The Price of Attention: Behavioural and EEG Responses to Advertising in LLM Conversations

When advertising enters a conversational assistant, its cost is paid by the user: in attention, trust, and overall experience. The central question is not simply whether advertising can be inserted, but when its cost is incurred, how large it is, and whether the way an advertisement enters the conversation changes that cost. This thesis therefore treats advertising not as an isolated interface feature, but as an intervention in the interaction itself, measuring how its presentation and timing affect the user, so that the cost any serving policy imposes on the user can be measured before it is accepted. It contributes a theory and a controlled experiment. The theory pairs a taxonomy of advertising interventions with an intent-trajectory formalism for defining annotation-free impact measures for advertising. In the experiment, N = 54 participants completed five four-turn tasks with a purpose-built retrieval-based assistant: an implicit mention and an explicit banner, each presented early (turn 2) or late (turn 4), plus an advertisement-free control. A subset of n = 18 also completed the experiment with synchronised 32-channel EEG. Any advertisement raised perceived manipulation by +1.27 on a seven-point scale, with stronger effects for the banner than for the mention and for early than for late insertion. The banner was noticed more often than the mention; about half of participants did not identify the mention as sponsored. Explicit banners were remembered far better than implicit mentions (81–87% versus 57–59%), and more than half of those who initially failed to notice the mention recognised it when shown again: a cost the user cannot name while it happens. Early insertion reduced credibility by about one third of a point relative to late insertion; trust did not differ on any planned contrast, and in a post hoc localisation it was lower only under the early banner; trust after re-exposure was lower for early than for late insertions. Presentation and timing did not interact on any of the four primary outcomes, and no moderation by personality or by the recorded demographics was detected. No advertisement effects were detected in the intent trajectories; the proposed theory remains to be validated as a human-free metric for advertising. In the EEG, aggregated over each condition, posterior α power was lower for early than for late insertion, a contrast that also differs in conversational depth. At early-banner onset, an exploratory transient slow-power tilt (absolute δ and θ up, relative α and β down) was observed, and none after the implicit mention (the two formats were not compared directly). The onset-locked posterior α response covaried with the change in reported trust (ρ = .80, n = 18), making it a candidate neural correlate of a trust drop to be tested in replication. The theory, platform, analysis code, and multimodal datasets are released openly. Together, they provide a foundation for measuring the user-side cost of conversational advertising and for holding serving policies to it. The price of attention is part of the cost of monetising conversation; it should be measured and minimised in the design of advertising systems.