the audience.

AL: I am Andrew Lipsman, an independent analyst and consultant. I write at the Substack Media Ads and Commerce. Previously, some folks may know me as an eMarketer analyst where I covered retail, e-commerce, and specifically retail media. A lot of people know me more as a retail media expert.

ES: I will link to the blog in the show notes. I recommend that anybody following this space follow your blog. It is great, and you provide a lot of fantastic insights, particularly in the retail media space, but also broadly in the digital ad space.

We spoke just over six months ago. If I remember correctly, the genesis of that conversation was that I had written on LinkedIn professing some skepticism toward agentic commerce as it was being depicted. You jumped in on that, and we shared a similar sentiment. I then wrote a piece that went fairly viral called “Agentic Commerce is a Myth.” The title was not fully reflective of the point I was making. It was not that agentic commerce will never happen or that it is fundamentally impossible; my point was that independent agentic commerce faced a lot of headwinds and frictions because of natural tensions with the big retail platforms, and also the superiority of the advertising model relative to the affiliate links model that OpenAI ended up introducing and then walking away from.

We had shared an outlook at that time that agentic commerce seemed unlikely to take root. That was a very unpopular position. Now, enthusiasm for agentic commerce does seem to have waned, especially now that ChatGPT has abandoned instant checkout and Walmart, which was the largest retail platform to partner with ChatGPT, has integrated Sparky into that platform as an app. They have walked away from that affiliate integration. What is the state of agentic commerce now?

AL: Let me rewind really quickly to six months ago. When you finally articulated the skepticism, I was just coming out of conference season where I was talking to a lot of retailers. They had become fully convinced in a few weeks’ time that all their traffic was going to go away in the next couple of years, that their retail media businesses were going to be vanquished, and that this crisis of epic proportions was coming. Everybody seemed to buy into this hook, line, and sinker so quickly. I felt that the consumer behavior has to happen first. My lens is always consumer behavior, and I just did not see it because I had seen so many trends where the technology always led the consumer behavior and it never manifested, voice commerce being the most obvious one in this space.

I voiced some skepticism as well. My article is titled “Agentic Commerce is a Collective Hallucination” because I saw the industry collectively hallucinating this thing into existence that I did not think was going to happen. A lot of the conversation became a definitional debate over what is agentic commerce. The definition that you and I used, and that a lot of the pro-agentic people believe in, is this fully autonomous, end-to-end decision-making on behalf of the bot and bot-to-bot communication and transaction. Even if there is minimal human intervention, I just do not think this will ever happen.

What has happened since then is a series of things showing that the behavior really is not taking off and probably not likely to take off. The pro-agentic crowd has moved the goalposts to include anything that is AI-assisted. Any transaction that might start in an LLM or even on Amazon Rufus or Walmart Sparky is now agentic commerce. I do not totally buy it. I think we have all started to use agentic commerce as a shorthand for all of this stuff, but I am a proponent of using terms like AI assistance.

Even more recently, the big news was that OpenAI, about six months into this instant checkout experiment, backed away from it. Walmart had said that it just was not converting and was not working. No surprise here. Now OpenAI has moved to a new strategy, which is the super app strategy. It was not working plugging into the LLMs organically to have checkout there. Now they are asking the retailers to take extra steps and actually build mini-apps within the ChatGPT app. This is another thing that has been a fever dream of all the big tech companies over the years. They have always wanted to create the super app and emulate what happened in China with WeChat, but that is another thing I am very skeptical of. WeChat came out of a moment in time in which the majority of Chinese internet users were not on the internet. You had mobile, social, and e-commerce all happening at once. As those things evolved together, you actually had the potential for a super app experience. In Western markets, those things have evolved completely separately. There is no way to take those independent things and tangle them back together. That is why I think this super app strategy is a pipe dream.

ES: You wrote a really well-argued piece about that recently. You are very skeptical of the super app approach that ChatGPT is taking.

AL: It creates this chicken-and-egg scenario now where every retailer already has a very well-designed e-commerce experience. It is built for purpose. Now you are asking retailers to start building a new experience and manage essentially a new e-commerce touchpoint when the consumer behavior has not been demonstrated. Why am I going to invest all those resources until the actual transaction behavior is there? We have actually seen this happen before and it failed. We saw Facebook Shops, Instagram Shops, and these other e-commerce efforts on social networks, and they never worked.

