Photo: Pexels
Ask an AI assistant to buy you a pair of running shoes this Christmas and you may get two very different answers depending on where it tries to shop. On Shopify, the assistant can browse more than a billion listings, compare options and head to checkout through Shop Pay, with you approving each purchase. On Amazon or eBay, the same assistant may be turned away at the door. That split, which hardened in October 2026, is the most important thing happening in online shopping right now, and it has almost nothing to do with how clever the AI is.
The fight is about who controls the sale. For thirty years, shopping online meant a person visiting a store: browsing, reading reviews, clicking ads and checking out. An AI shopping agent breaks that chain. The shopper states a goal, the agent does the browsing, and the store never gets to show its ads, its sponsored listings or its carefully arranged homepage. Some of the biggest names in commerce have decided they will not accept that bargain on someone else's terms. Others have decided that refusing it is the surest way to lose the next decade of customers. Both sides cannot be right, but both can be rational, and understanding why is the key to getting ready for a holiday season in which machines do more of the buying than ever before.
What an AI shopping agent actually is
Strip away the marketing and an AI shopping agent is software that acts on a goal rather than a click. You do not tell it to open a website. You tell it what you want, under what conditions: a waterproof jacket under £80, in a size medium, delivered by Friday, from a seller with decent reviews. The agent searches, compares prices and delivery times, reads product details, and in the more advanced versions completes the purchase, subject to rules you set and a final approval from you.
That last part matters, because the word "autonomous" gets overused. The agents actually being offered to shoppers in 2026 are mostly supervised, not free-running. Shopify's version, for example, is marketed as a "personal shopper" that takes budget guardrails and delivery preferences from the shopper and still requires the shopper to approve each purchase, according to reporting of its October announcement by The Verge and others. Google's agentic checkout works in a similar way: a shopper sets a target price and authorises a purchase through Google Pay when the conditions are met. The fearsome image of an AI draining a bank account overnight belongs to science fiction and to badly configured business tools, not to the consumer products on sale today.
There are, broadly, three kinds of agent in play. First, a store's own agent: Amazon's shopping assistant, Walmart's Sparky, a retailer's chatbot that lives inside its own app. Second, an independent agent the shopper brings along: ChatGPT, Gemini, Perplexity or Claude, which can compare across stores. Third, an agent built into a platform's checkout system, such as Shopify's Shop Pay route or Google's Universal Commerce Protocol, which standardises how discovery, payment and orders work. The current fight is mostly about the second kind, the shopper's own agent, because that is the kind a store cannot control.
The split, in one picture
On one side of the split sit the blockers. Amazon has blocked outside shopping agents and sued Perplexity over its shopping assistant, a case that is still running. eBay rewrote its user agreement in February 2026 to prohibit automated agents placing orders without human review unless eBay approves, according to reporting of the change. JPMorgan Chase and Charles Schwab have blocked unauthorised agents from customer accounts, and JPMorgan chief executive Jamie Dimon has said publicly that the bank will not let outside agents act inside its systems because it is, in his words, too risky for clients and for the bank. Delta Air Lines restricts agents booking flights without authorisation, Yelp requires agents to come through its paid data programme, and other names including Amtrak, Cars.com and Kohl's have reportedly put up barriers too.
On the other side sit the welcomers. Shopify opened its billion-plus listing catalogue to independently configured AI agents on October 13, 2026, letting shoppers connect an agent to their Shop account and check out with Shop Pay. Google launched its Universal Commerce Protocol in January 2026 and has rolled out agentic checkout in AI Mode and the Gemini app, starting with US merchants including Etsy and Wayfair. Meta is building licensed shopping experiences with retailers including Walmart, Gap, Sephora and Best Buy. TikTok Shop has gone further, with its own standalone app for sellers. Mastercard, Visa and Stripe are all building payment rails designed for agents, because whoever controls the payment step controls a toll booth on every agent purchase.
It would be a mistake to read this as innovators versus dinosaurs. Amazon is not behind on AI shopping; it runs one of the most advanced shopping assistants in the business, it just wants that assistant to be its own. Shopify is not being reckless; its version keeps a human approval on every purchase and keeps the merchant as the seller of record. The split is not about whether agents will shop. It is about whose agent, on whose terms, and who gets paid for the introduction.
