Advertising Need to Convince Your AI, Not You

By Javier Surasky

AI assistants can help us compare products, but sellers are already working to influence their choices, opening a new struggle over our purchasing decisions.

A promotional phrase in a product description can make an AI assistant prefer it, even when it costs more. That is one finding from a study published on October 19, 2026, examining how much influence sellers can exert over systems that choose products on our behalf (Buecheler et al., 2026).

The issue starts with an everyday expectation: when we ask an assistant to compare offers, we expect it to weigh the options against our needs, whether we want to save money or avoid a bad purchase because we do not understand the technical differences.

In either case, we entrust part of the choice to a tool whose usefulness depends on how it interprets the information available. Yet we forget that this information comes from someone trying to sell us the product. Delegating the task therefore introduces an intermediary that sellers can also target.

I see a question here that we rarely ask, although the answer matters in our everyday lives: how can we tell whether a recommendation reflects our criteria when other actors can influence the system making it?

The Seller Is Part of the Comparison

Buecheler et al. (2026) conducted 113,472 trials using seven models and products from two categories: headphones and coffee machines. They changed commercial cues, such as popularity claims or expert endorsements, and varied prices to measure how far a preference for one option would hold. Their findings showed that certain cues could favor a product despite its higher price, with effects varying by model. Available as a preprint, the study examines controlled choices between two alternatives: it measures responses under those conditions, rather than the full behavior of an assistant browsing an online store.

The authors caution that a cue’s influence does not, by itself, demonstrate irrational bias, and point out that a claim may contain information relevant to the choice (Buecheler et al., 2026). An expert recommendation may indeed justify paying more if it rests on verifiable evidence. But that is precisely where the difficulty lies: buyers do not know whether the claim has been verified or how much weight it carried relative to other features of the product an AI recommends or buys through an agent.

From that perspective, I do not think it is enough to ask whether the assistant carries paid advertising. Nor am I claiming that commercial agreements exist between those who train the model and particular suppliers. I do argue, however, that AI-mediated purchasing decisions should be more transparent.

Choosing Like Us Does Not Mean Deciding Like Us

The relationship between personalized assistance and commercial interests has long featured in critical thinking about digital technologies. Sadin (2020) examines how personal assistants build a sense of familiarity that brings brands and consumers closer together. In his discussion of conversational commerce, product recommendations and transactions form part of an architecture that steers behavior (section “El estadio incitativo de la verdad: estar lo más cerca posible de los cuerpos y las mentes”).

His analysis brings into focus the interests shaping the relationship between an assistant and the person consulting it. A tool that remembers our preferences and responds to our requests can be convenient and valuable, but that familiarity encourages a perception that the user’s interests come first, implying an independence the service has not demonstrated. A recommendation that feels tailored to us does not establish that the criteria behind it protect our interests.

When we delegate, we expect the assistant to compensate for our limitations by reviewing more options, comparing them more closely, and handling other purchasing tasks that we cannot do or do not want to spend time on. That expectation needs to be tested.

Bini et al. (2026) found that more advanced or larger models were not consistently less biased. In preference-based decisions, their responses tended to become more human-like, including departures from the standard of rationality used by the researchers. Instructing the models to adopt rational criteria reduced some biases, although the improvements were modest (pp. 2–4).

The suggestion that AI responds as we do brings us to Sun et al. (2026), who studied product choices and found that framing a purchase as a treat or a practical necessity changed the options selected. Yet those shifts did not consistently reproduce the changes in sensitivity to price and quality predicted by the theory of rational human behavior (sections 5 and 6).

Context can justify different priorities: buying something to meet a need and buying it as a treat are not equivalent situations. What this study calls into question is whether similar choices allow us to infer a shared mechanism. When discussing human and algorithmic biases, we therefore need to distinguish what behavior we observe, under what conditions, how we evaluate it, and how the final decisions are reached.

Recommendations Can Also Change Us

The assistant’s assessment and recommendation are one part of the process; what people do with those responses is the other.

Glickman and Sharot (2025) examined this issue in a series of experiments involving 1,401 participants and perceptual, emotional, and social judgments. They found that interacting with biased systems could amplify human biases, while participants were not always aware of how strongly they were being influenced. Conversely, accurate AI could improve decision-making.

The interaction can therefore lead to different outcomes. This is why I propose examining whether commercial influence over an assistant could combine with the trust buyers place in its recommendations when making a purchase. None of the studies discussed follows that entire sequence.

The question deserves attention because someone could take a conclusion shaped partly by the presentation of an offer to be an independent assessment. If that conclusion also reinforces an existing preference, they may see fewer reasons to question it.

What we need to understand is when assistance helps us challenge our assumptions and when it gives us additional, perhaps false, confidence in a choice we have not examined closely enough.

What It Means to Act on the Buyer’s Behalf

Comparing options takes money, time, and knowledge that we do not always have. A tool that can interpret specifications and assess offers can help us understand our options before choosing. That usefulness makes it important to define what service providers must commit to.

Protecting buyers means specifying what each participant should be required to do. Assistant providers should respect buyers’ priorities and seek authorization before departing from them, disclose commissions and commercial agreements that apply to a purchase, and explain the range of options compared and why those options were selected.

Assistants should also distinguish between claims they repeat and claims they have verified, while allowing buyers to fine-tune the factors used to choose a product: price, quality, the seller’s verified credentials, installment payment options, and so on.

Finally, assistant providers should be required to test their services’ susceptibility to commercial content that changes their choices without improving the offer, and to publish the results.

Sellers should have obligations too: claims about popularity, scarcity, or expert endorsement should be backed by verifiable evidence, and information about the total price, availability, and terms should be consistent whether it is read by a person or by a system assisting them. Sellers must be prohibited from including instructions in their offers that tell an assistant to ignore the buyer’s preferences or hide alternatives.

I propose allocating these responsibilities according to what each actor can control: sellers should answer for the information and instructions they include in their offers; assistant providers, for how they process and use them to recommend or buy. Both should maintain complaint channels that allow users, as buyers, to trace the steps each party took.

Delegating a purchase can save us effort, but it should not limit our ability to understand how the decision was made or hold a service accountable when it fails to deliver what it promises.

References

Bini, P., Cong, L. W., Huang, X., & Jin, L. J. (2026). Behavioral economics of AI: LLM biases and corrections (Working Paper No. 34745). National Bureau of Economic Research. https://doi.org/10.3386/w34745

Buecheler, F., Berbeglia, G., & Guo, M. (2026). A cue’s hidden price: Auditing seller influence on AI shoppers [Preprint]. arXiv. https://doi.org/10.48550/arXiv.2610.25843

Glickman, M., & Sharot, T. (2025). How human–AI feedback loops alter human perceptual, emotional and social judgements. Nature Human Behaviour, 9, 345–359. https://doi.org/10.1038/s41562-024-02077-2

Sadin, É. (2020). La inteligencia artificial o el desafío del siglo: Anatomía de un antihumanismo radical (M. Martínez, Trans.). Caja Negra. (Original work published 2018).

Sun, J., Shleifer, A., & Belinkov, Y. (2026). Do LLMs choose like humans? Using cognitive theory to evaluate LLM decision-making [Preprint]. arXiv. https://arxiv.org/abs/2609.22225