Market Commentary

AI Shopping Agents at the Retailer's Door: a choice the hotel industry has already made once

Shopping agents built by Meta, OpenAI and others have started arriving at retailers' websites to buy on someone else's behalf. In the space of a month Amazon shut them out and Walmart wrote them a set of house rules. The decision looks technical. It is a decision about who owns the customer, and the hotel industry has already shown what each answer costs.

By Commerce desk × Folka

1. The booking sites came first

In the early 2010s, a hotel that wanted to fill its rooms had little choice but to list them on the large booking sites. The sites brought guests the hotel could not reach alone. In return they took a commission on each booking and, for a time, a say over what the hotel could charge elsewhere. Their contracts often required the hotel not to offer a lower price on any other channel, including its own website.

The arrangement was so one-sided that regulators stepped in. France passed a law in July 2015 removing these "rate parity" terms from hotel contracts, which the hotel trade body HOTREC called the first explicit legislative ban in Europe. In December 2015 Germany's competition authority prohibited Booking.com's best-price clauses, its president noting that he could see "no apparent benefit for the consumer". The hotels then went after their guests directly. Marriott launched "It Pays to Book Direct" in August 2015, and Hilton followed with "Stop Clicking Around" in February 2016, offering members cheaper rates for booking on the hotel's own site.

It worked, but it was not free. Kalibri Labs, a hotel-data firm, tracked about 52 million bookings across roughly 12,000 hotels after the direct-booking campaigns and found more guests than ever choosing to book direct. The hotels had won back a relationship they once held for nothing, and they had paid for it in discounts, loyalty rates and marketing. A decade on, the argument has not ended: in April 2026 the European Parliament was still noting reports of persistently high commissions and of Booking.com practices it described as equivalent to the clauses that had been prohibited.

What the hotels never really had was the option of staying off the sites. The fight was always over the terms. Retail is now being offered a version of the same bargain by a new kind of intermediary.

2. Two doors, one month

The new intermediary is the shopping agent: software that takes a shopper's instruction, compares, and buys. On 8 September Meta released Muse, an agent that shops on its users' behalf. Within two weeks Amazon had blocked it, saying continued access by "an unauthorized AI agent" broke its conditions of use and that Meta had never asked permission. eBay has also blocked agents. Semafor reported both companies' moves as a way to keep control of their advertising businesses, which depend on people, not software, looking at the page.

Walmart made the opposite choice. On 21 October it opened an Agent Gateway, a technical door through which approved agents can search its stock and buy. From November, shoppers can give each agent its own profile, much like an authorised user on a credit card, and decide whether it needs their approval for every purchase or can spend up to a limit. The terms are strict: an agent may not go beyond what the shopper asked, may not see payment details, and is shut out if it breaks the rules.

It is tempting to read this as protection against reach. The fuller picture is that both companies want to be the agent, or at least to write its rules. Since April 2025 Amazon has run its own agent, Buy for Me, which buys from brands' own websites inside the Amazon app, with the brand handling delivery and returns. Amazon objects to other companies' agents at its door, not to agents at doors. Walmart, for its part, was a launch partner for checkout inside ChatGPT in October 2025, and now runs its own shopping assistant, Sparky, inside ChatGPT alongside the open gateway. The market leader closes its door; the challenger opens a controlled one. That is a familiar pattern in platform markets, and it suggests the real choice is not open or closed. It is whether a retailer means to be the agent, or the stock list an agent searches.

Call the set of rules a retailer writes for these visitors its door policy. Until this autumn almost no retailer needed one.

3. Why the door needs a policy now

Agents could shop before September. What changed is that enough of them now do. Cheq, a company that sells tools to verify agent traffic, reported a hundredfold increase in agent visits to the shops it monitors over 45 days, while noting that agents are still a small share of shoppers. Meta says more than ten million people in the US and Canada have set up a Muse agent since launch, and Mark Zuckerberg told investors that demand was high enough that Meta had to ask Stripe to raise its credit limits. The plumbing has followed: Shopify now lets browser-based agents complete a checkout once the buyer has authorised it, and Google has launched a shopping cart that works across Search and its Gemini assistant.