Why Amazon is fighting Perplexity in court
The clearest window into the blockers' case is Amazon's lawsuit against Perplexity. Amazon began blocking Perplexity's shopping assistants last year and then sued, arguing that Perplexity had repeatedly been asked to stop operating its agents on Amazon's site. Perplexity's answer, in essence, is that its software is an extension of the shopper, like an assistant or an employee running an errand, and that a customer should be free to delegate a task to software acting on their instructions.
The courts have not settled the argument. A federal district judge initially granted Amazon a preliminary injunction restricting Perplexity's access, according to reporting by Ynet and others. A federal appeals court overturned that injunction in August 2026, finding that Amazon was unlikely to prevail on the computer-access claims behind its request. That reversal did not end the case: Amazon filed an amended complaint, and the dispute is still active. The practical lesson for everyone else is that the law has not yet decided whether an agent is the customer or a stranger. Until it does, platforms will keep deciding for themselves, and their decisions will be enforced by code long before a judge rules.
Amazon's commercial motive is not hard to see. Amazon's store is, among other things, an advertising business. Sponsored listings, deals placements and the arrangement of search results are how sellers pay to be seen. An independent agent that reads a catalogue directly, compares on price and delivery alone, and skips the sponsored layer entirely removes the store's ability to charge for attention. Semafor's reporting, as summarised in October coverage, put it bluntly: Amazon and eBay blocked buying agents to reinforce control over their advertising businesses and checkout experiences. There is a revealing contrast inside Amazon itself. While its storefront closes to buying agents, its advertising arm has opened an MCP server, a standard machine-readable interface, to agents that want to buy ads, in open beta since February 2026 and across 36 countries by June. The front door is closed to agents that want to buy products; the side door is open to agents that want to buy advertising.
Why eBay and the banks joined the block
eBay's concern is different from Amazon's, and in some ways more understandable. A marketplace lives or dies on trust between strangers, and its rules, buyer protections and fee structure all assume a human is making the decision. An agent placing orders at machine speed, across many accounts, without review, breaks that assumption. It also breaks the fraud model: a human who buys ten identical phones looks suspicious; an agent buying ten identical phones for ten different people looks exactly the same in the data. eBay's February rule does not ban agents outright. It requires approval and human review, which is a way of saying: bring your agent through the front door, with a name on it, and we will talk.
The banks' worry is sharper still. A shopping agent needs payment credentials, and payment credentials are the keys to the kingdom. JPMorgan, Schwab and other large US banks have found unauthorised agents operating inside customer accounts, in some cases acting without clear instructions from the account holder, according to reporting by Axios summarised in CHEQ's October analysis. A bank that lets an outside agent move money on a customer's behalf takes on a liability question nobody has answered: if the agent buys the wrong thing, or buys at the wrong price, or is hijacked by a prompt injected into a product page, who pays? Dimon's answer, for now, is that the customer would, and that is not an answer most customers would accept after the fact. The banks are not being Luddites. They are refusing to underwrite a risk they cannot price.
There is also a quieter, technical reason for the blocks, and it may be the most important one. CHEQ's analysis of the split makes a point that every site owner should read twice: a website today sees five kinds of visitor, humans, AI agents, crawlers, scrapers and malicious bots, and many sites still cannot reliably tell a legitimate agent from a scraper. When you cannot tell the good robot from the bad one, the safe move is to block all of them. That is not a strategy, it is a confession, and it explains why so many blocks look blunt. The industry has standardised how agents reach catalogues and ad platforms. It has not standardised how an agent proves it is who it claims to be, acting for whom it claims to act for, within limits its owner actually set.
Why Shopify went the other way
Shopify's calculation runs in the opposite direction, and it starts from a different position. Shopify does not own the storefront in the way Amazon does; it sells the tools that power other people's storefronts. If shoppers stop visiting storefronts and start asking agents to shop for them, the storefront's homepage matters less and the product data matters more. A merchant whose catalogue is clean, complete and machine-readable wins; a merchant whose listings are thin, contradictory or keyword-stuffed loses, and no amount of advertising fixes that, because the agent does not see the ads.
That is why the October 13 announcement matters beyond Shopify itself. Before it, the main route for an agent into Shopify's world was a partnership, notably Meta's Muse agent, which gained access to Shopify merchants in September 2026. After it, any developer's agent can, in principle, search the catalogue of more than a billion listings, help a shopper choose, and move toward checkout through Shop Pay, without negotiating an individual deal with Shopify. The shopper connects the agent to a Shop account, sets rules such as a budget cap or a fast-delivery requirement, and approves each purchase. Orders flow back into the Shop app. The merchant remains the merchant of record, which means returns, refunds and customer relationships stay with the store, not with the agent's owner.
