Traditional research on political communication relies on domain-level proxies to measure ideological news exposure. This approach risks biases: it ignores within-outlet ideological heterogeneity and conflates passive exposure (clicks) with behavioral attention to news. We address these gaps using highly granular web-tracking data paired with survey responses across three General elections in Spain, Brazil, and the United States (N=950 users, 21,853 validated political visits). We classify the partisan tone of individual articles relative to each readers' ideology using automated content classification via Large Language Models (LLMs). We then analyze both content selection and real-time reading duration jointly.

Our findings challenge the generalizability of the "echo chamber" narrative and of symmetric polarization accounts. Spanish, Brazilian, and U.S. left-leaning media outlets exhibit substantial within-outlet ideological heterogeneity. Most consumption in these contexts is classified as neutral, and partisan framing receives no systematic reading-time premium. In contrast, among the U.S. right-leaning audience bloc, partisan "boosterism"—simultaneous in-party praise and out-party attack—is selected more frequently and elicits longer reading times (approximately 27\% relative to neutral content). The results suggest that strong ideological isolation and attention concentration do not occur uniformly across contexts. Instead, they appear concentrated in specific segments of the U.S. ecosystem during this electoral-window sample.