ES: We have also seen it with just alternative app stores. It just has not worked. Not to say that alternative app stores cannot work, and maybe there is an opportunity now with some of the rules being relaxed as a result of the Epic court cases, but the Amazon Fire App Store, for instance, never took off. The problem is exactly what you stated. It was just this dance that you would do as a developer every year. I was working at Rovio as the VP of User Acquisition, and Amazon would come to the headquarters in Finland and try to pitch Amazon Fire and how we should be on the Amazon App Store. We would ask what they were going to pay us to do that, because I cannot take a bunch of engineers and direct them to spend X months building for your platform. Then they would have to offer some money to make that attractive. I do not believe ChatGPT is doing that for the app store. I agree. I do not think the ChatGPT app store is going to take off. First of all, it is very clunky to use, and that could change over time, but I tried it with Zillow. We are being opportunistic with house hunting right now, so I am using Zillow every day. I thought if I could go and just use natural language to do house hunting, that would probably be much easier, but it just does not work. If you go to that app on ChatGPT, it is utterly dysfunctional.

AL: There is a lot of work to be done to get things up to par. That is the hard work in building these new markets, but you are hitting a larger point, which is that these are general-purpose apps. They are not built for purpose. Consumers know to make those trade-offs. There is a certain moment in time when the general purpose works for you and might help you narrow some things down in the research process, but once you are deeper into the funnel, you want the app that is built for purpose.

ES: It is not always the case that natural language is an easier or more convenient interface than clicking some checkboxes. Sometimes it takes longer to type out. If I am ordering a pizza, is it longer to type it out? There was a big announcement that Starbucks was integrating into ChatGPT to discover new drinks. Who is going to use that? If I am ordering a Starbucks drink, is it faster to click a button or is it faster to type out that I want a Grande Pike with a splash of milk? That is not faster.

AL: Good apps are already pretty simple. It is a couple of clicks. The agentic commerce use case always made me laugh because so many of the commodity purchases that people talk about being a use case for are two clicks on Amazon. This is not a hard thing to do already. The reality is that we are seeing search evolve, but as I said, these other apps are fit for purpose. Why do we think that the traffic and behavior is going to move upstream to the LLMs rather than the LLM technology moving onto the site? What we are seeing is Amazon Rufus take off because people know to go to Amazon for a purchase, and now Rufus is actually helping provide that assistance along the path to purchase.

ES: I want to be clear here because in our original conversation, we were addressing substantial issues like incentive misalignment and frictions with independent agentic commerce. It was not just a case of it taking longer to type out than to click. The actual substance of my resistance to this agentic commerce idea was that Rufus will capture the agentic commerce opportunity on-site. Where there is opportunity to meaningfully reduce friction in commerce through natural language, Amazon will capture that. It will not exist off of Amazon. There are too many places where the incentives break down. E-commerce is not fractured; it is very concentrated. Advertising is very concentrated. You are not going to see places where independent agents whittle away this engagement from the incumbents because they are fundamentally concentrated. That was the substance of my skepticism.

AL: We have seen this history play out twenty years ago with the dawn of search. It started with Google, and then that search technology moved onto the e-commerce sites. Now we have lived in this steady state for a long time where some of the activity will happen in Google initially and then it moves downstream onto the site. I think it is a direct parallel.

ES: I want to talk about what the demonstrated engagement problems were. Walmart noted that it saw three times worse conversion from instant checkout than with click-out purchase. That defies what a lot of people cite as the advantage of agentic commerce in the abstract—that you are going to be having more intent as you discover these products and therefore you are going to be more primed to convert. That is actually not what was observed. What do you think explains that experience? People might say these are early days, but my rebuttal is that it is late. If what we saw with the instant checkout experience was so disastrous that OpenAI walked away from it in a matter of months in wholesale favor of ads, then it is over. Why do you think Walmart had that experience?

AL: Two main reasons jump out to me. The first is that we have seen the buy button experiment many times over, and consumers have already weighed in with their vote. They do not like buy buttons. They do not like to transact through an intermediary. They tend to like the direct relationship with the retailer. They want to know that they have a throat to choke if something goes wrong. There is just a lack of trust with things like financial security. People ultimately will get over those sorts of inhibitions, but people just do not like to transact in those environments as much.

