When AI does the shopping, who gets the customer?

For a read-only check on 23 October, Marketer Fizz opened bol’s Belgian, Dutch-language Shophulp while signed out. The request was ordinary: a cabin suitcase for a four-day train trip, under €120, with quiet wheels, low weight and a laptop compartment. It asked for up to three options, including the seller, delivery and returns, and explicitly ruled out adding anything to a basket or buying it.

The assistant returned three product cards and a comparison. For the ONYX suitcase, its prose said “wordt verkocht door bol”: sold by bol. The card itself named Qumax Brands. Opening the linked product page confirmed Qumax Brands as the seller, with payment through bol and dispatch by bol. The recommendation had compressed three different roles into one familiar name.

There was another handoff hiding in the link. Although the question began on the Belgian storefront, the product opened on bol’s Dutch storefront. That page displayed a 30-day consideration period and free returns through the bol account. Those were the terms visible on that page at that moment; they do not verify the Belgian purchase or return experience.

This was one query, without a purchase, delivery or attempted return. It is an observation about the information shown, not a ranking of shopping assistants. But it exposes something worth following: a helpful shortlist still leaves a shopper needing to know who is actually making the promise.

The clearest consumer evidence in this reporting comes from PYMNTS Intelligence’s October report. Among respondents describing their most recent AI-assisted purchase, 59.7% bought on Amazon; 19.5% used big-box or department stores; 6.3% bought on brands’ own sites or apps. These are purchase destinations after AI help, not shares of purchases completed inside chatbots.

The survey covered 2,191 US consumers on 10–13 August, balanced against census characteristics. It records reported behaviour rather than transaction logs and cannot tell us how Belgian or Dutch shoppers behave. It also cannot establish that AI caused Amazon to gain sales.

For a marketer, the distinction changes the story. A product can win the comparison while its own website loses the visit. A retailer can lose the initial question yet keep the sale. A shopper can enjoy an unfamiliar assistant’s advice and still prefer a familiar place to spend money. The journey need not have a single winner.

The first contest is for a place in the answer. A suitcase brand is no longer talking only to someone browsing pictures. It also needs its product to make sense when a person asks about a laptop pocket, train travel or weight. A lovely campaign and an accurate product description have different jobs; this shopping question needs both.

OpenAI’s current shopping guide says ChatGPT selects product results independently of partnerships. It distinguishes those results from advertising. The same guide says the first price shown may not be the lowest and that merchant updates can take time to appear. A confident sentence is still a reason to open the offer, not a substitute for it.

That distinction matters for paid visibility. OpenAI’s advertising guidance says ads are labelled and separate from answers; advertisers cannot buy changes to the response. Ads may appear for eligible Free and Go users, while Plus, Pro, Business, Enterprise and Edu accounts remain ad-free. Buying an ad and earning an ordinary product recommendation are therefore different routes to attention under the company’s stated rules.

This is not a rule for every assistant. A retailer building its own shopping adviser can make different commercial choices. The useful question for a reader is whose selection they are seeing: a shop’s assortment, an independently selected product result, or a paid placement. A conversational voice should make those boundaries easier to understand.

On 29 September 2025, OpenAI announced a turning point: eligible US users could buy from US Etsy sellers inside ChatGPT. The shopper confirmed the order there, while the merchant handled payment and fulfilment. That was the launch’s scope, not a description of every product or country available today. Current developer guidance still places order acceptance and charging on the merchant’s systems.

Google’s Universal Cart, launched in the US on 15 October 2026, tackles another familiar nuisance: products scattered across tabs and shops. Its announcement describes saving items across Search and Gemini, following price changes and finishing either with Google Pay at participating shops or on the merchant’s website. Both routes leave the merchant sending confirmation and providing support.

For the shopper, staying in one place can feel like dealing with one shop. The organisations behind the screen need not merge. Google’s merchant guidance explicitly keeps the participating retailer as the seller of record: the business making the sale. A basket is a convenient place to collect choices; it is not, by itself, a completed order.

Shopify’s documentation makes that last point unusually tangible. In its Shop shopping connection, carts belong to individual merchants: products from three shops need three checkouts. Creating the cart places no order. A conversational front door can hide several shop counters behind it.

