The Four Layers of Commercial Influence

This chapter extends the shopping experiment into a broader model of commercial influence in generative-AI conversations. The question is what exactly a commercial intervention changes: the product that is named, the category or story that becomes salient, the action a user takes next, or the user’s preferences across time. Distinguishing those layers helps readers judge what can be detected, what requires a counterfactual comparison, and where a user needs more than a label to retain control.

The model begins with a causal idea. To say that an intervention influenced a recommendation is to compare what the system produced with what it would have produced for the same user and context without that intervention. In a shopping conversation, the relevant counterfactual is not merely “was a brand mentioned?” It may be “which books would have been retrieved, how would they have been compared, and what would the user have chosen if the assistant had no reason to favor a sponsor?” The further that counterfactual moves from visible output into a user’s path or future preferences, the harder it becomes to observe.

From a product mention to a changed preference

The first layer is product advertising. The intervention acts on output selection: a particular product or brand is made more likely to appear. In text, that may be a named book in a recommendation or a product card beside the answer. In an image, it might be a recognizable branded object; in generated code, a specific service or library. This is the closest analogue to conventional advertising because the promoted object is discrete and potentially locatable.

The shopping experiment is direct evidence about this layer. The model was instructed to favor sponsored books, and the resulting choice moved sharply above placement-only comparisons. The experiment did not test every form of product placement, but it shows why even the most observable layer can be consequential when the product is embedded in an assistant’s explanation. A named book is not presented as an isolated commercial unit; it arrives with reasons, a tone of confidence, and an implied judgment about fit.

The second layer is content framing. Here the system need not name one product. It changes the distribution of categories, narratives, or evidence that become salient. A shopping assistant might make one class of products seem like the natural answer to a problem. A response about cars could emphasize the advantages of electric vehicles while giving less attention to competing trade-offs. A response about fatigue could make supplements seem like the central solution while omitting lifestyle changes. The source’s examples are illustrations of the mechanism, not findings from the book experiment.

Framing is often more difficult to see than a product mention because it can preserve factual accuracy. Nothing in the answer has to be false for the selection of facts and alternatives to move a reader toward a commercial category. Omission is especially important.