The other thing is that if agentic commerce is working as it was proposed, then they are being delivered a single option. Consumers actually do not want that single option. It sounds nice and efficient, but you actually get less satisfaction when you only have one choice. There has been a lot of academic research around the paradox of choice and this inverted U-curve where on one end, if you have too many choices, it is overwhelming and you end up not liking that result, but if you only have a single option, you also are not satisfied with the purchase because you do not know what you are missing out on. You can end up with the exact same product, but if you have three or four options versus one option, you are going to be more satisfied when you have three or four options. Consumers just do not want that single option at the end of the day. They want to feel confident and build conviction in that purchase when they make it.

ES: I think there were other issues too, including the point that this only facilitates one purchase. The goal of an Amazon or a Walmart is to get you to fill a cart with lots of stuff and adjacent related products, not just to drive a single purchase.

AL: Again, it is not fit for purpose. Then you say it is not working because you cannot basket build, and then you start building all these additional bells and whistles and utility. All of a sudden, what have you built? An e-commerce site. Why don’t I just click through and go to the e-commerce site in the first place?

ES: If you want to address the functional shortcomings of this, you are not just building a better discovery engine on top of these retail platforms; you are building a better version of these retail platforms. If that was possible, that would have already happened. If you could unseat Amazon, you would have already done that. The prize for doing that is too great not to incentivize people to try, and it has not happened. These companies are too unfocused. They are all over the place, throwing spaghetti at the wall. What makes us think they are going to focus enough to actually execute on this? Yes, they have a lot of resources, but they have also got to start generating revenue pretty damn quickly or they may not exist a couple of years from now.

AL: Thus ads. People would say OpenAI can go back to the trough whenever it wants, and it just did. It just completed the single largest private fundraise in history. But they do not have as much money as Amazon. These companies that it purportedly wants to compete against in this use case are far larger and richer and have more machine learning engineers than they do, plus a massive bank of data which is the actual asset. OpenAI only has a couple thousand employees and you are going up against companies with hundreds of thousands.

ES: What I found really interesting about that experience was that it was the most controversial thing I had ever written. I got more pushback on that than anything I ever wrote related to privacy or anything else. It was an exceptional amount of ire. I had spoken about this with Ben Thompson because he had a similar experience with one of the pieces he wrote. His point was that he thinks the stock market is being propped up right now by AI and there is a lot of fear and uncertainty around what happens if we just cannot prove out a commercial use case here beyond companies integrating on the enterprise side. On the consumer side, what if there is just no robust use case? I read in the Wall Street Journal this weekend that they have the AI-adjusted Mag 7, so they swap out Tesla with Broadcom. If you remove the AI-adjusted Mag 7 from the S&P, the S&P is down on the year. I do think that maybe explains a lot of the hostility.

AL: It is possible. I took a lot of arrows after my article as well. I have never really experienced that. One of the reasons maybe is that I am not on Twitter, but you see the reports of what happens when people are anti-Bitcoin or thought NFTs were BS. It was kind of wild. Maybe it does ultimately come down to that. The whole stock market thing—first off, I do not invest in any individual stocks. I have no horse in this race. You and I as analysts, and Ben Thompson as an analyst, if there is one thing that we all do the same, it is that we compulsively look at both sides of the issue. We are testing our own assumptions and trying to see what argues in favor of this happening and what argues against. If all the analysts in this ecosystem ultimately come down on the side of being skeptical, I think it is a good time for the rest of the industry to start questioning things as well. It was redolent of the NFT phase, this unbridled enthusiasm. How could you be against this? This is so obviously the future. I kind of question what kind of future you were even depicting if you think this is where commerce is going to head. I have been doing my Prosperous Society podcast series, and part two was about agentic commerce. I made the point that if you think AI is going to obviate the need to work, why would you want it to obviate the need to use the fruits of this labor to express yourself? If AI is taking my job or doing my job for me, do I also want it to do the thing that I use the benefits of that job for, which is buying stuff? I do not think so. Even for the staples, for the groceries and the eggs, you can do that now. You do not need AI. You do not need to train a neural network to know when I am going to run out of toothpaste. I brush my teeth twice a day. That is not what AI is bringing value to. The big stuff is actually the thing that I take joy in researching and spending time there. I have been looking at buying a car for six months. Part of it is just because it is fun to do. I like to go through AutoTrader and pick different years and models. It is enjoyable. That is a part of human expression. If you want to suppress that, I do not know what we are left with.

