Do people use AI assistants before or after searching?

A cross-surface panel study linking captured assistant prompts and responses to the same panelists' observed searches and pageviews finds that conversational-assistant use more often comes after prior web activity than before it, reversing the sequence assumed by "answer engine" framings of AI. The paper states its thesis directly: "Search tends to open the observed journey toward content, while assistant use more often follows prior web activity." The headline estimate is a paired within-user difference of directions — "(assistant before minus after) minus (search before minus after)" — at +20.6 percentage points [19.9, 21.3], replicated in a second, adjacent month. Assistant sessions are also more self-contained: 34.1% of assistant-containing sessions show no external web step at all, against 19.5% for search-centered sessions of the same users, a within-user gap of +13.0 points [12.5, 13.6].

The mechanism offered is a measurement discipline, not a causal claim. The authors build a "cross-surface session reconstruction" that counts prompt/response events as first-class alongside searches and pageviews, then sort each AI-containing temporal session into one of four observable positions relative to the assistant span: contained (no external step anywhere), AI-first (web only after), AI-last (web only before), and bridge/interleaved (web on both sides, or between assistant events). They name the resulting difference "recomposition" — activity distributed differently across dialogue, search, and browsing, "without implying that assistant use caused the difference." They are explicit that containment is not resolution: "timestamps alone cannot establish one task, satisfaction, or completion." Search-centered sessions also interleave with content more tightly (bridge 53.5%) than assistant-centered sessions do, which instead concentrate consecutive assistant turns.

This complicates any reading of assistant adoption as pure substitution for search, and it sits usefully beside Does chat delegation actually save time on task completion?, which found that lower interaction effort inside one hybrid interface did not translate into shorter task duration. Both papers warn against inferring resolution or efficiency from interaction counts alone: that study showed effort metrics and completion time can move independently within a single session, while this one shows that a contained, no-external-step assistant session does not by itself mean the user's task was finished. It also qualifies Can conversation structure predict dialogue success better than content?: that paper reads satisfaction out of within-dialogue structure, while this one is explicit that cross-surface position — contained, AI-first, AI-last, bridge — cannot establish satisfaction or task completion on its own.

The excerpt does not establish what the assistant turns actually accomplished, whether a contained session corresponds to a satisfied information need, or how task type shapes the before/after pattern — the authors explicitly reserve "workbench-versus-seeking, topical continuity, satisfaction, and per-task splits" for a separately governed validation pass. It also covers only standalone assistant surfaces (ChatGPT, Claude, Perplexity, Gemini's own surface) over one month in the US and Great Britain on an opt-in panel, excluding search-embedded AI (Google's AI Overviews and AI Mode) entirely. What follows at this strength: the "assistant replaces search" story is premature, and reasoning about AI's effect on information-seeking should treat dialogue, search, and browsing as one cross-surface system rather than assume a synthesized answer closes the episode.