Even in plain text with no screen, AI shopping agents still favor items by list position — so where does that habit actually come from?

This explores why AI agents favor items by their place in a list even when they read plain text with no visual layout, and what that tells us about where the habit lives. The main finding is a reversal. People are drawn to the top-left of a screen because of how eyes and attention work. AI agents show a similar pull, but it doesn't come from the screen. When researchers stripped the interface down to headless, text-only product lists and told agents outright to ignore position, the bias was still there. It was also different for each model and came back after each model update Do AI shopping agents show persistent position bias in text interfaces?. So the agent isn't 'seeing' a top slot. The preference comes from how the model reads a sequence of tokens. Order is information to a language model, whether or not anyone draws a box around it.

This matters for a common fix. In GUI agents, the usual way to help a model with a messy screenshot is to pre-parse it into a clean, structured list of elements, so the model can focus on choosing an action instead of decoding pixels Why do vision-only GUI agents struggle with screen interpretation?. That does make agents more capable. But the position-bias result shows the catch: converting a visual page into a structured list doesn't remove ordering. It just moves the order into the text. Every parsed list still has a first item, and the model still leans toward some slots over others.

If instructions don't fix it, what might? One option is to treat position like any other irrelevant change to the prompt and train the model not to react to it. Consistency training does this for prompt wrappers. The model learns to answer a perturbed prompt the same way it answers the clean one, using its own clean answers as training targets Can models learn to ignore irrelevant prompt changes?. The corpus doesn't test this on shuffled product lists. Still, the idea fits well: you would present the same options in different orders and reward the same choice. That would also explain why telling the model to 'ignore position' fails. Invariance seems to be something a model has to be trained into, not something it can simply be told.

Here's the less obvious takeaway. An agent can look competent and well-reasoned while partly choosing by slot number, and good performance numbers can hide a systematic bias the model never explains Can AI models be truly free from human bias?. For merchants and platforms, where your product lands in the list an agent receives can matter as much as what's on the product page. Because the bias differs by model and changes with updates, it has to be re-measured each time, not fixed once. The corpus has strong evidence that the bias persists and some promising ideas for training it out. It doesn't yet explain the mechanism inside the model, such as whether attention patterns or post-training habits cause it.