AL: We are getting philosophical here, but I completely agree that there is this whole techno-determinist mindset that thinks that the technology dictates how we will behave in the future and that essentially we are going to let the bots take over. At some point, we just become the humans from Wall-E and we are fat and we just sit in front of screens all day. We are abstracting humanity out of the process. I am pro-humanity. I want to get meaning out of life, and I get that from doing the work that it takes to enjoy things and achieve things.

ES: I do not know that I could be any more optimistic about the impact of AI on society. I do not know if I could be any more of an AI accelerationist, either, but I also think we can pick and choose where we want to apply this. We should apply it to the things that make our lives better and not apply it in ways that strip us of our sense of self. It is weird that people got so upset that I was making the case that I do not think this is going to be something that people adopt broadly. Then, where it does make sense, it makes sense to tack it on as a feature to Amazon or to Walmart and not as a broad product category. I guess one of the reasons people got upset is because they are building startups in the space that presupposes it becomes a big new product category.

AL: Yeah, and at the end of the day, those who seem to be most pro-agentic I have found have a huge vested interest in it. I do not, so I do not care. I would much rather put thoughtful analysis into it and be directionally correct than make a fortune on it.

ES: Let’s talk about Rufus. How is Rufus evolving? Amazon moved sponsored prompts to general availability last month, so those are ads in Rufus. They are finding new ways to monetize it. The last touchpoint we had was ten billion in annual incremental sales as a result of Rufus. What is the latest on Rufus?

AL: There has been some good third-party analysis. Sensor Tower has been doing a bunch of it. They did some analysis around the holiday season which said something like 40% of transactions that occurred had some interaction with Rufus along the way. That is interesting. It is certainly a key touchpoint. When you have that much activity flowing through that sort of behavior, it is natural that you would find some sort of advertising use case. We are going to see some sort of evolution of what that right ad format is. But one of the analyses they did that was so interesting to me was they looked at all the different Rufus searchers. They classified them into ten different consumer segments and how they were using it essentially. Some of those segments were search assistants, research conversationalists, product validators, quick checkers, things of that nature. If you look at all ten segments, the interesting thing is none of them were of the “set it and forget it” variety where they just type something in and tell Rufus to do the work for them. They are actually doing things to take them through the middle part of the funnel. They are trying to build that confidence towards the purchase. They go in with very high intent. They are more likely to convert. To me, the whole use case of the AI tool is about moving through the middle of the funnel, which is probably the most important part of the funnel and the most undervalued part of the funnel. The interesting irony of all of this is that the advertising industry right now is built on traditional brand building advertising where the KPIs are all about things like reach and impressions, and then the bottom of the funnel where it is clicks and conversions and ROAS. We have never had a great language around the middle of the funnel, but that is actually where the heavy lifting happens.

That is the interesting dynamic that I am wrestling with in my head right now. We could be at the dawn of a really new era of advertising where we get better KPIs and understanding of what moves people through the middle of the funnel. If that is true, there is a ton of economic value associated with it. But what is that magical metric? What’s that single unifying metric that makes the whole industry start to unify around? I think that is yet to be determined. That is what I am watching for right now—how does that all manifest? From OpenAI’s perspective, if they have a mega-ad business, I think it is going to be in that middle of the funnel. They can basically do your Yahoo homepage business because they have a lot of traffic. They can have that search, Google-like click-through business, but the problem there is if they are actually giving you that one perfect recommendation for a product, then that comes into conflict with whatever they are advertising. It actually needs to work where people are in that early exploration phase. That is where you introduce the ads and that is in the middle of the funnel. But then what does that look like and what are the KPIs associated with that?

ES: I agree that it should not just be a search model. I wrote a piece recently that search is just a flawed mental model for how these ads should look. I think it should just be display that is based on behavioral history that is collected. My sense is that is probably what they are building. The search model is ill-suited for that engagement profile because of the questions around who you are serving. Am I serving the advertiser or am I serving the user? If there is ever any doubt, then you just poison the well. In terms of guiding people and massaging that transaction in the right direction that leaves the consumer really satisfied, I agree that is the opportunity here. But I do want to revisit that 40% metric because I think you are underselling it a little bit. It is more than interesting; I think that is astonishing. That is again why I say it is not too early, it is too late. Amazon won already because if you have got 40% of Amazon’s transactions during the holidays flowing through Rufus or being influenced by Rufus, then they have dominated agentic commerce. That is an insurmountable advantage.

