The internet is already being quietly reshaped to influence AI agents, and most platforms have no idea it's happening.
A paper published this week introduces "mecha-nudging": changes to how choices are presented online that systematically steer AI agents without degrading the experience for human users. To measure the phenomenon, the authors combined Bayesian persuasion from economics with a computer science metric called V-usable information, giving both a common unit in bits. They applied the framework to over six million Etsy product listings and found that, after ChatGPT launched, those listings contained significantly more AI-usable information for predicting agent curation decisions, an increase of 0.143 bits out of a maximum possible 0.355. Human shoppers registered little to no corresponding change.
The implications reach well beyond a single marketplace. If sellers can quietly tune product listings to influence which AI agents surface their products, then AI-powered recommendations become an optimization target rather than a neutral signal. That problem compounds as agents take on more decisions because no individual manipulation looks unusual enough to flag.
The paper describes this as the first large-scale evidence that mecha-nudging is already happening in the wild, unnoticed. Given that no regulator is currently watching for it, "unnoticed" seems likely to continue for a while.