Shopify's bet is that openness wins the next channel the way openness won the web. It is a plausible bet, and the numbers behind it are striking: Shopify's own 2026 reporting, as summarised in October coverage, says 65 percent of shoppers will use AI when shopping, and independent analysis has put AI-driven orders across Shopify stores at 15 times their January 2025 level by January 2026. But openness has a cost, and October supplied an example. Days before the catalogue announcement made headlines, attackers used a Shopify app's trusted settings to fake checkout across a reported 49,000 stores, and the cost landed on merchants rather than on the platform, according to The Inference's reporting. Open doors let good agents in. They also let bad actors study the lock.
The money question nobody wants to say out loud
Follow the money and the split stops looking mysterious. In the old model, the store charged for attention: ads, sponsored placements, affiliate fees. In the agent model, attention is free, because the agent reads data, not advertising. The store's new revenue must come from somewhere else: a checkout fee, a payment toll, a data licence, or a partnership in which the agent's owner is paid for the introduction. Each platform is choosing the toll booth it can defend.
Amazon can defend its own assistant and its advertising business, so it defends them. Google can defend search intent and the payment step, so it built a protocol and a checkout. Shopify can defend the merchant relationship and the catalogue, so it opened the catalogue and kept the merchant of record. The payment networks can defend the transaction, so Visa, Mastercard and Stripe are building agent payment protocols. The losers in this reshuffle, if anyone loses, are the middlemen whose only value was arranging attention: comparison sites that add no data, affiliate farms, and review operations that an agent cannot verify. There is a reason the same October coverage that reported the agent split also reported that fake reviews may cost shoppers hundreds of billions: an agent that cannot trust what it reads is a liability, not a convenience.
There is a second money question, for sellers. If an agent chooses on price, delivery time and return terms, then the listing fields that used to be afterthoughts, exact dimensions, materials, delivery promises, return windows, stock levels, become the whole pitch. A seller who writes "premium quality, best price, fast shipping" in every listing is telling an agent nothing it can use. A seller who writes the actual weight, the actual delivery time to a specific country, and the actual return policy is telling it everything. The agent does not haggle. It filters.
What this means for shoppers this holiday season
For shoppers, the honest advice for Black Friday 2026, which falls on November 27, with Cyber Monday on November 30, is to use agents as researchers before using them as buyers. Let an agent compare prices, track a price drop, read the delivery terms and flag a suspicious listing. Approve the purchase yourself, on a payment method with buyer protection, and check the seller before the agent checks out. The scams that flooded the last holiday season, fake stores, counterfeit goods, non-delivery, do not disappear because an agent is doing the clicking; an agent can be fooled by a polished fake store just as a human can, and arguably faster.
Two practical rules help. First, set the guardrails before you set the agent loose: a hard budget cap, a short list of acceptable sellers or platforms, and a rule that nothing over a certain amount is bought without your explicit approval. Second, keep the agent inside a licensed channel where you can: a store's own assistant, a platform checkout such as Shop Pay or Google Pay, or an agent partnership the retailer has actually agreed to. An agent scraping a store that has banned agents is not a clever workaround; it is a purchase with no reliable customer service behind it if anything goes wrong.
What this means for sellers: a five-move playbook
For sellers, the shift rewards boring excellence. First, fix one product record and use it everywhere: the same title, specifications, price and delivery promise on every channel, because an agent that finds contradictions will either pick a competitor or drop the product entirely. Second, fill the fields agents actually read: materials, dimensions, compatibility, delivery time by destination, return window and stock status, in structured fields rather than buried in a paragraph. Cross-border sellers should also watch the import-duty rules as the de minimis exemption ends. Third, price honestly, including delivery, because an agent that discovers a delivery surcharge at checkout will abandon the cart and may not come back. Fourth, decide your agent policy before a customer asks: whether you welcome agents, through which channels, and what happens on a return. Fifth, watch your own store's agent traffic, not just its human traffic, because the first sign of the shift will be a rising share of visits that never see your homepage at all.