AL: It is huge. I suspect that you work backwards from the transaction. A transaction is going to happen, and you are more likely when you are in that very high intent mindset to get yourself to the point where you can actually feel confident clicking. I do not mean to suggest necessarily that 40% when you go in is going to be 40% of the transaction. It is probably quite a bit higher during the holiday season because we know intent is the highest. That said, there is no doubt it is going to touch a very high percentage of transactions and be a key part of that path to purchase going forward.

ES: It kind of reminds me when I was getting pitched a lot by AI creative production startups. I was saying Facebook is going to own this. You cannot beat their data, but they already have access to this massive group of advertisers. They just have to flip the switch and turn on that tool. They did, and they announced one million of their advertising clients using their AI-enabled creative generation tools. Then six months later it was two million. If you are a startup competing there, then you are talking about excluding Facebook as a place to deploy these. What are you left with? It goes back to that concentration and that access issue. But going back to the middle of the funnel thing, if you look at that as the opportunity space, first of all, I agree with you that we are talking about something new here. This is not SEM that just has a different front end. This is something fundamentally new, and that is exciting. It is much more exciting than I think a lot of people are giving this credit for. Also, I do not think you actually exclude the notion here that this is display. It could be display, it could be typical behaviorally targeted display if I am guiding you through this transaction using world knowledge and not inserting any sort of influence from advertisers just in the way that I help you figure out what to buy. But I can use advertising as a way to give you recommendations on adjacent things that you might fill up the basket with. That is where I think Rufus is doing it. It is those follow-on queries. It is, okay, I found you the thing that you want to buy and that I think best suits your needs, and therefore everyone wins because you complete the transaction, I am the retail platform, you get the best possible product, and you do not have to worry about whether my recommendation was influenced by an advertiser bidding against it or not. Then, I can make some suggestions that are mediated by bids, by ads, as ads.

AL: That is where incrementality happens. You have related purchases, maybe you upsell people into a more premium version or a larger assortment. All those are places of incrementality. But also, the thing I think about a lot is there is a ton of advertising value in when you find that person who is in your market, they are in the early exploration process, and you introduce your brand. Especially if you are a new brand, a challenger brand, just make them know that your brand exists and what your brand is about. They may not be completing that purchase or intending to move through the path to purchase at that time at all. I think that is an interaction that gets very undervalued because we do not have that performance data. You do not have the click-through at the end of the day and the attribution. If I am an advertiser right now, I am actually looking for all the moments online where that is happening and I do not have to really pay for it. Now, this creates a huge measurement challenge. How do you start to introduce the attribution or the credit or the KPI so that I understand that impact is happening and accruing value to my brand even if it does not translate to a purchase immediately?

ES: Really good teams, really sophisticated teams have mechanisms for doing that, but I think it is also a benefit of these big fully integrated retail platforms because they know. They can say your brand was mentioned as a participant in this category, and maybe not that day and maybe not three months later, but six months later when that person was in market for that thing, they bought your product. They have all sorts of ways of testing that to see if that was actually incremental and if that would have happened anyway or assessing the influence that original brand ad had. They are much better positioned to do that than an independent agent who cannot actually verify against this long sequence of behavior.

AL: And they are going to have to do the research to prove out those effects and then develop the nomenclature around it, develop the metrics, and then teach an entire ecosystem. This happens whenever we have new ad formats. We always had these cottage industries that pop up around them. It is just new, and we have to accept the fact that there is probably value there. If we are going to capture that value as an industry, a lot of this new stuff is going to have to take hold. I think about Pinterest a lot in this context because to me Pinterest is almost a perfect advertising channel in some ways because what you are doing on that site is so—you can draw a direct line from the behavior you are doing to a product purchase or commercial behavior. Yet it has been a very undervalued channel for a long time because it actually excels in the middle of the funnel, but it has always gotten under-attributed. They have actually had to steer more towards performance advertising just to get credit so then they can capture those ad dollars. It is not an easy challenge at all.

