Walmart’s MCP server is generally misunderstood; it doesn’t provide carte blanche catalog / transaction access and, if anything, is further support for the idea that Agentic Commerce is a Mirage. First, Walmart is allowing third-party agents to *connect to Sparky*, its own on-platform agent. This is an agent-to-agent gateway. Walmart retains the consumer data and first-party consumer relationship. And Walmart exercises fundamental control over product discovery and recommendations if Sparky is the agent interface. Second, autonomous checkout is permitted only under Walmart's conditions. Consumers must authorize the agent through their Walmart account and establish permissions and spending limits, which Walmart enforces. To use the Gateway, the consumer must connect their agent to their Walmart account by authenticating directly with Walmart. So Walmart continues to accumulate user-level purchase data for use in a) the product recommendations returned by Sparky and b) Walmart Connect. Third, third-party agents are restricted from using Walmart’s data for any purpose that could possibly intermediate Walmart. From its terms: An Agent or its Deployer must not: (a) conduct systematic or bulk scraping, collection, storage, or other use of Walmart Proprietary Data; (b) retain Walmart Proprietary Data beyond what the Permitted Purpose reasonably requires; (c) use Walmart Proprietary Data to generate content, target advertising, create synthetic data, train models, or produce commercial analytics, reporting, or data-licensing products; or (d) license, resell, syndicate, or redistribute Walmart Proprietary Data to any third party. Walmaet’s agent gateway relegates external agents to distribution interfaces, and it preserves Walmart's first-party consumer relationship and advertising economics. This isn’t a strategic embrace of third-party agentic commerce. I was going to publish about this last week but got sidetracked by $META earnings. Post going live tomorrow.

Meta Q3 2026 earnings: interpreting early Muse metrics I don’t think I’ve ever been tagged in as many X posts as I was yesterday when Mark Zuckerberg stated on the Meta earnings call that shopping is the most common use case for Muse, and that the average Muse transaction size is $200. At first blush, this seems to undermine my 'Agentic commerce is a mirage' thesis. But I believe that deconstructing Zuckerberg's statements reveals a different interpretation. I've made my Q3 2026 earnings analysis for Meta free; no paywall. mdm.co/meta-q3-2026-earnings…

The catalyst for this tweet was a conversation I had last week with a friend who is the CMO of an F100 company. I didn't want to share any details that could leak identifiable information, but he gave me permission this morning: Like every major brand / retailer, he's considering just how permissively and fulsomely to integrate with third-party agents for agentic discovery, and how much functionality should be made available outside of his owned environment. This is a salient and critical question! Agentic discovery could become an important part of the funnel. But in some implementations, it could also be corrosive. If agents are merely recommending his products / company among a relevant set and delivering incrementally net-new users, integrating / cooperating is essentially costless. But that very obviously and predictably won't be the case. Some users would have been reachable with ads. Will they behave differently when exposed to an in-agent recommendation versus an ad? Will their AOV change as a cohort? What about their product journey? Will they be as reachable after their first checkout? Of course, we don't know, so we have to test this. How do you design an experiment that captures this? How do you construct the exposure and holdout groups? How long should this experiment be run? How do we accommodate the upcoming holiday season, the paid promotion of new agents that influences their usage in ways that are unlikely to persist, and differences across categories and geographies? These are genuine, serious challenges, and his board has been extraordinarily accommodating in allowing him to put together a thoughtful, comprehensive "agentic" strategy. But it occurred to me in speaking with him that many CMOs are absolutely about to be replaced because they either aren't given that time / space to experiment (and the CEO has a friend who is confident that he can 3X the business with his "agentic playbook") or because they say yes to everything without any clear vision for how the business will be impacted.

30% agent share with a 3% take rate by 2030 is extraordinarily aggressive. Repeat this arithmetic with 5-10% agent share and the entire endeavor looks futile / indefensible. Unless, of course, agents don't implement a take rate, which, for many reasons, can't scale to a stable equilibrium. Rather, if agents implement ad partnerships like the one Amazon has already struck with OpenAI, allowing the largest retail platforms to integrate directly into off-platform agentic purchasing workflows while retaining control over user relationships and their advertising revenue, then everyone's incentives align, and DAU (not eCommerce share or take rate) is the only determinant of revenue.

