In September 2025, OpenAI announced that people could buy eligible products inside ChatGPT. Merchants would pay a fee on completed purchases. The same announcement said those fees would not influence which products appeared in the results.[1]
Both statements belong in the story. Leave out the fee, and you conceal a commercial relationship. Leave out the stated separation, and you imply a conclusion the evidence has not earned. A serious inquiry has to hold both in view, then ask how the arrangement works and what would show whether it works as promised.
A commission can fund a useful service. It can also give the service a reason to favor completed sales over other outcomes. Whether that reason enters a particular recommendation depends on choices made inside the business: what the software is asked to do, what information it receives, how performance is judged, and which boundaries the company maintains.
The existence of a payment gives us a place to begin. It does not finish the investigation. That may seem a less exciting position than announcing that your assistant has betrayed you. It is a more useful one. You cannot protect yourself from an arrangement you have misdescribed.
Follow the payment far enough
Imagine the purchase of a hundred-dollar pair of shoes. The buyer sees the price and decides whether the shoes are worth it. Elsewhere, the transaction may pay for manufacturing, delivery, a marketplace, payment processing and the service that helped the buyer find the item. Different firms can be paid at different stages. The buyer does not need to know every supplier’s profit to make a sensible choice.
Still, some payments are directly relevant to the advice. If a service earns money only when you purchase, it has a commercial interest in a purchase occurring. If it earns different amounts from different sellers, the choice of seller may matter to its revenue. These are possible pressures within the arrangement. We need evidence before claiming the service acted on them.
A useful question is what happens when helping you requires ending the sale. You might decide to repair the shoes you already own. You might borrow a piece of equipment rather than buy it. You might ask an assistant to cancel a subscription or to explain why the cheaper version is adequate. Sometimes the best outcome for the customer produces no new transaction.
A purchasing assistant should be capable of recognizing those outcomes. If its definition of success always ends at checkout, the customer’s interest has been made too narrow.
Consider how an ordinary business target might work. A team could set out to reduce the time between a query and a completed purchase because customers find checkout tedious. That may be a good improvement. But if the team measures only completion, it can miss cases in which an extra question would have prevented a bad choice. The measure rewards smooth motion, while the person still needs a suitable destination. This is a general problem in choosing targets, not a claim about a particular company’s private dashboard. What a business chooses to count affects what it can notice. A service may need measures for successful returns, unsuitable purchases avoided or errors caught before payment, alongside its sales figures.
Of course, these measures can be gamed too. An assistant could make a show of caution, asking unnecessary questions to display how responsible it is. Counting refusals would not solve the problem any more than counting purchases. We need to connect the measure to the user’s actual purpose.
That purpose can be awkward for the business without being hostile to it. A customer who is well advised may return and buy something else. A retailer may earn trust by suggesting a repair. A subscription service may accept a cancellation cleanly because its reputation has value. Revenue and customer welfare often support each other. The important word is often. A sound relationship needs to survive the moments when they do not.
Paying for help does not buy the whole company
A subscription seems to offer a clean answer. Pay the assistant directly. Let it work for you.
There is real value in this arrangement. The service can earn revenue without turning each conversation into a sale. A person asking for help with a poem, a family plan or a difficult decision need not become an immediate source of advertising income.
Anthropic made this distinction part of its public position in February 2026. It said Claude would remain ad-free and discussed subscriptions and business customers as its sources of revenue. Its argument was that advertising could create pressures inconsistent with the kind of help it wanted to offer.[2]
That is a business choice worth examining. It is also a company explaining the merits of its own choice.
Direct payment does not mean the provider has no interests of its own. It still has costs. It may prefer tasks that are cheaper to perform. It may want you to remain a subscriber, use other products or recommend it to friends. It has legal obligations and a reputation to defend. It may decide that a feature is too expensive or risky to provide. A monthly payment purchases a service under certain terms. It does not make you the owner of the company or give you the authority to override every rule.
This matters because people sometimes speak of a paid assistant as if payment settles the question of loyalty. Think about a subscription that promises extensive research but must cover the cost of every search. There can be tension between the depth you would like and the cost the provider can sustain. A responsible service can manage that tension through clear limits, appropriate pricing and honest explanations. An irresponsible one might simply do less work than the user is led to expect. The difference is in the conduct, not in the mere existence of a subscription.
There is a second problem with declaring paid access the answer to every concern. Some people cannot afford it. If the safest, clearest and most accountable forms of assistance are reserved for those who can pay more, the burden of uncertainty falls on those with less room to absorb a mistake. A market can offer different levels of speed or capacity. Basic honesty about what a recommendation represents should not be a luxury feature.