ES: Pinterest is a retail discovery platform. Fundamentally, that is what that product is. I am going there to find things to buy. It is almost tragic the degree to which they have just failed to gain traction or failed to grow that ads business because I imagine it is driving a lot of value. The problem is they are subject to the measurement limitations of their clients. That is why their client base are very large advertisers. It is not the long tail, because the long tail cannot convince themselves quantitatively of the value. The large clients can. They have the tools. They have the infrastructure. They have the measurement apparatus. The small clients cannot, and then they cannot directly make the case as a Meta can or as a Google can just through this massive corpus of data that they receive through all of their pixel and the CAPI. That is what I have called small platform syndrome. It is just budding up against that limitation. But Pinterest is just a perfect example of how our industry, because of data and attribution, we have abstracted away the humanity of it. You cannot tell me that seeing an ad in that context is intuitively valuable to you. Figure out what the metrics are that help articulate that if that is what you need to drive investment, but we need to start thinking a little bit bigger picture because the metrics actually are increasingly steering advertisers in the exact wrong direction. Those metrics are increasingly gamed and they are steering us away from a lot of the most valuable marketing channels today.

I want to talk about Rufus. You shared some slides with me from a keynote that you gave, and in one of the slides you had indicated that AI search tops out at roughly 2% of referrals for large retailers. Do you think the steady state of agentic discovery, AI-assisted commerce, is a discovery tool for click-out? If so, how can these platforms capture value outside of their own ads and through subscriptions? I am talking about the independent side. If it is just click-out purchase, if that is the steady state, how do they extract value from that?

AL: I think it is the steady state. I expect those numbers will go up, but it is a new channel. We have added channels in digital over the years, so this becomes a new one. One point is just to show right now it is not a huge referral part of referral to top retailers. It goes anywhere from 0.1% to about 2% for top retailers. About 60% of top retailers’ traffic is direct anyway, so that is not going away to begin with. Then what is interesting is where you see the percentages higher. It is Best Buy, Home Depot, Lowe’s. These are higher, very high consideration categories. Again, when you have really high involvement, you are much more likely to do some of that exploratory research and then end up clicking through. If it was true that higher Gen AI activity was eating into e-commerce traffic, you would see those higher referral retailers declining in traffic, but you are not. You are seeing the exact opposite. You actually see a positive correlation between the share of Gen AI referral and traffic growth to retailers. There is probably a better argument that Gen AI and having new forms of search is an incremental behavior and something that actually will add to e-commerce growth if anything.

ES: Another slide that I found to be really fascinating was just the search share by searcher category. You had the early adopters at 41.4% as the most recent data point in the series. But that was flat for three months running. Then you had the all-searchers, which was also flat but it seemed to be on a sort of upward trend, but it was flat for the last two—essentially flat for the last three. Is this the steady state? Because I would have imagined first of all that the early adopters number would have been higher if this truly was this new consumer behavior that was going to be established as a norm. For this to have flatlined already, again, this just reinforces my point. This is not too early. This is too late. Amazon owns it. Walmart owns it with Sparky. All these retail platforms, 2% of their clicks are from out, who cares? They will build their own agents. They are going to integrate directly. They are going to get you to the transaction and service that messy middle. Talk to me about that flattening.

AL: This is looking at LLM search, the share of search. It goes from about 2% to 6% overall of searches, which is non-trivial in a massive market like search, but it is not overtaking the market. There is a subset that we look at here of early adopters, basically people who were already searching on LLMs back in early 2024. In that, you see a sharp uptick from about 8% of searches. It goes up pretty quickly into the 30% and 40% range, but then as you mentioned, it starts to level out at around 40% in late 2025. There is another cohort analysis to see if this is just an early adopter phenomenon or are newer cohorts doing the same thing. The newer cohorts are doing the same thing. It is not as sharp of an uptick to like 40%, but it is still fairly pronounced. That suggests that activity should be moving in this direction for search overall. The interesting thing is though it did flatline and that flatlining has continued. What a lot of folks do not understand is that if you were to look at ChatGPT data going back to around July or August of last year, it has not budged in terms of weekly active users. It has not budged in terms of total sessions. It is flat in the US at least, and maybe incrementally growing globally. This is no longer a hockey stick phenomenon. It started out that way, but then it hits a natural limit. I have seen this growth curve before where everyone assumes that the exponential curve goes up and to the right forever. It was the early days of Twitter. Twitter had this huge hockey stick and then all of a sudden it came to a screeching halt and it just went flat for a while. Then it started to build in a very steady incline over a longer period of time. I expect something similar will happen here with ChatGPT, but we just need to acknowledge the fact that we are no longer in this exponential growth curve.