A US shopping launch is easy to mistake for the future arriving everywhere at once. Shopify’s current Google checkout instructions are more specific: participating stores and customers must be in the US. Customers elsewhere can discover products but are sent to the shop’s website to buy. Google’s broader merchant guidance covers eligible products in the US, Canada and Australia; that does not establish a European Universal Cart rollout.

Bol offers a closer example for readers in Belgium and the Netherlands. Its 15 September announcement opened Shopper Agent in beta across its website and app, making recommendations from bol’s assortment and leaving the final purchase in its familiar shopping environment. Our Shophulp check showed that local front door in use, while also exposing a country change when following a product.

Retailers are making deliberate choices about that door. In BCG’s 21 October report, bol’s Joris Scheepens says it intends to keep external agents focused on discovery and retain the transaction. BCG also reports that bol built advertising into its own agent from launch. This is a retailer’s commercial service, not a neutral tour of every shop on the web.

Nordstrom’s Miguel Almeida describes another position in the same BCG reporting: willingness to offer loyalty benefits outside its website if Nordstrom remains the merchant. Its checkout capability was largely built but waiting for customer demand. These are executives’ accounts to a consultancy that advises retailers, not interviews conducted by Marketer Fizz or independent conversion experiments.

The choices reveal a disagreement more interesting than a race to add a chat box. One retailer wants the customer to come back through its own door. Another can imagine serving the customer somewhere else, provided the sale remains its own. Neither position tells us yet which shoppers will return.

A new wrinkle arrived after this feature was first published. In her 22 October test for The Verge, reporter Mia Sato instructed Meta’s Muse to put ten newly listed items from a favourite brand into a RealReal shopping cart, then add them again whenever the 20-minute reservation expired. She reports that the repeated holds kept those items unavailable to other shoppers for hours. This was a deliberate test, with no intention to buy, rather than an agent spontaneously deciding to hoard clothes.

In her 26 October follow-up, Sato reports that The RealReal tightened its bot defences. Chief technology officer Luke Friang told her that logging in, adding items to a cart and checking out would become much harder for all agents. They could still find listings and direct people to product pages. Meta did not immediately comment on that update; in the earlier report, its spokesperson said Muse asks permission before buying and users can take over.

The distinction matters before anybody pays. A cart can reserve scarce stock and affect another shopper even when no purchase follows. This case also shows a retailer keeping product discovery open while making the later steps harder. It is one reported test and a company’s stated response, not our own checkout trial or proof that every route is blocked. For European readers, it does not establish local Muse availability.

The question for a brand is therefore more concrete than whether it welcomes AI shoppers: which actions may an outside assistant take, and what happens to everyone else when it repeats them? A recommendation can cross the shop’s threshold. Permission to hold the goods is another matter.

Return to the suitcase. It is easy to imagine a smooth recommendation followed by an awkward question: the zip has broken, who can fix this? That is an illustration, not something our check tested. It is also the moment when the identity of the seller stops being a detail in a product card.

OpenAI’s original checkout announcement assigned fulfilment, returns and support to the merchant. Google’s cart launch similarly leaves confirmation and customer support with the seller. Neither promise demonstrates that an assistant will successfully arrange a return. Explaining a policy, sending someone to an order page and getting money back are different experiences.

There are practical trade-offs even before a complaint. Shopify says some familiar extras, including product bundles, subscriptions and certain loyalty or rewards screens, are unsupported in its Google direct checkout. A shorter path to payment may therefore present less of the shop’s usual experience. For a retailer, the question is which parts are useful service and which are merely another obstacle before paying.

Our editorial view is that the most durable advantage will be a promise a shopper can recognise and verify across those handoffs. That means a seller with a name, an offer that survives the click, and somewhere dependable to go when the purchase needs attention. Accurate product information helps an assistant make a choice. Keeping the promise gives a person a reason to choose the shop again.

The next battle for the customer will not be settled by the most fluent recommendation alone. Watch what happens when someone changes their mind, asks for help or buys again. AI can move the conversation. The relationship still has to be earned.