November 9, 2026

Search Sensitivity vs. Search Precision: Building a Defensible Literature Search

Every literature search sits on a trade-off. Push for sensitivity and you retrieve more of what exists, including material you will discard. Push for precision and most of what you retrieve is relevant, but you risk miss

Written byWisPaper TeamAI Research Workflow Team
Editorial illustration for Search Sensitivity vs. Search Precision: Building a Defensible Literature Search

Every literature search sits on a trade-off. Push for sensitivity and you retrieve more of what exists, including material you will discard. Push for precision and most of what you retrieve is relevant, but you risk missing studies that should have shaped your argument. The practical question is not which value is "better" but which one your current task requires, and how you will document the choice so a coauthor, reviewer, or future lab member can retrace it.

This guide diagnoses where you are starting from, walks through the decisions that define a search, and shows how to tell when the result is ready. It uses one running example: a manuscript revision in which a reviewer asks whether the review missed relevant work.

Sensitivity is the share of relevant records your search actually retrieves. Precision is the share of retrieved records that are relevant. The two pull against each other: broadening a query to catch more relevant work usually brings in more irrelevant work, and tightening a query to raise the relevant share usually drops some relevant records.

Cochrane's guidance on searching for and selecting studies frames the search as something planned from eligibility criteria rather than improvised. That framing matters here, because sensitivity and precision are only meaningful relative to a defined target: what counts as relevant for this review, in this manuscript, at this stage.

A useful way to hold the distinction:

  • Sensitivity-oriented search — built to reduce the chance that eligible work is left out. Appropriate when you are scoping a field, responding to a completeness challenge, or preparing a review that must account for the literature broadly.
  • Precision-oriented search — built to return a high proportion of on-target records. Appropriate when you already know the field and need specific evidence, a particular study, or a narrow methodological point.

Neither is a permanent setting. The same researcher may need both within one project.

Most search problems are actually specification problems. Before adjusting terms, identify which situation you are in.

You are revising a manuscript under review. A reviewer has questioned coverage. Your task is sensitivity-leaning: demonstrate that you looked widely, and document what you did. The artifact you need is a search record a coauthor can inspect.

You are closing a project or handing off to the lab. Your task is reproducibility. Someone else must be able to rerun the search and reach the same candidate set. The artifact is a documented method trail, not just a reference list.

You are starting a new review. Your task is staged: broad discovery first, then narrowing against eligibility criteria. Cochrane's process supports this shape — plan from eligibility criteria, use multiple sources, combine free-text terms with subject headings, document the process, deduplicate, and screen in stages.

You are chasing one known paper or author. Your task is precision. You want the specific record, its versions, and its citation trail, not a field-wide sweep.

Write down which of these you are in. It determines every later choice.

What decisions and artifacts does a defensible search require?

A search that can survive scrutiny is built from explicit decisions, each producing something you keep.

Decisions:

  1. Which eligibility criteria define relevance.
  2. Which sources to search.
  3. How to combine free-text terms with subject headings or field tags.
  4. Whether to run one broad pass or staged passes.
  5. How to handle duplicate records and multiple reports of the same study.
  6. Whether to rerun the search before submission.

Artifacts:

  • The eligibility criteria, written down.
  • The exact queries used, per source.
  • The date each search was run.
  • The raw result set before screening.
  • A screening log showing what was excluded and why.
  • A deduplicated, linked set of included records.

Cochrane's chapter explicitly calls for documenting the process, deduplicating records, screening in stages, linking multiple reports of one study, and rerunning searches before publication. Each of those is an artifact, not just a habit.

A sensitivity-first search is designed so that the cost of a miss is treated as higher than the cost of extra screening.

Start from eligibility criteria and translate them into concepts. For each concept, combine free-text terms with subject headings where the source supports them. Cochrane recommends combining free-text terms with subject headings and searching multiple sources rather than relying on one.

In practice, for the manuscript-revision example:

  1. List the concepts your review actually covers.
  2. For each concept, gather synonyms, spelling variants, and controlled vocabulary.
  3. Run the combined query in more than one source.
  4. Keep the raw result set before you filter anything.
  5. Record the query and the run date.

PubMed's documentation is useful here because it exposes how a query was interpreted. Field tags, phrase searching, filters, and Advanced Search let you construct a deliberate query, and the search details show how terms were translated and combined. That translation view is itself an artifact: it tells a coauthor what the database actually did with your terms.

Where does this fail? Sensitivity-first searches produce large candidate sets that require real screening effort. If you run one without a screening plan, the extra recall becomes unmanageable noise. The failure mode is not too many results; it is too many results with no log.

A precision-first search assumes you already know roughly what you need and want the retrieved set to be mostly on target.

This is where source-specific tools earn their place. Google Scholar documents searching by author and title, date sorting and limits, Cited by and Related articles, citation export, and locating alternative versions of a work. Those features suit a precision task: find a known paper, follow its citation trail, or check whether a newer version exists.

PubMed supports citation lookup and Similar Articles, which serve the same precision purpose — starting from a record you trust and moving outward deliberately.