Almost every one of these figures comes from a company with something to sell, and that is itself a finding. Retailers are writing door policies before anyone independent of the agent makers can measure the traffic coming through.

The more important change is speed. OpenAI says its latest models cost half as much to use as their predecessors and make about half as many mistakes on its own factual tests. If agents keep improving at the pace the labs claim, and this is a scenario, not a forecast, the purchases they can handle will move from the routine towards the considered choice within a few product cycles. A door policy written for an agent that can restock washing powder will not hold for one that can furnish a flat.

The counter-evidence deserves the same weight. In October Anthropic cut its internal agent tests off the live internet because it could not reliably control what its agents did there, and OpenAI said it would stop training its models on the live web, citing significant risks. Capability is rising faster than reliability. A retailer deciding its door policy is deciding under both facts at once: the visitors are getting more capable, and nobody, including the companies that build them, can yet promise how they will behave.

4. Who the agent works for

The usual picture of a shopping agent is a loyal servant, acting only for the shopper. The money says something less simple. OpenAI has charged merchants a small fee on purchases completed through checkout in ChatGPT since September 2025, while describing its product results as "organic and unsponsored". Meta plans to charge businesses a fee when Muse helps complete a sale. OpenAI is now adding labelled advertisements beside images generated in ChatGPT and says they will not influence its answers. Instinct, a fast-growing agent platform, would not say whether it earns money from the products it recommends.

The hotel story shows where a commission tends to end up: in the ranking. In 2019 the UK's Competition and Markets Authority secured commitments from six booking sites, including Booking.com and Expedia, to tell people when commission had affected the order of their search results. In April 2026 Italy's competition authority opened an investigation into whether Booking.com's Preferred Partner programme rewards higher commission with more visibility. An agent paid by sellers and acting for buyers carries the same pull. The question is less whether agents will face the conflict than whether paid placement inside an agent's answer will be disclosed, and who will insist that it is.

In the hotel story the booking sites at least worked for the hotels' guests while charging the hotels. A shopping agent may end up charging the retailer, serving the shopper, and answering to its own advertisers, all in the same transaction.

5. What crosses the threshold, and who carries it

When an agent buys on a shopper's behalf, three things change hands that a retailer used to keep. The first is the customer record: who bought, what else they looked at, how they paid. The second is the shelf, meaning which products are seen at all. The third is advertising. Amazon's advertising business earned more than $68bn in 2025, and nearly all of it depends on a person looking at a page. An agent does not look at sponsored listings. Amazon's block is, among other things, the defence of one of the most profitable businesses in retail.

It is also a decision taken on others' behalf. About six in ten units sold on Amazon come from independent sellers, by Amazon's own count. When Amazon shuts its door, it decides for them whether agents can find their goods, and not all of them would decide the same way.

Then there is the purchase that goes wrong: the wrong size, the duplicate order, the disputed charge. Each door policy answers the question of who carries it differently. Walmart's terms keep its own returns process in charge. Amazon's Buy for Me sends returns to the brand that made the sale. Visa and Mastercard announced agent payment programmes a day apart at the end of April 2025, each setting rules for what an agent can be trusted to spend. India points to a further possibility, in which the payment system itself sets the limits. In October 2025 NPCI, which runs the UPI payments network, piloted purchases inside ChatGPT with Razorpay and OpenAI, using a feature that sets funds aside for a single merchant, with BigBasket among the first shops. In that design it is the payment rail, not the retailer, that decides how far an agent can go.

