01 · The Question
Are you still asking the question you started with because it remains important, or because you already became attached to it?
Researchers often approach the literature with a question already forming in their minds.
That is useful. A preliminary question gives the search direction. It tells you which concepts, populations, theories, and evidence to investigate.
But the question you begin with is based on what you knew before conducting the review. Once you understand the literature, you may discover that the question is too broad, conceptually confused, methodologically unanswerable, already well answered, or simply less important than another question revealed by the evidence.
If none of those discoveries are permitted to alter the research question, the literature review becomes strangely ceremonial. You searched extensively, learned considerably, and then asked exactly what you would have asked before doing any of it.
The important question is therefore: after understanding the evidence, is your original research question still the question that most deserves to be asked?
03 · What You Need to Know
How can the literature reveal that you are asking the wrong question?
Your first question is necessarily based on incomplete knowledge
Before reviewing the literature, you do not yet know the evidence base well. That is partly why you are reviewing it.
Your initial question may therefore contain assumptions about how constructs should be defined, which variables matter, whether an effect exists, which population is understudied, or what kind of evidence is missing.
As those assumptions change, the question built from them may need to change too.
This follows naturally from asking whether the literature changed any of the assumptions you began with. A revised assumption that leaves every downstream research decision untouched deserves another look.
The literature may reveal that your question is too broad
Suppose you begin with:
“How does generative AI affect student learning?”
The literature quickly reveals that “generative AI” includes substantially different uses, while “learning” could refer to immediate performance, retention, transfer, metacognition, conceptual understanding, or numerous other outcomes.
The original question may be useful for initial exploration but too broad to support one coherent empirical study.
A more mature question might ask whether a specific form of AI-generated formative feedback affects students' later unaided writing performance.
The topic has not necessarily changed. The question has become answerable.
The literature may show that two concepts you combined need separating
Researchers frequently begin with everyday terminology that the literature later reveals to contain distinct constructs.
“AI use” may differ from AI reliance. Reliance may differ from uncritical acceptance. Engagement may differ from participation. Achievement may differ from learning. Confidence may differ from competence.
If your question treats distinct constructs as interchangeable, additional data will not repair the conceptual problem.
Broad topic
“Generative AI and critical thinking.”
Researchable question
Specifies which form of AI use, which aspect or measure of critical thinking, which population, and which relationship or effect is actually being investigated.
The question may already have a good answer
This is one of the most consequential outcomes of literature review.
Perhaps you planned to investigate whether an intervention improves an outcome, only to discover several rigorous recent studies and a high-quality synthesis already addressing essentially the same comparison.
You could still conduct another study. The stronger question is why you should.
If the literature already answers the consequential question adequately, your original question may no longer represent a meaningful evidence gap.
The next step is to determine whether your proposed study would be redundant.
The literature may reveal that the real uncertainty is somewhere else
Suppose studies consistently show that an intervention improves immediate performance. Your original question asks whether the intervention works.
After reviewing the evidence, that is no longer the interesting uncertainty.
Perhaps almost no study measures retention six months later. Perhaps effects differ according to baseline ability. Perhaps the intervention works under researcher supervision but routine implementation remains poorly studied.
The question can then move from “Does it work?” toward “Does the effect persist?”, “For whom does it work?”, or “Does it work under ordinary implementation conditions?”
A mature research question begins where the evidence becomes uncertain, not where the topic begins.
Your causal question may outrun the available design
You may want to ask whether X causes Y but only have access to a cross-sectional survey.
The literature may show that dozens of similar surveys already document the association while leaving causality unresolved.
You now face an important choice. Change the design so that it addresses the causal question more credibly, or change the question to one the feasible design can actually answer.
Watch Out
Do not preserve a causal research question while quietly downgrading the study to a design that cannot answer it, then hope regression coefficients will negotiate the difference.
The literature may change which population matters
Your original plan may focus on a population because it is accessible or because no study appears to have examined that exact group.