ES: I have a lot to say about ChatGPT’s self-reported WAU numbers. First of all, WAU is a non-standard metric. People say it makes more sense to better reflect the usage, but that is not the point of a metric. A point of a metric is to be comparable. If they are using that internally because they think that better captures behavior patterns, fine, but that is not how they are using it. They are using it as an external reporting metric, and it is non-standard. What is the point of adopting a non-standard metric? To obfuscate the actual underlying dynamics, which is I think why they have adopted it. They said 900 million WAU when they announced the fundraise in February. You would have expected them to have updated that when it hit a billion. Obviously, the growth is slowing down. They had been engaging in this steady drumbeat of metrics reveals, and now they are not. A billion is a pretty big milestone. Obviously, the growth is slowing down.

AL: Third-party data speaks to this. It is pretty similar trends, whether you are looking at MAU, WAU, or DAU. They are literally in parallel. Some of them might show very minor increases. Depending if you are looking at weekly, there is even some suggestion that you are seeing some declines in some of those, very incremental, but they seem to align with post-Super Bowl and some of Anthropic catching, gaining steam. No matter how you cut it, it is not a strong growth story anymore. I think we are seeing that both from the external reporting and what third-party data shows.

ES: That is why you would use WAU because it is more sensitive to big swings after a model release. Even if it is backward-looking WAU, if you are averaging out these very large numbers because there was a model release, then that is why you would use it. You could time it to show higher numbers. There is nothing inherently wrong with WAU as a metric, especially an internal reporting metric. If that is more aligned with usage patterns, you should use that, but you cannot compare it with anything else. It might be used to game the releases. That was my issue with it.

AL: Just lacking the comparability I think is a key point, although I could also argue maybe it makes sense because long term this is behavior that should very much be a daily habit for people. Long term, you have to be reporting in terms of daily active users. In the interim, as you are still building that user base, maybe it is fair to say weekly is a decent metric for now. The bigger factor is probably just lack of comparability.

ES: Let’s get to performance TV. You wrote a great piece recently, “Will CMOs finally wake up to Retail Media in 2026?” You made the case that performance TV and in-store retail media are the two significant new opportunities in 2026. Explain what both of those are.

AL: Performance TV is—a lot of times people think it is shoppable TV, whether I have a QR code or an add-to-cart or I have to click on my remote to buy something right away. I think that is always going to be a very tiny behavior. When I say performance TV, I mean TV ads, specifically CTV ads, underpinned by closed-loop targeting and attribution. Think about primarily what Amazon is building here with Prime Video and also now that Amazon is plugged into all the other ad-supported CTV networks. That is performance TV and that is a big market opportunity and will really change the dynamics of TV advertising because you still have the benefits of TV ads which have always been about brand building and reaching the right audiences, but now you can actually see if they are driving sales. There are all sorts of interesting dynamics that come into play when you force-fit performance metrics into this equation.

In-store retail media is just we are seeing a rise now of in-store digital ads into physical store environments. CVS has moved in this direction, a lot of regional grocers and now Kroger and Albertsons. This is a big thing in every other market but the US; we are very late to the game here, but it has a ton of potential. The reason why these are increasingly in the crosshairs of the CMO is a lot of CMOs have kind of pushed retail media to the side as just this e-commerce performance channel. It is not what they have traditionally cared about. They are brand builders, creatives, brand marketers at the end of the day. My argument is the CMO of the future and advertising of the future, you have got to be able to do both. You have to understand brand building and performance at the same time. These are almost the two perfect venues for the CMO to care about because they are scaled media that can do unbelievable brand building with great creative. I think in-store digital is a creative opportunity unlike we have ever seen because you can have executions happening in three dimensions across the store in really unique and interesting ways. Also they are going to drive performance. How do you blend these disciplines and be able to think about it and have reporting that really captures the big picture of reaching audiences at scale, how valuable are these audiences, who are these audiences, and then within the short-term view of the world, is it leading to incremental sales? It cannot be one or the other or I think you miss a lot of the value.

