What's this about?
New research from Mollick's group asks whether you can consistently predict — or, through marketing, influence — what an AI agent chooses when shopping. The answer is no. Small differences like the order pages were viewed in, or what happened to be in the agent's memory, changed its preferences in unpredictable ways.
Why would anyone expect an agent to be predictable?
Because the entire discipline of marketing assumes preferences are influenceable in stable ways: placement, framing and social proof move human choices reliably enough to build an industry on. The hope, or fear, was that agents would be even more manipulable — a deterministic system you could reverse-engineer. What the research finds is worse for everyone: not resistant to influence, just *inconsistent* under it, which means neither shoppers nor sellers can rely on the outcome.
What's the community saying?
- The practical implication for retailers is that optimising a product page for agents is currently not a stable strategy.
- The implication for consumers is less comfortable: an agent that shops inconsistently is an agent whose choices you cannot audit or reproduce.
- It sits neatly beside AnTrap in today's digest — both are findings that agent behaviour is far more sensitive to incidental environment state than anyone would like.