One warning belongs in any playbook written in 2026: do not buy an "agentic SEO" tool before fixing the data. October coverage of the Shopify shift carried a useful anecdote, of a merchant who stacked three such tools until the catalogue contradicted itself across surfaces. Agents read the same structured fields search engines have read for a decade. Keyword-stuffing is keyword-stuffing in any decade, and a catalogue that contradicts itself is worse than a thin one, because the agent cannot tell which version is true.
What happens next: three scenarios
The most likely near-term outcome is a messy middle. Licensed agents, the store's own assistant, a platform checkout, a named partnership, become the normal way agents shop, while unlicensed scraping stays blocked and litigated. Shoppers get most of the convenience, platforms keep most of the control, and the Amazon-Perplexity case sets the legal boundaries slowly, case by case. This is the boring scenario, and boring scenarios are often the ones that happen.
The second scenario is a standards breakthrough. If the industry agrees on a way for an agent to prove its identity, its owner and its spending limits, in a form every store can verify, then the blunt blocks become unnecessary and the walled gardens lose their justification. The payment networks are the most likely authors of that standard, because identity plus payment is their home territory. Watch for a Visa, Mastercard or Stripe agent protocol gaining real merchant adoption, not just announcements, as the signal that this scenario is arriving.
The third scenario is a backlash. A high-profile agent failure, a large wrong purchase, a hijacked agent, a fraud wave traced to agent checkouts, could produce regulation that freezes the whole field, or a court ruling that gives platforms an absolute right to refuse agents. The banks' caution is, in part, preparation for this scenario: they would rather be criticised for moving slowly than be blamed for a failure they enabled. Any of the three could dominate by next Christmas. The only prediction worth making with confidence is that the shopper who never visits a storefront is not going away.
Frequently asked questions
Can an AI agent buy things on Amazon for me?
Not through an independent agent, in general. Amazon blocks outside shopping agents and is suing Perplexity over its shopping assistant, though the case is unresolved. Amazon's own shopping assistant, inside its app, is the licensed route, and it can buy from other stores for selected items through its Buy for Me feature.
Can an AI agent buy things on eBay for me?
Only with eBay's approval and human review. eBay's February 2026 user agreement prohibits automated agents placing orders without oversight unless eBay agrees. An agent that scrapes eBay and checks out without permission risks the order being cancelled and the account being restricted.
How does Shopify's AI shopping work?
A shopper connects an agent of their choice to a Shop account. The agent can search Shopify's catalogue of more than a billion listings, help choose a product, and move to checkout through Shop Pay. The shopper sets rules such as a budget cap and must approve each purchase, and orders flow back into the Shop app.
Is it safe to let an AI shop for me?
It is as safe as the channel and the guardrails. Licensed channels with human approval, buyer protection and a real merchant of record are far safer than an agent scraping a store that has banned agents. Set a budget cap, require approval above a threshold, and use a payment method with chargeback protection.
Will AI agents replace online stores?
No, but they are changing what a store is for. The storefront as a place to browse may matter less, while the product data, the price, the delivery promise and the return policy matter more. Stores that treat their catalogue as their real shopfront will do better than stores that treat it as an afterthought.
What should a small seller do first?
Fix one listing completely, with accurate specifications, delivery times and return terms in structured fields, and keep it identical on every channel. That single piece of work helps in ordinary search today and in agent search tomorrow, and it costs nothing but care.
The bottom line
The agent split of October 2026 is not a technology story. It is a story about who gets to stand between a shopper and a sale, and what they get to charge for standing there. Amazon and eBay are defending a model in which the store sells attention. Shopify and Google are building a model in which the store sells data and checkout. The banks are refusing to pay for a risk they cannot measure. All three positions make sense from inside the companies holding them, which is why the split will not be resolved by one court case or one product launch.
For shoppers, the practical move is to enjoy the research and keep the final approval. For sellers, it is to make the product data so clean that any agent, licensed or not, can only conclude one thing: this is the product, at this price, delivered by this date, with these terms. The storefront may fade from view. The truth of the listing will not. That is a change worth preparing for before the holiday rush, not after it.
For more explainers on the forces reshaping online business, visit abdulrahimmughal.com.
Sources
- Reuters, "AI shopping agents could threaten Amazon and Walmart's ad business" (4 November 2026) — reuters.com
- Reuters, "Amazon wins order blocking access for Perplexity's AI shopping agent" (10 March 2026) — reuters.com