The cost does not fall only on retailers. Vaer AI, working with the commerce platform Lightspeed, reviewed 460,000 shopping answers from AI assistants and found that when ChatGPT named a retailer, 97% of the names were large or mid-sized chains, and small or local shops made up between 1% and 4%. The study was commissioned by a company that sells to small retailers, so its framing has an interest, but the direction is hard to dismiss. A shopper who is shown three options instead of thirty does not experience a narrower market. The narrowing is invisible to the person it affects. The small seller who is never mentioned loses a customer it never knew it had, and the shop assistant whose job was to advise is left with fewer people to advise.

6. Where the agent gets its opening

Agents will not take over all of shopping at once, and the psychology of habit suggests where they start, though not quite where the obvious reading puts it. In well-known diary studies, Wendy Wood and her colleagues found that between a third and a half of what people do each day is habit: the same action, in the same setting, with little conscious thought. Shopping has a large habitual core, from the weekly groceries to the household refill, and it is tempting to conclude that this is what agents take first.

But the habitual core is already automated. Subscriptions, saved baskets and one-tap reorder buttons, most of them run by retailers themselves, handle the purchases that are decided before they are made. An agent adds least where the decision has already been taken. And habits are bound to the settings that cue them. In a 2006 review, Bas Verplanken and Wendy Wood argued that habits resist persuasion and become open to change mainly when people's circumstances shift: moving house, changing jobs. Read together, this suggests two openings for the agent. One is where comparison is tedious. The founder of Instinct says more than half of its transactions are travel, where the options are standardised and comparing them is a chore. The other is the moment when life breaks an old routine: a move, a new baby, a new job, when people are already rethinking how they do things.

Considered purchases are a different matter. Ron Johnson, who built Apple's stores, was asked whether someone would let an agent choose and buy a $1,000 or $2,000 laptop without visiting a site or a shop. "Honestly, nobody's going to do that," he said. Today he is probably right. How long he stays right depends on how fast the agents improve, and on whether shoppers ever come to trust them with a choice that matters.

7. Where the hotel story stops being a guide

The most important break is in whom the law protects. In the hotel story, regulators sided with the supplier against the intermediary. In this story, the first courts to look have started from the shopper's end. Amazon sued Perplexity in November 2025 over its Comet browser agent, and in March 2026 a district judge granted an injunction, finding that Comet acted with the user's permission but without Amazon's. On 4 August 2026 the Ninth Circuit vacated it. Writing for the panel, Judge Milan Smith said it is the user who accesses Amazon's computers, with the agent's help. The court limited its ruling to the record before it, and the case continues. For now, though, a US retailer cannot assume that a closed door is a policy it can simply choose. In Europe, where Amazon has been a designated gatekeeper under the Digital Markets Act since 2023, the Parliament's April resolution pointed to gatekeepers' own assistants, such as Amazon's Rufus, as the thing to examine. A retailer reading the hotel story as "the regulator will eventually rescue us from the intermediary" may find the roles reversed.

Closed doors have precedents of their own, and they cut both ways. In the late 2000s China's Taobao blocked Baidu's search crawler after Baidu launched a rival marketplace, and kept it blocked for years; it could afford to, because shoppers already began their search on Taobao. Ryanair spent years fighting booking sites that resold its seats. A Delaware jury found in 2024 that Booking.com had broken US computer-fraud law, but the judge set the verdict aside in January 2025 because Ryanair had not proved the minimum loss, and the airline's appeal is pending. A closed door can hold when the shop is already where people start. Enforcing it against an intermediary is slow and uncertain.

The strongest argument against this whole read is that retailers face many agents, which shoppers can switch between at no cost, so no single agent will gain the hold Booking.com had over a hotel. It deserves weight, but it is weaker than it looks. The agents arrive built into surfaces people already use by default: Meta's apps, Google Search, ChatGPT. They will remember sizes, addresses and card details, and that memory is a switching cost the booking sites never had. Search was free to switch too, and it concentrated anyway. Trust may slow all of this. Surveys suggest many shoppers are wary of handing an agent their card, and Muse has already given a Toronto user's home address to a stranger arranging a pickup. Seen this way, the door question may be early. It is not obviously wrong.

