Who Owns the Customer When AI Shopping Agents Do the Deciding
By M. Mahmood | Strategist & Consultant | mmmahmood.com
TL;DR / Summary:
When Walmart quietly tested AI-driven checkout across roughly 200,000 products, it discovered that purchases completed inside ChatGPT converted three times worse than purchases completed on its own website, a result that reframes the entire AI shopping narrative most executives have been sold over the past year. The real contest unfolding beneath the surface isn't a simple story about chatbots replacing storefronts, it's three interlocking battles happening simultaneously:
- Whether your product data is trustworthy enough for an AI agent to act on.
- Whether you understand that SEO, AEO, and GEO are fundamentally different games with different rules and different time horizons.
- Whether your brand or the platform running the agent ends up owning the customer relationship once that agent inserts itself into the buying decision.
This piece walks through what actually happened, why the conventional wisdom got the sequence backwards, and lays out a 90-to-180-day plan for making sure your organization, not a platform intermediary, still owns the customer once the current wave of hype settles into something more permanent.
The moment that exposed the whole industry's blind spot
The customer called it a bug. It wasn't.
Rebecca had asked ChatGPT to track down a birthday gift for her sister, a specific ceramic mug with a hairline blue glaze she remembered seeing months earlier and had never quite forgotten. The assistant found it almost instantly, confirmed the price without hesitation, and offered to complete the purchase then and there, without ever asking her to open a browser tab. She agreed, because the experience felt effortless in exactly the way these tools are designed to feel. A few days later, the wrong size arrived at her door. She abandoned the conversation entirely, opened Safari instead, and bought the correct item directly from the retailer's own site, the way she would have done a year earlier before any of this technology existed.
That small, almost forgettable moment, replicated across roughly 200,000 products in a controlled test, is the real story behind why Walmart pulled the plug on its ChatGPT checkout integration in March 2026. In-chat purchases converted at only one-third the rate of simple click-throughs to Walmart.com, a gap wide enough that Walmart's own product chief was willing to go on record and call the entire experience "unsatisfying." That kind of candor from a launch partner, rather than a critic, is usually the clearest signal available that a technology has been oversold relative to where it actually stands.
I have spent two decades building and evaluating enterprise AI and data platforms across telecommunications and portfolios collectively worth well over a billion dollars, and having sat through more vendor pitches on this exact topic than I care to count, my read on the Walmart number was never that it represented a UX failure to be patched with a better interface. It read instead as an early, unusually clean data point in a much larger structural fight that every retail and B2B organization is now being pulled into, whether its leadership has noticed yet or not, a fight over who actually owns the moment a customer decides to buy, and whether the AI standing between the brand and that decision is functioning as a bridge toward the sale or quietly becoming a toll booth that captures the value along the way.
None of what follows is really a story about chatbots, however much the headlines want it to be. It is three separate but tightly interconnected battles unfolding at once. AI shopping agents are rewriting how customers discover and evaluate products before they ever reach a checkout page. A parallel and less understood war between SEO, AEO, and GEO is rewriting how brands get found by anything at all, human or machine. And beneath both of those sits a quieter, more consequential fight over who retains the actual customer relationship once an intelligent intermediary inserts itself into the transaction. Leaders who treat these as three unrelated initiatives, assigned to three unrelated teams with three unrelated budgets, will find themselves losing ground on all three fronts simultaneously, often without fully understanding why their numbers are slipping.
The decision every retail and B2B leader is quietly avoiding!
Stated plainly, the choice in front of you is this: should your organization direct its scarce AI budget toward being discovered by AI shopping agents, toward converting the customer once discovery has already happened, or toward both simultaneously, and if both, in what sequence and under whose ownership? This is not a rhetorical question meant to set up a neat answer later in the piece. It is the actual resource-allocation dilemma that will determine whether 2026 becomes the year your organization built a durable advantage or the year it burned a budget cycle chasing visibility metrics that never converted into revenue.
