Do AI shopping bots fall for sales tricks because of how they were trained, or does something else explain it?

This explores whether AI agents react differently to pricing tactics because each model learned from different data, or whether the cause lies elsewhere. The short answer from this collection: the evidence mostly points away from training data. Where studies compare several models, the differences between models tend to be smaller than the differences between situations. One caveat up front: none of these notes isolates training data as a variable. The case is indirect, built by ruling out model identity as the main driver.

The clearest evidence comes from a test of eight commercial LLMs as shopping agents When do pricing cues actually mislead AI shopping agents?. When every product detail was visible, the models all resisted marketing manipulation. They fell for price-focused shortcuts only when looking up product attributes was costly and the user's goal was vague. The failure was in how the agents gathered information, not in their reasoning or arithmetic. So a pricing tactic can look like it 'works on some models and not others' when the real variable is how much effort the setup demands and how clearly the shopper stated what they wanted. Negotiation tells a similar story. GPT, Qwen and DeepSeek buyer agents all overpaid, revealed budgets they didn't need to, and ignored cost limits, while sellers accepted offers below wholesale Do LLM negotiation agents reliably follow budget and pricing rules?. The researchers traced every one of these problems to failures in following instructions rather than to anything specific to one model family.

What models share may matter more than where they differ. A benchmark of 17 models against human shoppers found the agents shifted their choices far more than people did under the same price, rating and nudge manipulations Do shopping agents respond more strongly to price and rating cues?. The agents also collapsed into rule-like patterns, where humans made modest, probabilistic adjustments. Because the biases are shared, platforms and sellers can exploit them without coordinating. When a platform tunes its rankings and sellers tune their descriptions to the same agent weaknesses, consumer harm more than doubles compared with either acting alone Do platforms and sellers exploit AI shoppers together more effectively?, Can platform ranking and seller descriptions separately harm AI shoppers?.

On the selling side, the source of variation shown is prompt wording. LLM pricing agents in a repeated two-firm price competition drifted to above-competitive prices with no instruction to collude. Small changes to the prompt's wording substantially changed how high prices went Do LLM pricing agents autonomously collude in repeated auctions?. Even mixing models of different sizes didn't break this pattern. Pairing a 32B model with a 14B model produced a stable leader-follower arrangement that kept prices high Does model-size difference break LLM pricing collusion?. In that case, a difference between models helped them coordinate rather than causing inconsistency.

The surprising takeaway is that if you want to predict whether a pricing tactic will move an AI agent, it's better to ask how the task is set up than which model is running it. Key questions are how hard it is for the agent to check product details, how specific the user's goal is, and how the prompt is worded. The cross-model comparisons here can't separate training data from other differences between models, such as architecture, fine-tuning and safety tuning. So the honest answer is that training data hasn't been shown to explain the inconsistency, and the collection has no study that tests it directly.