After reviewing the evidence, you may discover that the more consequential population is one for which generalizability is genuinely uncertain.
Alternatively, you may discover that your chosen population differs in no theoretically or empirically meaningful way from populations already studied extensively.
The question should follow the unresolved applicability problem rather than convenience alone.
The outcome may need to change
Perhaps most existing studies measure self-reported confidence while the important unresolved question concerns actual performance. Or perhaps immediate performance is well established while transfer remains unknown.
Changing the outcome can transform the contribution of the study without changing its general topic.
| What the literature reveals |
Possible change to the question |
| The original question is too broad |
Specify the relevant population, exposure or intervention, outcome, and context. |
| Key concepts are being conflated |
Separate constructs and ask which relationship actually matters. |
| The main effect is already established |
Shift toward durability, mechanism, implementation, harms, or meaningful boundary conditions. |
| Existing evidence is mainly correlational |
Change the design for causal inference or narrow the question to association. |
| One population dominates the literature |
Investigate a population that meaningfully tests generalizability. |
| Existing outcomes are poor proxies |
Ask the question using a more defensible outcome or measurement approach. |
| Studies disagree systematically |
Ask what moderator, condition, or mechanism explains the disagreement. |
| The proposed study is redundant |
Redirect the question toward a consequential unresolved problem. |
Disagreement in the literature can generate a better question
Conflicting studies are often treated as a nuisance that prevents a clean literature review conclusion.
They can instead reveal the research question.
If credible studies produce different effects under different conditions, the next useful question may concern what explains that variation.
After investigating why important studies disagree, you may discover that the original average-effect question is less useful than a question about effect modification or boundary conditions.
Strong evidence can turn an exploratory question into a mechanism question
Suppose the literature convincingly establishes that a phenomenon occurs.
Continuing to ask whether it occurs may add little. The more informative question may concern why.
But mechanism questions require their own evidence. Do not simply convert an established effect into an assumed mechanism. Ask what competing explanations exist and what study could distinguish among them.
A research question can become narrower without becoming less important
Researchers sometimes worry that refining a question makes the project seem smaller.
Often the opposite is true.
“Does technology affect learning?” sounds ambitious but is nearly impossible to answer coherently. “Does structured AI-generated formative feedback improve delayed unaided revision among novice academic writers?” is narrower but carries a much clearer empirical meaning.
Scope and significance are not synonyms.
The literature can also broaden a question
Revision does not always mean narrowing.
You may begin with one narrow intervention and discover that it belongs to a broader class of phenomena governed by a common mechanism. You may find that apparently separate literatures address the same conceptual problem under different terminology.
In such cases, a broader theoretical question may become more valuable than the initially local one.
Broadening should still follow evidence and conceptual coherence rather than ambition for its own sake.
Do not change the question every time you read a surprising paper
Research questions need stability as well as responsiveness.
One weak or unusual study should not force a redesign. Revision should follow the weight of evidence and the importance of the issue uncovered.
This is why you need to explain why some evidence deserves more weight than other evidence before allowing it to redirect the project.
Changes become harder once data collection begins
Question refinement is easiest and least problematic during planning.
Once hypotheses, outcomes, eligibility criteria, or analyses have been preregistered or data have been examined, changing the question can create risks of outcome switching, selective reporting, or post hoc hypothesis construction.
Exploratory changes can still be legitimate, but they should be distinguished transparently from confirmatory questions specified earlier.
This is one reason the literature review should do serious intellectual work before the data begin answering back.
The revised question should be traceable to the evidence
You should be able to explain why the final question differs from the initial one.
Original question
What did you initially want to know?
Evidence discovered
What did the literature establish that you had not known?
Remaining uncertainty
What important question survived the review?
Research consequence
What question now follows from that uncertainty?
This chain gives the research question an evidential rationale rather than merely a personal origin story.