Sequence this incorrectly and you will spend the better part of the year chasing AI visibility scores while a sharper competitor, one with cleaner structured product data and a clearer view of where the actual value sits, quietly walks away with the transaction, the customer record, and every repeat purchase that follows from it. Sequence it correctly, however, and what currently looks like a frightening architectural disruption becomes a genuine, compounding acquisition advantage, secured well before most of your category even understands what changed beneath them.
What actually happened in the agentic commerce experiment, and why the hype ran ahead of the evidence
In September 2025, OpenAI launched Instant Checkout, a feature that let ChatGPT users complete purchases from Etsy sellers without ever leaving the conversation, built on the Agentic Commerce Protocol that OpenAI co-developed with Stripe. The pitch that accompanied the launch was seductive precisely because it was simple: over a million Shopify merchants were said to be coming soon, and the implicit promise, repeated across breathless commentary at the time, was that the traditional storefront was already becoming a relic of the pre-agentic era.
By March 2026, the initiative was quietly winding down. Forrester analyst Emily Pfeiffer found that only about 30 Shopify merchants had ever actually gone live on the platform, a rounding error against the million once promised in the launch narrative. OpenAI itself later told Digital Commerce 360 that the phrase "complete purchases inside ChatGPT" had always really meant discoverability rather than transaction completion, a fairly significant reframing delivered well after the initial announcement had already shaped a year of industry expectations, and confirmed that Instant Checkout was simply moving into a different product surface called Apps.
Walmart's own numbers explain precisely why the retreat happened. Having tested roughly 200,000 products through Instant Checkout beginning in November 2025, EVP Daniel Danker reported that in-chat purchases converted at just one-third the rate of click-throughs that landed on Walmart.com. Walmart's own assistant, Sparky, performed considerably better, converting at around 70 percent of the website's rate, which suggests the gap is closing but has by no means disappeared.
Here is the part that no vendor whitepaper circulating this year will ever admit outright. The problem was never that shoppers fundamentally distrust artificial intelligence as a concept. The actual problem, once you look past the UX-focused postmortems, was that most retailers' underlying product data simply is not structured cleanly enough for an agent to confidently resolve a specific SKU, a current price, real-time inventory, and an accurate shipping promise, all without guessing at some point in the process. When an agent guesses, one of exactly two outcomes follows: it pauses to ask a clarifying question, introducing friction that a human shopper on a familiar website would never have encountered, or it simply picks wrong, which kills the cart outright. Analysts have started calling this the Product Truth Gap, and beneath its catchy name it is fundamentally a data infrastructure problem dressed up in UX language, one that a rigorous structured data playbook can genuinely fix, provided the organization treats it as an engineering priority rather than a marketing afterthought.
Notably, Google and Microsoft did not retreat the way OpenAI did. Google launched its Universal Commerce Protocol at the NRF conference in January 2026, and Microsoft's Copilot Checkout went live that same month, suggesting the two companies view the current friction as a temporary infrastructure gap rather than a fundamental flaw in the underlying premise. And the long-term forecasts have not meaningfully softened either: McKinsey continues to project that agentic commerce could orchestrate somewhere between 900 billion and one trillion dollars in US B2C retail revenue by 2030, with global projections extending as high as three to five trillion when you account for markets outside the United States.
The clearest loser in this early phase of the story is any retailer or B2B seller that continues to treat its product catalog primarily as a marketing asset rather than as a machine-readable contract that governing systems can actually parse. If your organization's SKUs, prices, product variants, shipping windows, and return policies still live mostly in loosely structured prose scattered across a webpage rather than in properly maintained schema and live, low-latency APIs, no amount of strategic partnership with an AI platform is going to rescue your conversion rate once agentic discovery becomes the norm rather than the novelty.
