At Trip, we’ve spent thirty years watching people search for clinical evidence. AskTrip lets us watch something we could never really see before: what people ask next.

When someone reads an answer and then comes back with a second question – a refinement – they are telling us how their thinking has moved. This has been made much easier, for the user, via our Explore further feature which was released during our recent major upgrade:

We pulled together a set of these refinement pairs and looked not mainly at their topics, but at their direction of travel. In less than a week, users generated more than 50 such refinements – enough to start seeing patterns in how questions evolve. Two patterns dominate.

Pattern one: from knowledge to action

The most common move is from understanding something to doing something about it. The opening question asks what a thing is, what causes it, or whether it works. The follow-up asks what to do, how to monitor it, or what to choose.

A nurse asks about the perceptions and contributing factors behind medication administration errors, then asks which training programmes actually reduce them. Someone asks what Bartter’s and Gitelman’s syndromes are, then asks about long-term management. A question about whether cladribine retreatment is safe in multiple sclerosis is followed by one about the monitoring protocols to use during it. A question about the weekend effect on organ procurement becomes a question about interventions that reduce weekend discard rates.

The follow-up rarely becomes more theoretical. Once the descriptive need is met, the pull is usually towards clinical application, the last mile of turning evidence into a decision. This maps neatly onto a common gap in evidence tools: they often answer the descriptive question better than the applied one.

Pattern two: from general to specific

The second common move is to sharpen a broad question by adding a constraint – a population, a comorbidity, a comparator, a subtype, or a more specific outcome.

“What is the best treatment for hypertension?” becomes “What is the best treatment for hypertension in patients with chronic kidney disease stage 4?” A broad question about dietary changes for weight loss narrows to whole-food plant-based versus omnivore diets, or to the effect of meal timing. A question about formula changes and infant growth narrows to prebiotics in sick preterm infants. A question about chair alarms for falls narrows to self-releasing chair-alarm belts specifically.

This is essentially the user reshaping a broad query into something closer to a well-formed PICO after seeing the first answer. Occasionally the narrowing is methodological rather than clinical: one user narrowed by evidence tier, asking which quality-of-life measures are highlighted specifically in systematic reviews.

Answer-induced follow-ups

Some refinements look different. Rather than simply narrowing a question the user already had, they pursue a concept that the first answer would plausibly have surfaced. We cannot prove that the concept was not already in the user’s mind without the answer text and a behaviour trace, but the pattern is suggestive.

A comparison of ivermectin versus permethrin for head lice is followed by a question about the prevalence of permethrin resistance – likely one reason the comparison matters. A question about folic acid flour fortification is followed by one about B12-deficiency risk groups, the classic masking harm. A question about whether LLM chatbots in the electronic medical record help clinicians pivots to how clinicians can improve their trust in the outputs.

These matter because they are the organic version of what a “suggest a follow-up” feature is trying to support. They show which answer-embedded concepts users spontaneously find worth chasing. That gives us a direct empirical rationale for building this support, rather than guessing at it.

Two things worth separating out

A related subset probes the evidence itself rather than the clinical content. Users ask whether there have been trials beyond the recommended stroke treatment windows; whether there is any direct evidence on cognitive recovery in brain injury with pre-existing ADHD; whether vitamin D injections deliver clinically relevant benefits; or, in one case, for supporting quotations on the timing of intravenous iron.

That last example is a request for provenance: a user wanting to see the source text, not just a synthesised answer. It is a small group here, but it speaks directly to the need for a show-your-working transparency layer.

Finally, a handful of questions are not clinical in the usual sense. They are about research methodology: predictors of PRISMA 2020 reporting completeness and the effect of AMSTAR 2; quality-of-life measures across reviews; how to structure a conference case presentation. These seem to come from an evidence-synthesis audience rather than from the clinician-facing applied questions. Their refinements deepen the methods question rather than moving towards clinical application. They should probably be analysed as a separate segment; mixing them with applied clinical queries risks blurring both signals.

Why this matters

The refinement pairs may be more informative than the initial questions alone because they capture the natural history of an information need. Initial questions tend to start broad, what something is, whether it works, what the evidence says. Refinements then move in two dominant directions: from knowledge to action, and from general to specific. A smaller but important subset appears answer-induced, where the first answer surfaces a concept the user then pursues.

That has a practical consequence. A follow-up is not necessarily a sign that the first answer failed. Often the answer has done its job: it has acted as a scaffold that helps the user find the sharper question they could not quite articulate at the start.

Refinements are therefore a useful signal not just for what users ask, but for how their uncertainty evolves after they receive an evidence summary. That is the behaviour we want to design around.