ES: Just for my own personal edification, what formats do these take? Is this genuinely a digital billboard in a grocery store?

AL: Think of TV screens. You can have them at the checkout aisle. Sometimes you have these six-foot tall digital screens at the entrance. You will see TV screens at the pharmacy counter or the deli counter. You can have some around the periphery of the store. I think the end cap at the end of the aisle is probably the most prominent place, maybe the highest impact format in the store. It is just increased digitization. You have to execute in a way that does not detract from the customer experience. I am also very concerned about introducing what I call the physical store equivalent of popup ads. You do not want to have ads shouting at you, it does not want to be promotion. But if you have high-quality creative dotting the store in interesting ways, it can actually elevate the store experience, modernize it, and introduce people to brands, new products, and drive sales.

ES: And this is bought through the existing Retail Media Network infrastructure? Target’s Roundel and Albertsons’ and Kroger’s—you are just buying that through their existing interface?

AL: Primarily how it is going to happen in the future, but there is also dollars driving through programmatic right now as well. Think of it as like an extension of programmatic digital out of home. That is where most of them come today. It probably makes more sense to plan and execute them like we used to do for linear TV buys because you can kind of plan by daypart, by geography and all these different factors that are much more of a scaled ad buy, and yet there is such an inclination to just buy everything programmatically. I just think you want to take a lot of care of what the creative is and how you execute it in-store. You cannot just throw any ad up there and expect that it is going to work. You have to make it native to the environment. This is how every new medium emerges where you just import ads from the last thing into the new thing and it does not really work that well at first, and then as the ecosystem builds around it, you get better creative and you start to have it realize its potential.

ES: Where is Shopify sitting in all of this? Performance CTV, agentic commerce—what is Shopify’s role here? They have been kind of slow to really embrace advertising, but they just launched their network.

AL: It is an audience network at the end of the day. They have a huge swath of potential advertisers with all these DTC brands. In that sense, they have a lot of potential and could be a front door to a lot of advertising. It is a platform I am scratching my head at a lot of the times. I am not sure that they have fully executed to their potential with respect to CTV. DTC brands are looking to diversify beyond Facebook and Google today, no doubt about it. CTV is probably going to be the next frontier, but it is not going to be easy. They have to get access to good inventory, inventory that is going to perform. I worry a lot about CTV ad fraud. If you are buying ads against fraudulent inventory, then it is not going to perform from an incrementality perspective and the investment will not be there. The other thing is CPMs are high for CTV, and so inherently that kills ROAS. ROAS is not a good metric, but CPM high, ROAS low, and so the typical direct-to-consumer advertiser will just optimize out of that. What they end up doing is they buy retargeting ads for TV, and that is the only way to make ROAS pencil out. I worry that is what is going to happen. It is not really incremental and it may not drive growth for their business over time. Can Shopify play a role there? Absolutely.

ES: When did we stop being marketers? You cannot tell me that seeing an ad in that context is intuitively valuable to you. Figure out what the metrics are that help articulate that if that is what you need to drive investment, but we need to start thinking a little bit bigger picture because the metrics actually are increasingly steering advertisers in the exact wrong direction. Those metrics are increasingly gamed and they are steering us away from a lot of the most valuable marketing channels today.

AL: Just lack of comparability I think is a key point. Another perfect example of how our industry, because of data and attribution, we have abstracted away the humanity of it. Sometimes I just have to remind people: when did we stop being marketers? You cannot tell me that seeing an ad in that context is intuitively valuable to you. Figure out what the metrics are that help articulate that if that is what you need to drive investment, but we need to start thinking a little bit bigger picture because the metrics actually are increasingly steering advertisers in the exact wrong direction. Those metrics are increasingly gamed and they are steering us away from a lot of the most valuable marketing channels today.

ES: Andrew, I appreciate you coming back on the podcast. How can people find you on the internet? How can they consume your content?

AL: I am pretty active on LinkedIn, not on Twitter. You can check out my Substack: mediaadsandcommerce.substack.com.

ES: Thank you, Andrew.

AL: Thanks, Eric.