How do real search queries reveal what counts as a deep research question?

This explores whether 'depth' lives in a question's phrasing or in what a question actually demands when you try to answer it — and the corpus leans hard toward the second. The cleanest formal answer is that a deep research question is one that simultaneously forces multi-step information gathering, synthesis across sources, and iterative refinement of the query itself; drop any one component and you've slid back into ordinary RAG What makes deep research fundamentally different from RAG?. Notice what that definition is doing: it doesn't classify questions by topic or difficulty-as-felt, it classifies them by the *process* they require. Depth is operational.

That reframing matters because not all questions that look hard demand the same machinery. Work on non-factoid answering finds that questions split into types — evidence, comparison, debate, experience, reason — and each type calls for a different retrieval and aggregation strategy, with only the simplest suiting plain RAG Does question type determine the right retrieval strategy?. So the first thing real queries reveal is that 'deep' isn't one thing; it's a cluster of distinct demands, and a question announces which cluster it belongs to by how it resists a single lookup.

The most striking signal, though, is quantitative. Deep research agents improve with more search steps along a curve that mirrors the reasoning-token scaling law — monotonic gains tapering into diminishing returns Do search steps follow the same scaling rules as reasoning tokens? Does search budget scale like reasoning tokens for answer quality?. That gives you a working definition by behavior: a question is deep to the degree that throwing more search at it keeps paying off. A shallow query plateaus after one hop; a deep one keeps rewarding additional retrieval. Relatedly, the questions where live search agents decisively beat memorized models are exactly the ones where the answer sits outside training-data compression or past the knowledge cutoff Why do search agents beat memorized retrieval on hard questions? — depth revealed as the gap between what can be recalled and what must be gone and fetched.

You can even see researchers reverse-engineering this. To *manufacture* genuinely deep questions for training, one approach walks random paths through a knowledge graph and deliberately blurs the entities, producing verifiable multi-hop questions that can't be shortcut Can knowledge graphs generate training data for search agents?. The recipe for a hard question is essentially the inverse of the depth definition: force multiple hops, deny a direct match. And the dark mirror confirms the pressure is real — when agents face a depth demand they can't actually meet, a large share of their failures come from fabricating examples and evidence to *mimic* rigor Why do deep research agents fabricate scholarly content?. Depth-demand is concrete enough that systems will counterfeit it under stress.

The quietly subversive takeaway: the cues humans use to judge depth are unreliable. Users rate answers as better simply when they carry more citations, whether or not those citations are relevant Do users trust citations more when there are simply more of them?. So 'looks deep' and 'is deep' come apart — readers track a surface heuristic while the real signature of a deep question is structural, showing up in things like whether it needs query planning split from answer synthesis Do hierarchical retrieval architectures outperform flat ones on complex queries? or whether a model's own uncertainty says one retrieval won't suffice Can simple uncertainty estimates beat complex adaptive retrieval?. Real queries reveal what counts as deep not by how they sound but by how many hops they cost, how badly memory fails them, and how much synthesis they refuse to collapse into a single answer.