Conversational AI reshapes search: short queries collapse, long ones scaffold

A new study from Iannelli and Ai, published on arXiv, examines how conversational AI changes the structure of information seeking. By linking real conversations with major AI assistants (e.g., ChatGPT, Claude) to users' subsequent search and browsing activity via an opt-in cross-surface panel, the researchers reconstructed entire episodes rather than isolated queries. Their core finding: AI does not uniformly shorten the seeking process—it bifurcates episodes into two distinct shapes.

Most AI-assisted episodes collapse in place: users receive an answer and take no onward search or content step. This collapse is heavily tied to query length—72% of queries with 1–3 words terminate immediately, dropping to 48% for queries longer than 20 words. Surprisingly, task type (lookup vs. learning vs. comparison) had no significant effect on collapse rate. Additionally, roughly 40% of assistant interactions are workbench use (drafting, coding, editing) rather than information seeking, and these collapse at the highest rate.

Crucially, conversational AI does not kill search. Search remains woven through roughly three-quarters of within-episode transitions—after reading a page, users return to the search box over the assistant by a 70/30 margin. Verification is exceedingly rare: explicit verification language follows only about 1% of episodes, and citation-forward interfaces don't measurably increase checking. The study concludes that AI recomposes seeking by answering brief asks in place while anchoring invested asks in longer, scaffolded journeys—adding a layer rather than replacing search.