The second battle: why SEO, AEO, and GEO are fighting three genuinely different wars
While commerce teams were consumed by the checkout debate, marketing organizations walked into an equally consequential fight that most of them still have not carved out a distinct budget line for, largely because the distinctions between these three disciplines are genuinely subtle and easy to collapse into a single, oversimplified "AI SEO" category. Traditional SEO earns you a ranked position within classic search results. AEO, or Answer Engine Optimization, earns you the quoted and directly linked answer that appears inside featured snippets, People Also Ask boxes, and AI Overviews. GEO, or Generative Engine Optimization, earns your brand something more elusive still, a simple mention embedded inside a generative answer produced by ChatGPT, Perplexity, or Gemini, frequently without any link back to your site at all, a nuance that retail-focused SEO guides are only now beginning to articulate with the precision that practitioners actually need.
These three disciplines reward fundamentally different behaviors, and they operate on fundamentally different timelines, which is precisely why treating them as a single workstream produces mediocre results across all three simultaneously. Traditional SEO rewards accumulated topical authority and earned backlinks, and meaningful movement typically plays out over a horizon measured in months. AEO rewards concise, cleanly structured, factual content that a machine can lift and repurpose almost verbatim, and it can produce visible movement within a matter of weeks once the underlying content and markup are in place. GEO rewards something considerably harder to manufacture on demand: a consistent pattern of entity signals distributed across the open web, built from reviews, comparison articles, and forum discussions that a large language model has already absorbed into its training and retrieval corpus long before your organization ever set out to influence it deliberately. You cannot purchase your way into strong GEO performance with a quick content sprint the way you sometimes still can with paid link placement in traditional SEO. The model has to already regard your brand as a trustworthy, category-relevant entity, and that kind of trust accumulates slowly, through consistency rather than campaigns.
This dynamic is precisely why AI shopping agents surfaced Etsy and Shopify merchants with clean, well-structured catalogs first, ahead of larger but less technically prepared competitors, and it is also why Walmart's own click-out traffic continued to outperform anything happening natively inside ChatGPT throughout the experiment. The website remained the trusted, GEO-anchored destination for the actual transaction even as raw product discovery began migrating steadily into conversational interfaces.
I have personally watched senior AI leaders inside large enterprises make precisely the same category error with GEO that an earlier generation made with the early days of SEO, namely assuming that a single team could own the entire discipline end to end with a tidy checklist and a quarterly review. GEO, properly understood, sits much closer to reputation management with a technical layer bolted on top of it than it does to conventional digital marketing. You cannot game your way into becoming the brand that ChatGPT spontaneously recommends for "best trail running shoes under 150 dollars." That kind of recommendation is earned through the same third-party credibility signals that brand equity researchers at the major strategy consultancies have written about for decades, only now those signals are being parsed, weighted, and scored by a language model rather than by a human analyst working from a survey.
The third battle: who actually owns the customer once an agent inserts itself into the decision
This is the fight that most boards are simply not having yet, and it is, in my assessment, the single dynamic most likely to quietly determine long-term enterprise value across the next several years. When a shopper asks an agent to "find running shoes for trail use, under 150 dollars, delivered by Friday," it is the agent, not the brand ultimately selected, that captures and retains the first-party relationship with that underlying intent and those specific comparison criteria, according to eMarketer's research on agentic shopping behavior. The brand only receives the transaction itself, and only on the condition that its underlying data is clean enough for the agent to select it with confidence in the first place.
The same IAB research cited in that eMarketer analysis found that 38 percent of consumers already use AI in some capacity while shopping, with roughly 80 percent expecting to rely on it even more going forward, though predominantly for comparison and decision support rather than fully autonomous purchasing at this stage. What that data actually tells executives, once you set aside the more sensational framing common in trade press coverage, is that the near-term danger is not that agents will suddenly replace your storefront overnight. The real and present danger is that these agents are already inserting themselves into the decision-making process itself, and every single decision reached without your brand physically present in that conversation quietly transfers a small increment of customer ownership toward whichever platform happens to be running the agent in question.
