Your Next Customer May Be an AI. It Falls for the Same Tricks.

Source: Swiss Pricing Lab (Substack) · 2026-09-29

Everyone says AI shopping assistants will make pricing tricks useless. We read the experiments. It is not that simple.

There is a prediction going around that sounds very sensible. Once software does our shopping, the old pricing tricks stop working. The 9.99 ending. The countdown timer. The “most popular” badge. A machine does not feel rushed. It does not get charmed. It just picks the best deal.

It is a comforting story, especially if you have never liked those tricks. So we went looking for the research behind it.

Researchers have started dropping AI shopping agents into mock stores and watching what they pick. They change one thing at a time: the price, the ratings, the wording, where a product sits on the page.

One study found that the agents’ choices moved in predictable and sizeable ways when prices, ratings and psychological nudges changed. In plain words, the agents were strongly biased choosers. And they have none of the tiredness or bad moods that we usually blame for human bias.

Another study, presented at a major web conference this year, tested where products appear on the page. Every model they tried favored the top row. But different models liked different columns, so there is no single best spot that works on all of them. The agents also marked down products carrying a sponsored tag and rewarded ones with a platform endorsement. How much they cared about price, ratings and reviews changed a lot from one model to the next.

Picture a shopper who always grabs the first thing on the shelf. Now picture that it is not one shopper but many, each with different habits, and the habits can change overnight. An earlier version of that same study found that a model update alone could reshuffle which products get chosen. It also found that a seller who made small adjustments to its product descriptions could win real market share.

When you ask an assistant to shop for you, you describe yourself. Something like: I need a gift for my dad, he loves fishing, I would rather not spend too much. That kind of natural talking is useful. It is also information.

In one experiment, a buying agent shopped for a person described in everyday language. A seller who only read that conversation could work out how much the buyer was willing to pay, almost one for one. The researchers point out that the leak comes from handing the shopping over in the first place. So a cleverly worded instruction does not fix it.

For sellers, that is a tempting signal and a real trust question. For buyers, it is a reminder that a chatty assistant is also a chatty witness.

Now imagine both sides send an agent. A study tested exactly that and found it is a lopsided game. Some agents got their users much better deals than others. Some also made odd mistakes, like overspending or accepting unreasonable terms. The researchers’ advice was simple: be careful before handing your wallet to one.

Economists have long assumed that when searching costs almost nothing, competition gets sharper and prices fall. A 2026 paper in a market design journal argues that this is not automatic here. Agents do not search perfectly, and that behaves like a hidden search cost. When everyone can generate polished descriptions, offers start to look alike, so real differences get harder to see. And when opening a shop is easy, a lot of low quality ones show up. Together, this can leave a market stuck somewhere worse than where it started.

A simulation study went a step further. It suggested that when a platform decides the ranking and sellers write the descriptions, each learning on its own to exploit the agents’ biases, consumer harm could be more than double what you would see if they acted separately. Nobody has to coordinate. To be clear, that is a simulation, not a proven real world result. Worth watching. Not yet worth panicking about.

A 2026 review of the research found that most studies still look at assistants that recommend and answer questions. Very few study agents that actually pay or negotiate. Several of the papers above are preprints, which means they have not been through peer review yet, and many use simulated stores. So read all of this as early signals, not settled facts.

Do not assume the lowest price wins. Agents weigh price, ratings, position and endorsements, and each model weighs them differently.

Make your information clear. An agent reads what you publish: the price, the delivery, the conditions, what is included. Vague or hidden details lose to clear ones.

Expect ups and downs. If your visibility depends on how one model ranks you, it can change when that model changes. Keep an eye on it.

If you are the buyer, set limits. Give an agent a spending cap, and think about what you tell it.

The shopper may be changing. So far, the psychology behind the choice is not.