Using natural language user profiles in RecSys "Spotify recently published a paper, Textual User Taste: Natural-Language User Context for Foundation-Model Recommender System at Scale (Fazelnia et al., 2026), that touches on a fascinating and timely topic of broader research at the intersection of RecSys and language modeling: whether the dense user embeddings used in candidate retrieval and ranking can be complemented with, or perhaps even replaced by, a natural language description of a user’s preferences." mdm.co/using-natural-languag…

Agentic commerce is not a mirage for travel "But the travel use case of agentic commerce presenting an obvious and fairly unobjectionable alignment between retailer, agent, and consumer also hides a sleight-of-hand that some of the proponents of agentic commerce would prefer to go unnoticed: that this alignment contrasts substantially with the dynamics of retail eCommerce, and so championing agentic travel booking — which is convenient, and which does provide unambiguous value to hotel chains, airlines, and consumers alike — doesn’t serve as a blanket endorsement of the agentic commerce enterprise." mdm.co/agentic-commerce-is-n…

Anecdotes and personal experiences can be useful tools for understanding the utility of products for specific use cases, but I think it’s a mistake to use such a narrowly-scoped instrument to try to explain the business models of trillion-dollar companies — and, more contemporaneously, why they’ll be utterly disrupted by AI. Certainly this is true for a company like Google $GOOGL, as the mechanics of digital advertising are more complex than they might appear when looking at the results for any given query. The notion that Search merely, exclusively monetizes the results that would have otherwise been exposed organically, almost like a protection racket, is flawed. Search ads offer keyword adjacency / expansion tools, and shopping journeys start long before a consumer knows exactly which product they want to buy. Further, marketing teams can test their spend for revenue incrementality; doing so with Google Search is straightforward. Certainly some Search ad spend is wasted but that’s true for every ad channel. “Google Search dies because of Muse” ignores the incremental product discovery value of search advertising, the fact that Google has launched its own AI products that currently reach in aggregate billions of people per month (AI Overviews: 2.5BN, AI Mode: 1BN), the fact that Google owns the most popular browser, the fact that Google operates a mobile platform, and the fact that Google operates a suite of consumer products that are all deeply integrated. And in the context of this specific argument, it also ignores the fact that Muse will need to generate meaningful revenue at scale and will almost certainly do so with advertising. In some ways, Google Search has already died (and been resurrected in a different form). This is the premise of my Google’s Gambit series, which I started publishing long before Muse. The company is adapting in real time to new consumer interaction models. If consumers find agents useful for product discovery, there is no reason to believe that Google *can’t* accommodate that. But to reduce a ~$250BN annualized business to a consumer- and advertiser-hostile protection racket, and then assume that Google cannot adapt its distribution and monetization machinery to an agentic consumer workflow (which it already partially has, through Universal Commerce Protocol, Universal Cart, agentic checkout, etc.), is, in my view, overly simplistic and naïve.

