What makes a narrative engaging? Sentiment arcs — the emotional trajectory of a story — have been widely studied. Topic progressions capture thematic structure. But information density — how much each paragraph departs from what came before — has not been systematically measured across a large corpus.
Bamman and Piper (arXiv:2602.20647) measure semantic novelty in 28,606 pre-1920 English books. Each paragraph's embedding is compared to the accumulated centroid of all preceding paragraphs. The resulting trajectory — the novelty curve — captures how predictable or surprising the text is at each point. Eight narrative archetypes emerge from clustering these trajectories, from steep descent (rapid convergence to a stable topic) to steep ascent (escalating unpredictability).
The strongest predictor of reader engagement is not the shape of the novelty curve but its volume — the variance. Books that oscillate between high and low novelty, that repeatedly surprise and then settle, that maintain information-density dynamics, are more engaging than books with either consistently high or consistently low novelty. The variance, not the mean, predicts engagement.
Circuitousness — how much the trajectory wanders — has a strong raw correlation with engagement (rho = 0.41) but is 93% confounded with book length. Control for length and the correlation drops to 0.11. The raw correlation was measuring that longer books tend to be more popular, not that wandering narratives are better.
Genre constrains the possible shapes: fiction maintains plateau profiles; nonfiction front-loads information. Historical analysis shows books became progressively more predictable between 1840 and 1910.
The general observation: engagement comes from oscillation, not from sustained novelty or sustained familiarity. The reader wants surprise AND resolution, repeatedly. The volume of the signal — its dynamic range — matters more than its direction or destination.