This is an opinion about what providers should owe their users. It does not follow automatically from a price tag. It follows from the role the service asks to occupy. If it invites a person to rely on its judgment, it should make material conflicts and limits understandable to that person. Paying directly may improve some incentives, but the obligation remains.
Three things that can look like one
Advertising is especially easy to discuss badly because several different activities can appear in the same small space on a screen. One is an answer to your question. Another is a paid advertisement displayed nearby. A third is a shopping feature that selects products or ranks sellers. They can be connected, but they are not interchangeable.
OpenAI’s current help page says that ads may appear for eligible Free and Go users, while the other listed paid and institutional plans are ad-free. It says ads run separately from the chat model and do not shape its answers. The ad system can use the context of the conversation, with additional signals where personalization is enabled. OpenAI also says advertisers do not receive users’ chats; a person who sends a direct message to an advertiser is sharing that message separately.[3]
These distinctions are important enough to state plainly. Selecting an advertisement using conversational context is a use of information. Sending the conversation itself to the advertiser would be a different use. Placing an ad below an answer is different again from changing the answer to favor the advertiser. Calling every one of these things “selling your conversation” would make the story vivid and inaccurate.
The differences do not remove every question. They help us ask better ones. How clearly is the advertisement identified? Can a hurried reader distinguish it from the assistant’s recommendation? What happens when the person asks a follow-up about it? Which information is used to choose the ad, and can the person control that use?
We can examine those questions without claiming that a paid placement has secretly rewritten the answer. The same care applies to shopping results. OpenAI’s 2025 checkout announcement distinguished selecting a product from ranking merchants that sell that product. It said checkout availability could be one factor in choosing among merchants, while saying products themselves were not favored for offering checkout. Its current shopping help page describes product selection and merchant ranking separately.[1][4]
A person may reasonably prefer the seller with an easier checkout. Another may prefer to leave the conversation and buy directly. Convenience is a genuine benefit, but it has a value that the customer should be allowed to judge.
Suppose two sellers offer an identical item. One costs a little more but makes returns easy. The other is cheaper and difficult to contact. Ranking the first seller higher need not be corrupt. It could reflect a sensible judgment about the whole purchase. But the assistant should be able to explain the relevant tradeoff rather than hide it inside the word best.
Now change the facts. The more convenient seller pays the assistant’s provider. Is the ranking still justified? It might be. The presence of a commercial relationship raises the need to inspect the reason; it does not by itself invalidate the reason.
The difficult case is where a benefit to the user and a benefit to the provider travel together. A faster checkout can genuinely save you time while also making a revenue-generating transaction more likely. It would be quite reasonable to choose it, provided you could see the alternatives and decide that the time saved was worth any difference in price. The problem begins when the provider treats its preferred route as if you had instructed it to take that route. A good recommendation needs room for your own judgment about the tradeoff.
The helpful shopkeeper
A retailer’s own assistant presents the relationship more openly. You are in a shop. The shop would like to sell you something.
Amazon’s May 2026 description of Alexa for Shopping includes product comparisons, price history, cart building and supported purchases from other stores. It also describes using information from Amazon and Alexa to personalize the experience. The stated range is broader than a tool limited to recommending whatever happens to be on one Amazon shelf.[5]
A retailer can know a great deal about its products and provide excellent help. It may find the right replacement part, explain an unfamiliar feature or help avoid buying something unsuitable. None of that ceases to be useful because the business hopes to make a sale.
The question is what kind of representation the shopper should expect. If you ask a shop employee whether the store has a suitable coat, you can interpret the answer within the setting. You know where you are. When the same guidance arrives through a personal voice that also helps manage your day, the setting can become less obvious. A sentence that sounds like advice from your own assistant may also be an offer from a seller.
We should not have to infer the commercial role from the warmth of the language.
An assistant can serve a retailer and a customer in many compatible ways. It can also face a choice between them. If a cheaper adequate item sits outside the accessible catalogue, can it say so? If a repair would be enough, will it explain the repair? If it lacks reliable information, will it refrain from decorating a sales pitch with false certainty?
Those are questions about the service’s scope and behavior. The answers should come from its documentation and suitable tests, rather than from the assumption that every shopkeeper is dishonest or every helpful voice is independent.
A useful retail assistant could make the commercial boundary quite clear: these are the products I can compare; this is how sellers enter the results; here is what I could not check. That would not make the service less useful. It would make its usefulness easier to judge. A shop can help you without claiming to stand outside commerce.
A promise needs somewhere to stand
Meta’s Muse announcement makes a specific privacy commitment: conversations and data in the user’s virtual machine are not shared with Meta’s advertising systems. Its technical explanation also distinguishes that commitment from the ordinary traces left when an assistant visits or buys from another website.[6]
That is a reason to avoid transferring assumptions about one product’s business model onto every service made by the same company. It is also a reason to examine the actual boundary being promised.