Here is the point that virtually no vendor sales deck circulating today will ever put directly in writing, because doing so would undercut the entire premise of the partnership being pitched. Platform-run shopping agents are not neutral arbiters standing dispassionately between buyer and seller. They are built and operated by companies that also generate substantial revenue from advertising and that frequently run their own competing retail products alongside the very marketplace they are ostensibly helping you access. An agent that presents itself as simply "helping you compare options" is simultaneously a commercial entity quietly deciding which comparisons actually surface and which merchants receive preferential checkout eligibility or favorable ranking, a structural dynamic embedded directly into how OpenAI itself describes its broader agentic commerce ambitions. Handing over your entire comparison layer to a platform-run agent is therefore never a neutral UX convenience, however it gets marketed internally. It represents a slow, steady, and largely irreversible transfer of pricing power and customer data away from your brand and toward whichever organization ultimately controls the agent mediating the relationship.
The winning strategic posture for 2026, then, is not to race toward enabling in-chat checkout everywhere technically possible, following the same instinct that led many retailers to overcommit to Instant Checkout in the first place. It is instead to make your underlying product data clean and consistent enough that agents choose to surface your brand on genuine merit, and then to route that resulting traffic deliberately back toward a conversion experience your organization fully owns, one where you can actually capture durable first-party data rather than watching that value evaporate into a third-party platform's proprietary dataset.
A working framework: the AI Commerce Ownership Matrix
Three examples that illustrate how this plays out in practice
Consider a mid-size apparel brand with consistently strong third-party reviews and product descriptions that remain identical across its own site and every retailer feed it supplies. That brand found itself named, entirely unprompted, inside ChatGPT's comparison answers without ever having run a dedicated GEO campaign, precisely the entity consistency effect that GEO analysts describe as the underlying mechanism. The explanation, once you look closely, is almost mundane in its simplicity. Every mention of that brand across the open web described the same materials, the same sizing philosophy, and the same price tier, giving the underlying model a stable, internally consistent pattern that it could confidently trust and cite when generating a response.
Now consider a home goods retailer that, after Walmart's conversion data became public, ran its own disciplined 30-day test comparing fully structured data, spanning Product, Offer, ShippingDetails, and ReturnPolicy schema, against a plain, unstructured baseline, following essentially the same structured-data remediation that analysts recommended in the immediate aftermath of Walmart's disclosure. The treatment group produced measurably fewer clarification prompts from the agent, meaningfully faster time-to-cart, and noticeably better click-through conversion overall, a result consistent enough to justify a full rollout across the retailer's broader catalog.
And then there is the industrial parts distributor that simply ignored all of it. Its product catalog sits behind a login wall accessible only to existing customers, its specifications live exclusively inside static PDF documents, and it carries no schema markup anywhere on its digital properties. That distributor is now, quite literally, invisible to every AI shopping agent currently evaluating its category, not because its products are inferior to the competition in any meaningful sense, but simply because no model attempting to resolve a purchasing decision has any structured way of determining what the company actually sells.
A few honest edge cases worth acknowledging
High-consideration and B2B purchases will resist full automation for considerably longer than everyday retail categories will. Fully autonomous, end-to-end shopping journeys spanning multiple brands remain meaningfully underdeveloped through 2026 for anything genuinely complex or high-stakes in nature. Regulated categories continue to carry unresolved liability questions surrounding agent-completed transactions, and Fortune's 2026 reporting explicitly flagged fraud protections and return standards that simply have not been settled industry-wide at this point. GEO itself also remains inherently slower to influence than either SEO or AEO, since you cannot directly engineer your way into a model's trained corpus through deliberate effort alone. Brands operating in genuinely new categories may therefore see very little measurable GEO benefit until sufficient third-party content has organically accumulated around them, a timeline that resists most executives' natural preference for quarter-over-quarter results.
Frequently asked questions
Do AI shopping agents actually complete purchases today, or do they mostly just help people decide what to buy?
Most AI shopping activity throughout 2026 continues to function as decision support rather than full autonomous completion. Thirty-eight percent of consumers already use AI in some form while shopping, primarily for comparison purposes, while checkout products such as Instant Checkout were meaningfully scaled back after converting three times worse than conventional click-out traffic.
What is the genuine, practical difference between SEO, AEO, and GEO for a commerce brand trying to allocate budget sensibly?