There are two fundamental problems with this set of scenarios related to Muse’s revenue contributions to Meta by 2030: - A and B (worst and base), where Muse is 0.5% and 2.3% of revenue, respectively, are economically irrelevant to Meta and mean the company should shut the product down absent some second-order effects; - the revenue assumptions in the only scenario that is commercially meaningful to Meta (C, best case, 7.8%) presume a uniformity of direction in take rate, interchange fees, and GMV share with scale that is improbable, and they ignore the opportunity cost of making Muse free and monetizing it with ads. The first point is obvious: even with the lofty assumption of 15% eComm addressability by third-party agents (!), and 40% (!!) of Muse GMV passing through partners willing to accept a 4% transaction take rate (!!!), Muse accounts for 2.3% of Meta’s revenue by 2030 in scenario B. Immaterial. Who cares. The second point is more nuanced. In scenario C, 60% of Muse GMV (!!!!) runs through partners willing to pay a 5% take rate (!!!!!). That doesn’t follow from scenario B; surely the largest retail platform holdouts would negotiate a lower transaction fee. That is to say: the highest fees are likely realizable by the smaller players first, and the average take rate probably decreases as total eComm penetration increases. But more importantly: why would Meta prefer a uniform transaction fee to the auction mechanism it has spent two decades perfecting? Merchants have heterogeneous margins, conversion probabilities, customer values, product lifecycles, and willingness to pay for access to a consumer. An auction allows them to express that willingness to pay; a 5% transaction fee reduces all of that heterogeneity into a fixed percentage. This is the opportunity cost missing from the model. Meta already generates enormous advertising revenue per user by auctioning access to consumer attention and intent, without needing to intermediate the transaction or secure a commercial agreement covering the resulting GMV. The model implies ~76MM Muse users in scenario C. Transaction fees + interchange generate roughly $396 per user per year. For context, while Meta no longer reports geographic ARPU, its earnings disclosures imply current US/Canada Family ARPP somewhere around ~$340/year: UCAN revenue is running at roughly $95–100BN annually against an estimated ~275–300MM daily users. And UCAN revenue grew 32% YoY last quarter. So even in scenario C, Muse needs each US user to push $12,784 of annual GMV through the product (>$1,000 / month), 60% of that GMV to flow through fee-paying partners, those partners to accept a 5% rake on average, and Meta to capture 10bps of interchange just to generate ~$396 of commerce revenue per user in 2030. Whereas Meta’s existing advertising machine, today, is generating something like ~$340 per UCAN daily user. So Meta could just show those users ads. And the latter mechanism allows merchants to bid according to the actual value of each impression rather than Meta attempting to impose a uniform fee. So the model ultimately demonstrates just how heroic the assumptions have to become before transaction monetization can approach competition with Meta’s existing advertising machine.

Apple blocks alternative IDs in iOS 27 "Another reason to be dubious about the long-term durability of alternative identifiers is simply that Apple likely views them as inimical to its privacy goals. I wrote about the interminable game of whack-a-mole that Apple plays with advertising platforms and intermediaries alike in April 2023, when Apple closed a loophole that allowed third-party cookies to be presented as if they were first-party cookies through IP-address cloaking. Ultimately, Apple has the resources and the patience to outlast any practices it simply doesn’t want taking place on its platforms. In that sense, every workaround or loophole always has an expiration date for the Apple ecosystem." mobiledevmemo.com/apple-bloc…

There are two fundamental misapprehensions behind the belief that independent, general-purpose agents will broadly disrupt the eCommerce landscape: 1. That merchants would prefer organic agent recommendations and transaction fees to auction-based advertising because branded search advertising is a zero-value "tax" or "toll." This isn't the case; branded search advertising has been observed to deliver incremental value. In “Competition and Crowd-Out for Brand Keywords in Sponsored Search," Simonov et al. found that own-brand ads on Bing increased traffic even without competing ads, while competitors captured 18–42% of clicks when brands facing competition did not advertise. Even in the specific advertising context of Amazon, the available research supports incrementality. In “Data-driven budget allocation of retail media by ad product, funnel metric, and brand size,” Koen Pauwels (friend of the podcast!) and coauthors estimate positive contributions from Sponsored Products to awareness, consideration, and revenue across more than 122,000 brand/category combinations. Characterizing branded search advertising as a tax doesn't really address the core economic issue, which is whether it can deliver incremental value. Obviously, like all advertising, branded search can be wasteful. But the research supports the idea that branded search can generate incremental returns. So the question is why a retailer would prefer to integrate with an agent and pay a fixed, percentage fee on each transaction rather than bid for distribution according to what reaching a particular customer is worth to its business. I don't think it would. Even if a retailer can accommodate an agent's fixed transaction fee (OpenAI's was 4% on Instant Checkout), wouldn't it sometimes prefer to pay more for the opportunity to reach a relevant customer? A bid allows a retailer to express its own derived willingness to pay. So if agents become important sources of demand, why wouldn't merchants want to compete for exposure on those surfaces, too? 2. That consumers will demand a single general-purpose agent that handles all of their needs, crowding out platform-owned agents while also retaining all of the functionality that the platforms provide. Suppose the mono-agent succeeds. Amazon loses its consumer traffic and becomes middleware that fulfills agentic purchases. Why assume that Amazon continues operating at its current scale if the agent has subsumed the customer relationship and much of the revenue opportunity ($69BN in advertising revenue in 2025)? Fulfillment, customer service, returns, fraud prevention: who manages these if the revenue opportunity from the consumer relationship is undercut? The mono-agent thesis assumes that an agent can dismantle the platform's business model yet benefit from the infrastructure upon which those businesses were built. But the more totally any agent displaces the platform, the more of the platform it needs to replicate. I think the most concrete and immediate example is reviews. If people stop buying on retail platforms, where are the reviews submitted? In the agent? That defeats the purpose. But if product reviews evaporate, how does the agent rank products? This is obviously why Amazon has been reluctant to integrate with third-party agents. And no, Walmart's cooperation with various agents isn't instructive here; the two companies operate against fundamentally different incentives. Walmart has a much smaller eCommerce footprint and advertising business: it has more demand to acquire and less existing demand to defend. And it also operates its own agent.