The word private can cover several questions. Who stores the information? Who may use it to improve a model? Which party receives it when a task is carried out? What does an outside merchant learn from a purchase? We will examine these flows in a later chapter. For now, the point is that a useful promise identifies what it covers.
“Your interests come first” expresses an aspiration. A statement that a particular category of data is excluded from a particular system gives us something more definite to check.
Even a clear promise leaves the question of how it is maintained. A company might separate systems, limit employee access, constrain data use, commission testing or provide records. Different measures protect different things. A promise can be sincere and still fail in implementation. A safeguard can reduce a risk without eliminating it. Readers do not need to become auditors of every provider. The companies that build these services are in a much better position to test their own systems and supply evidence. Outside scrutiny matters too, particularly where users cannot observe the relevant process themselves.
We should expect claims of user control to be matched by controls people can use. We should expect claims of independence to be matched by procedures that preserve it. The burden of explaining a relationship should grow when a service asks to take on more of the user’s decisions.
This is where the familiar instruction to read the fine print becomes inadequate. A person can spend an evening reading terms and still have no evidence about how a recommendation was selected. The terms may describe broad rights without explaining the choice in front of them.
Good disclosure should reach the decision it is meant to improve. If the assistant has compared only participating sellers, that information belongs with the comparison. If a displayed item is a paid placement, the label belongs with the item. A disclosure buried elsewhere cannot help a user distinguish two things that look the same at the moment of choice.
The FTC’s September 2025 inquiry into companion chatbots included questions about monetization, disclosures and conversation data. Its focus was especially on children and teens, and the inquiry itself was not a finding of wrongdoing.[7] The relevance here is the shape of the questions. A service’s intended relationship with the user, its source of revenue and its use of information need to be examined together.
Which interests are allowed in the room?
It is tempting, after tracing the money, to demand an assistant that follows only the user. The idea needs a boundary. A user can ask for something that harms another person. They can want access they have no right to obtain. They can authorize a message that discloses someone else’s private information. A provider may properly refuse such requests.
Commercial self-interest is only one reason a service might depart from the user’s wishes. Safety rules, legal duties and the rights of other people also shape what it can do. A refusal should be examined in context; it is not automatic evidence that the company has chosen profits over the customer.
There is an equally important boundary in the other direction. Invoking safety should not end every question about a provider’s choices. A restriction can serve a legitimate purpose and still be broader than necessary. An error can block a harmless request. People need intelligible explanations and, where appropriate, a way to challenge the result. Making those distinctions visible gives the user a more useful kind of power than a vague assurance that the service is on their side.
Consider the various meanings of the word boss. You set the task. The provider sets many of the operating rules. A website may control access to its service. A payment company may reject the transaction. The seller decides whether to accept the order. None of these powers is identical to the others, and no single cheerful reply makes them disappear.
A trustworthy service helps you understand the parts that matter to your decision. It does not need to expose every business secret. It does need to avoid pretending that its scope is the whole world, that its interests are absent, or that your final approval resolves every earlier choice.
We can now ask more of the title’s question. Who AI serves will not always be revealed by identifying who pays for it. We must also examine who chooses its goal, who supplies its options, and who can correct its result.
The next chapter starts with something that never reaches the checkout page: the option you did not get to consider.
Sources
[1] OpenAI, “Buy it in ChatGPT: Instant Checkout and the Agentic Commerce Protocol,” September 29, 2025. Historical launch account of fees, product selection, merchant ranking and transaction roles. It is not a statement of every current commercial term.
[2] Anthropic, “Claude is a space to think,” February 4, 2026. The company’s stated ad-free position and argument about incentives; not independent evidence about competitors.
[3] OpenAI, “Ads in ChatGPT,” accessed September 30, 2026. Current help page on plan eligibility, answer independence, personalization and advertiser access. Availability and controls vary by region and account.
[4] OpenAI, “Shopping with ChatGPT Search,” accessed September 30, 2026. The current page and the 2025 launch post describe different dated snapshots; omission of a launch factor from the newer page alone does not prove that factor was removed.
[5] Amazon, “Meet Alexa for Shopping, your personalized, agentic AI assistant on Amazon,” May 2026. Company description, not independent testing or a claim of universal availability.
[6] Meta, Muse announcement, September 8, 2026, and “How We Built Safety Into Muse,” September 8, 2026, especially “Our Policy Around Data.” The discussion distinguishes direct sharing with Meta advertising systems from effects of activity at outside services.
[7] Federal Trade Commission, “FTC Launches Inquiry into AI Chatbots Acting as Companions,” September 2025. Section 6(b) information gathering; neither a finding against the recipients nor a specific inquiry into the shopping products discussed here.