8. Our view

Strip away the technology and what is happening at the retailer's door is an old argument about who owns a customer, reopened by companies with more reach than any intermediary retail has faced before. Meta and OpenAI are not building shopping agents out of an interest in shopping. An agent stands at the most valuable point in commerce, the moment of decision, and each company already has a way to be paid for standing there. Amazon, which shuts its door to them, is building the same thing for itself. The contest is less between open and closed retailers than over who gets to be the agent.

The hotel story is usually told as a warning about letting an intermediary in. Read more carefully, it is a story about having no real choice to keep one out. The hotels that mattered were all on the booking sites; what they fought over, and partly won back with regulators' help, were the terms. This time the law may sit on the other side of the argument, and the time is shorter. The hotels had a decade to learn what they had given away. If agents keep improving at the pace the labs claim, retailers may have a few years, and a policy that suits this year's agents may not suit next year's.

The people on the other side of the door are hard to read, partly because almost everything known about them comes from the companies selling agents. The surveys say shoppers are wary; the platforms say they sign up anyway. Neither is solid ground. What is clearer is that the cost of an agent's convenience is mostly invisible to the person enjoying it. An answer with three options does not feel like a narrower market, and the small seller who is left out is a loss to nobody but the seller. If the psychology is right, the change will arrive less through the weekly shop than at life's turning points, when people are already letting go of old routines.

We think the retailers treating this as a security question are asking it too narrowly. Underneath is an older question: whether a shop wants to remain a place people visit or become a stock list that agents query. Each answer carries a cost that will not show for some years. A retailer that keeps agents out may find it has kept out a growing share of its own customers, and that the courts will not hold the door for it. A retailer that lets them in may find itself paying, as the hotels did, to win back customers it once had for nothing.

9. The forks, and what to watch

There are four realistic door policies. None is safe; each rests on an assumption that could turn out to be wrong.

Closed. Block outside agents and offer an assistant of one's own, as Amazon has. This assumes shoppers will come to the retailer rather than send someone else, and that the door can be held. Both assumptions are at risk: if agent shopping grows, the closed retailer is invisible to a growing share of buyers, and after the Ninth Circuit's ruling a block may be a contest in court rather than a setting.

Open to chosen partners. Work with one or two agent makers and keep the rest out. This assumes the retailer can pick the agent shoppers will actually use. The assumption most at risk is the choice itself: the partner may not be the agent people end up with, or may become a landlord rather than a partner.

Open to all on terms. Admit any agent that accepts the rules, as Walmart's gateway does. This assumes the retailer can keep enough of the relationship, through profiles, records and its own terms, to stay more than a warehouse. The assumption most at risk is the balance of power: if a few agents come to carry most of the traffic, they will set the terms, as the booking sites did. The fee per sale is already here; the pressure to rank by fee is what follows.

Open by default. Let any agent in and compete on price and stock. This assumes the retailer wins on the things agents compare well: price, availability, speed. The assumption most at risk is margin. A shop that competes only on what an agent can measure is competing on price alone.

Walmart, notably, is running two of these at once.

---

How we read this

This commentary draws on reporting by Fortune, TechCrunch, Semafor, Retail Dive and Reuters; the court record in Amazon v. Perplexity (district court, March 2026; Ninth Circuit, 4 August 2026); decisions by regulators in France, Germany, the UK and Italy and the European Parliament's resolution of 30 April 2026; company disclosures from Amazon, Meta, OpenAI, Walmart, Visa and Mastercard; and Verplanken and Wood (2006) on habit. Figures supplied by companies, including Meta's sign-ups, OpenAI's error rates and Cheq's traffic count, are their own claims. The capability scenario and the 2029 projection are Folka's editorial read, not observed fact. The one number that would most change this read is an independent measure of the share of online sales completed by agents.

Grounded in Folka’s corpus: Checkout Moves to the Agent · Shopping Through the Agent · Wiring the Agent Web · Shopping, One Sentence In