A precision-first workflow for the same manuscript:

  1. Identify the anchor papers your argument depends on.
  2. Use author and title search to locate each one precisely.
  3. Follow Cited by and Related articles to find adjacent work.
  4. Check for alternative versions so you cite the right one.
  5. Export citations into your reference manager.

Where does this fail? Precision-first searching inherits your current picture of the field. If your anchor papers are incomplete, the citation trail will be too. It is a poor tool for answering "what did we miss?" — which is exactly the reviewer's question.

Where does each approach break down on the same task?

Run both against the reviewer's challenge and the failure modes separate cleanly.

CriterionSensitivity-firstPrecision-first
Primary goalAvoid missing eligible workReturn mostly relevant records
Best starting pointEligibility criteria and concept listKnown anchor papers
Source strategyMultiple sourcesTargeted lookups and citation trails
Query styleFree-text plus subject headings, broadAuthor, title, citation links
Screening burdenHighLow
Main failureUnmanageable volume without a logInherits gaps in your existing picture
Artifact producedRaw result set plus query recordCitation trail from trusted records
Fits reviewer challengeYesPartially
Fits lab handoffYes, if documentedOnly if the trail is recorded

On the revision task, precision-first alone cannot answer the reviewer, because it cannot show coverage beyond what you already knew. Sensitivity-first alone can answer the reviewer but will bury you unless you screen in stages and keep a log.

What does a combined workflow look like?

The two approaches are sequential, not competing. Use sensitivity to establish coverage, precision to deepen the records that matter.

Stage 1 — Specify. Write eligibility criteria and the concept list. This is the artifact everything else depends on.

Stage 2 — Sweep broadly. Run the sensitivity-first query across multiple sources. Keep the raw set. Record queries and dates. If you are working inside a project workspace, this is a natural point to set up the review in Research Projects so the search stage and later stages stay together.

Stage 3 — Screen in stages. Deduplicate, then screen titles and abstracts, then full text. Log exclusions. Link multiple reports of the same study, as Cochrane recommends.

Stage 4 — Deepen precisely. For included records, run precision moves: citation trails, related articles, alternative versions. Use WisPaper search to locate specific papers you need to pull in, and save the ones you keep into My Library so the included set is a maintained collection rather than a folder of downloads.

Stage 5 — Verify references. Before submission, check that every citation in the manuscript points to a real record. TrueCite screens BibTeX against academic databases to flag hallucinated references — a concrete check at the point where reference lists are finalized.

Stage 6 — Rerun and record. Cochrane recommends rerunning searches before publication. Rerun, note the date, and add anything new through the same screening log.

How do you keep an auditable method trail?

An auditable trail is one a coauthor can follow without asking you questions. Build it as you go, not afterward.

Keep, at minimum:

  • Eligibility criteria, versioned.
  • Each query, per source, with the run date.
  • The raw result count before screening.
  • The screening log with exclusion reasons.
  • The included set, deduplicated and linked.
  • The rerun record before submission.

Two habits make this practical. First, treat the query itself as a deliverable — PubMed's search details view, which shows how terms were translated and combined, is a good model for what to preserve. Second, keep the included set in one maintained place rather than scattered across downloads, so the trail and the collection stay in sync.

When you need to interrogate the included set — for example, to check what the collected papers say about a specific mechanism a reviewer raised — Library QA lets you ask questions based on papers uploaded to My Library or added from search results. That keeps the question grounded in the set you actually documented.

How do you know the search is ready?

Readiness is a checklist, not a feeling.

  • Eligibility criteria are written and unchanged since the search was run, or changes are logged.
  • More than one source was searched.
  • Free-text terms and subject headings were combined where the source supports it.
  • Queries and run dates are recorded.
  • Records were deduplicated.
  • Screening happened in stages with logged exclusions.
  • Multiple reports of the same study are linked.
  • The search was rerun before submission.
  • Every manuscript citation resolves to a real record.

If any line is blank, the search is not ready — regardless of how confident you feel about coverage.

When should you hand the search to a longer-running agent?

Some search tasks outgrow a single session: a full literature review, a reproduction attempt, or a broad sweep that needs to be organized before screening. Scholar Agent is built for longer research tasks in the literature review, inspiration, and reproduction space. Handing off at that point makes sense once your eligibility criteria and concept list exist, because the agent needs the specification you built in Stage 1.

For a new project where the search is meant to open up directions rather than close them, Idea Discovery takes research topics, background, references, and experimental constraints as inputs. That is a different job from verifying coverage, and it belongs before the sensitivity sweep, not instead of it.

If you are new to the platform, the WisPaper homepage introduces it as an AI research assistant for finding, screening, organizing, and working with academic papers — the four activities this workflow has moved through.

FAQs

Sensitivity-first is the right default when the cost of a miss is high and you have a screening plan to absorb the volume. But sensitivity without documentation does not produce a defensible review. Cochrane's guidance ties the search to eligibility criteria, multiple sources, staged screening, deduplication, and a rerun before publication — the recall only counts if the process is recorded.