SEO earns a ranked position within search results over a horizon measured in months, AEO earns the quoted and linked answer inside snippets and AI Overviews within a matter of weeks, and GEO earns your brand an unlinked mention inside a generative answer, built slowly through entity consistency that no single campaign can fully control or accelerate.
If an AI agent discovers my product on my behalf, who ultimately owns that resulting customer relationship?
Ownership currently splits along a fairly consistent line. The platform running the agent retains the underlying intent and comparison signal, according to eMarketer's research on shopping agents, while your brand only captures the durable relationship and first-party data if you deliberately route the resulting transaction to a checkout experience you own, rather than allowing it to complete quietly inside the conversational interface itself.
A 90-to-180-day playbook for reclaiming customer ownership
Days 1 through 30, owned by the Head of Ecommerce or the CIO: Audit your top 50 revenue-generating SKUs for stable identifiers, current pricing accuracy, live availability data, structured shipping windows, and machine-readable return policies, applying essentially the same checklist logic underpinning the structured-data fix that industry analysts recommended following Walmart's disclosure. Flag any SKU scoring below five out of seven fields as an immediate remediation priority.
Days 15 through 45, owned by the Head of SEO or Content: Split your content roadmap explicitly into three distinct lanes, SEO, AEO, and GEO, each assigned a single accountable metric owner, so that these fundamentally different disciplines stop quietly competing for the same undifferentiated budget line and the same undifferentiated success metrics.
Days 30 through 60, owned by the CIO or the Data Platform Lead: Implement Schema.org Product, Offer, ShippingDetails, and MerchantReturnPolicy markup consistently across your top-performing categories, aligned closely with Google's own agentic commerce tooling specifications, with underlying feeds refreshing at intervals no longer than 15 minutes.
Days 60 through 90, owned by the CMO or the VP of Growth: Run a disciplined A/B test comparing a discovery-in-AI, buy-on-site approach against any active in-chat checkout pilots, tracking click-out conversion rates and clarification-prompt frequency closely before committing additional budget to scaling either approach further.
Days 90 through 150, owned by the CFO or the CDO: Instrument reliable first-party data capture at the precise moment of click-out conversion, tying this workstream explicitly to the same governance discipline already established in your organization's AI governance framework.
Days 150 through 180, owned by the CEO or the Board: Review category-level agent-readiness scores across the organization and reallocate budget decisively away from teams still treating SEO, AEO, and GEO as a single undifferentiated line item, applying the same AI vendor evaluation framework discipline already in use elsewhere to hold agency and platform partners accountable for measurable share-of-answer performance rather than vague promises of AI visibility.
The bottom line
Walmart did not fail at agentic commerce in any meaningful sense. It ran the experiment honestly, measured the resulting conversion gap without spin or defensiveness, and adjusted course accordingly, which is precisely the discipline every retail and B2B leader should be applying right now, rather than either dismissing AI shopping agents outright or rushing uncritically to hand them the checkout button before the underlying data infrastructure can support that decision. The brands that ultimately win this decade will not be the ones boasting the flashiest AI integration on their homepage. They will be the organizations whose product data is clean enough to be confidently chosen by an agent, whose content is structured clearly enough to be quoted directly by an answer engine, and whose brand presence across the web is consistent enough to be named unprompted by a generative model, all while the actual transaction, and the durable customer record it produces, remains firmly on ground that the brand itself continues to own.
If you want the more comprehensive playbook I use directly with enterprise clients to build resilient AI commerce and data strategy from first principles, my AI Strategy Book walks through these frameworks in considerably greater depth, and my Entrepreneurship Book addresses how founders specifically should think about defensibility once large platforms begin mediating an increasing share of their customer relationships.
Related reading: AI vendor evaluation framework vs traditional RFPs, AI governance framework for boards, AI cost allocation framework, AI vendor consolidation framework, the AI co-innovation trap, and open-weight AI operating model. If your business needs a hands-on assessment, MD-Konsult Consulting can run the Product Truth Gap audit outlined above directly with your own team.
0 Comments