Transaction fees simply won't scale in an agentic commerce use case, for reasons that exceed just the GMV requirement. I made this case in Affiliate links, personalized ads, and chatbot revenue optimization, which I published after OpenAI introduced Instant Checkout but before it launched advertising. I argued that OpenAI would ultimately relent and adopt advertising simply because it is economically superior to running an affiliate network, which their proposal would have become. Affiliate fees / transaction fees may play some small role in agentic shopping, but they can't possibly be the base case business model, for the reasons I lay out in that piece. mobiledevmemo.com/affiliate-…

On this week's episode of the @MobileDevMemo podcast, I speak to OpenAI's Asad Awan about the company's ambitions with advertising in ChatGPT. mobiledevmemo.com/podcast-op…

How can ultra-long sequences of events be evaluated efficiently for content recommendation? ByteDance released a paper earlier this month, accepted for presentation at RecSys '26, that introduces SequenceO1, a recommendation model architecture that incorporates ultra-long sequences of up to 100,000 user events. The authors accomplish this through what they call Sketch Attention. Rather than constructing a single user vector from an event sequence, Sketch Attention compresses a very long stream of event embeddings (up to 100k) into 1,024 learned representations of user activity, or a "sketch." The model uses 1,024 learned prototype vectors, trained end-to-end with the ranking model, to generate user-specific sketches through cross-attention against each user's event history. Notably, the attention weights are normalized across prototypes for each event, so every event distributes its contribution across the summary slots. The sketch can be re-used across candidates contemporaneously, so ranking is request-centric rather than candidate-centric. This is similar conceptually to what Meta discusses in its Adaptive Ranking Model blog post from a few months ago: instead of processing full attention across the candidate-conditioned user embeddings, it runs a lighter-weight comparison on the pre-processed sketch to produce the candidate scores. SequenceO1 then utilizes the raw embeddings of the user's most recent 10,000 events in a separate branch that produces a candidate-conditioned user embedding, and the candidate-conditioned outputs of both branches are fused in MixFormer, the downstream ranking backbone. Both branches (ultra-long and 10k) are processed with Stacked Target-to-History Cross Attention (STCA). The sketch is cached with a 1-hour TTL for inference and a 3-hour TTL for training. The authors observed statistically significant improvements to engagement metrics on Douyin and Douyin Lite across the board, including +2.33% for video finishes and -6.98% for dislikes on Douyin. And SequenceO1's ultra-long history branch was estimated to require 50x fewer training-side FLOPs and 64x fewer inference-side FLOPs than computing the full 100k-event sequence directly with STCA. Paper linked below.

What if everything turns out alright? Among the many disastrous consequences predicted of artificial intelligence is a jobs famine: widespread worker displacement that devastates the economy and brings about a second Great Depression. But that hasn’t happened. In fact, the US economy is doing well and, with a 4.1% unemployment rate, operating roughly at full employment, despite other headwinds like persistent inflation. And while whether humanity is “early” in the adoption phase of artificial intelligence is debatable — I argue that, certainly, from an exposure standpoint, AI is a familiar technology at this point — the notion that AI has displaced large swaths of the workforce is not supported in the current data. Given the buoyancy of the economy, should we reduce the odds of that happening? We can’t definitively say that it won’t, of course, but we can use the data at our disposal now — in September 2026, nearly four years after ChatGPT’s launch — to update our likelihoods. And one likelihood that I believe is underestimated, sitting somewhere between The Prosperous Society and total economic Armageddon, is that everything turns out alright. mobiledevmemo.com/what-if-ev…