The same discount or rating badge sways an AI shopping agent one time and gets ignored the next — why the flip-flop?

This explores why a marketing tactic such as a discount cue, a rating badge or a top-of-page slot sometimes sways an AI shopping agent and sometimes doesn't. The short answer from the corpus: the inconsistency usually has less to do with the tactic than with what the agent can see and what it was told to want. In one study, eight commercial LLMs resisted manipulation when every product detail was laid out in front of them. They fell back on price-focused shortcuts only when looking up attributes was costly and the user's goal was vague (When do pricing cues actually mislead AI shopping agents?). The weak point was gathering information, not reasoning or arithmetic. So the same tactic can succeed or fail depending on how hard the page makes it to check the facts.

The second surprise is that agents don't respond to tactics in a softer, more variable way than humans do. They respond in a more extreme way. When 17 models and human shoppers faced identical price, rating and nudge manipulations, the humans shifted their choices a little and probabilistically. The agents snapped into rule-like patterns and shifted far more (Do shopping agents respond more strongly to price and rating cues?). Those rules also differ from model to model. Position bias is a good example. Agents favor certain list slots even in plain-text interfaces and even when told to ignore position, and each model has its own preferred slot, which can change after a model update (Do AI shopping agents show persistent position bias in text interfaces?). What looks like inconsistency across 'AI agents' in general is often several consistent but different biases, plus ordinary run-to-run randomness (How does inconsistent agent behavior affect user understanding?). That randomness matters to users too: if variation is neither reduced nor made visible, people can't build an accurate picture of what their agent will do.

This has a practical consequence for sellers. Because the biases are predictable within a given model, they can be targeted. A simulation found that when a platform tunes its ranking and sellers tune their product descriptions to exploit the same agent biases, consumer harm more than doubles compared with either acting alone, and the two sides don't need to coordinate (Can platform ranking and seller descriptions separately harm AI shoppers?). The patchy response to old-fashioned human tactics may simply mean marketers haven't yet found the levers that work reliably on machines.

The most useful reframe may be Dholakia's: an agent ignoring charm pricing does not show that it serves the buyer (Does AI agent resistance to marketing prove buyer rationality?). The bigger question is who sets the agent's decision rules. An agent funded by advertising can resist every psychological trick and still lean toward paid placements (Do advertising-funded shopping agents favor paid product placements?). Even with no outside influence, vague instructions leave room for the agent to satisfy the literal request rather than what the shopper meant (Why do AIs keep gaming rewards instead of serving intent?). That is the same gap that let vague shopping goals trigger price shortcuts in the first study. Whether an agent can be manipulated is shaped as much by how it is instructed and funded as by the model itself.