LITERATURE SEARCH · RESEARCH DISCOVERY · DATABASE SEARCHING · KEYWORDS · CONTROLLED VOCABULARIES · BOOLEAN LOGIC · CITATION CHAINING · EVIDENCE RETRIEVAL

How Literature Search and Research Discovery Work

A literature search is not typing a topic into a box and accepting whatever appears first. It is the controlled conversion of a question into a set of retrievable concepts, sources, search expressions, result sets and follow-up routes that let a researcher find enough relevant evidence to make a defensible judgement.

The search box is only the visible surface. The real search is the reasoning that decides what the box should be asked to find.

This article owns the researcher-side discovery workflow. It is deliberately distinct from How Scholarly Indexing Works, which explains how databases ingest and organise published research, and from How Systematic Reviews and Evidence Synthesis Work, which owns the downstream synthesis process after evidence has been identified.

The central problem is simple to state and difficult to solve: the literature does not use one vocabulary, live in one database, arrive in one version, or rank itself according to your question.

The short answer

QUESTION
  → PURPOSE OF SEARCH
  → CONCEPTS
  → SYNONYMS / SPELLINGS / ACRONYMS
  → CONTROLLED VOCABULARY
  → DATABASE SELECTION
  → SEARCH LOGIC
  → TEST SEARCH
  → KNOWN-ITEM CHECK
  → RUN SEARCH
  → EXPORT RESULTS
  → DEDUPLICATE
  → SCREEN / INSPECT
  → CITATION CHAINING
  → GREY / REGISTER / REPOSITORY SEARCH
  → GAP CHECK
  → DOCUMENT SEARCH
  → SET ALERT / UPDATE SEARCH
  → RETURN TO EVIDENCE

1. Every search begins with a purpose

An undergraduate exploring a new topic, a clinician answering a bedside question, a doctoral student mapping a field and a systematic-review team all search differently.

The required precision, completeness, documentation and stopping rule depend on the decision the search is meant to support.

2. Exploratory search and systematic search are different jobs

Exploratory searching helps a researcher learn the vocabulary and structure of a field. Systematic searching aims to identify evidence according to a predefined question and method with enough transparency that another researcher can understand or repeat the process.

Trying to make the first exploratory search fully systematic can freeze discovery too early. Treating a systematic review like casual exploration can create invisible selection bias.

3. Discovery is iterative

Early searches teach you new terminology. New terminology improves later searches. Important papers reveal authors, institutions, methods and references. Those become new routes into the literature.

A strong search is therefore designed to learn from its own results without quietly changing the original research question.

4. The research question is not yet the database query

A natural-language question can contain several concepts, implied assumptions, outcomes and contextual qualifiers. Databases need those concepts converted into searchable terms.

The researcher first decomposes the question, then decides which concepts genuinely need to be represented in the search.

5. Searching every word in the question can reduce recall

Cochrane’s current searching guidance warns that it is often unnecessary—and sometimes undesirable—to search every aspect of a review question because comparators and outcomes may not be mentioned consistently in titles, abstracts or indexing.

The useful search concepts are the ones likely to exist reliably in the records you need to retrieve.

6. Question frameworks are scaffolds, not prisons

PICO, PICOS, PEO, SPIDER, PCC and other frameworks help researchers identify populations, interventions, exposures, comparisons, outcomes, contexts and study designs.

The framework helps structure thinking. The final search may use only a subset of its elements.

7. PICO is powerful for intervention questions

P = Population / Problem
I = Intervention
C = Comparator
O = Outcome

For many intervention reviews, the search concentrates on population and intervention, with a validated study-design filter where appropriate, rather than requiring every comparator and outcome term.

8. Different research questions need different concept models

A qualitative experience study, diagnostic-accuracy review, scoping review and historical literature search do not share one optimal question framework.

Choose the framework that clarifies the information need rather than the one most familiar to the researcher.

9. Concepts become term families

One concept can appear under many names.

CONCEPT: artificial intelligence in education

free-text candidates:
artificial intelligence
AI
machine learning
large language model*
LLM*
generative AI
generative artificial intelligence
intelligent tutoring system*

The exact term family depends on scope. Adding every related technology can increase noise and concept drift.

10. Synonyms matter because authors disagree about labels

Researchers may describe the same idea with formal terminology, abbreviations, older names, regional spellings or emerging jargon.

A comprehensive search anticipates meaningful vocabulary variation instead of assuming the researcher’s preferred term is universal.

11. British and American spelling can change retrieval

Behaviour/behavior, organisation/organization, paediatric/pediatric and tumour/tumor are obvious examples.

Some databases map variants automatically. Others require explicit search terms or truncation. Never assume cross-database behaviour is identical.

12. Acronyms are dangerous without context

Short abbreviations can represent many unrelated concepts. Searching AI, ML or GP without context can retrieve huge amounts of noise.

Use acronyms only when the field genuinely uses them and the database syntax can constrain them intelligently.

13. Older terminology can preserve older evidence

Concepts change names over decades. A historical literature search that uses only current terminology can systematically miss older work.

Terminology history is part of retrieval design.

14. Controlled vocabulary solves a different problem from keywords

A controlled vocabulary assigns standard subject terms to records so related literature can be found even when authors used different words.

The National Library of Medicine defines controlled vocabulary as a specified list of terms used to reduce the scattering of related subjects under different headings.

15. MeSH is the controlled vocabulary of MEDLINE

Medical Subject Headings (MeSH) is the NLM thesaurus used to describe biomedical concepts in a hierarchical structure.

Searching a MeSH term can retrieve records whose authors used different surface language for the same indexed concept.

16. Controlled vocabulary is database specific

MEDLINE uses MeSH. Embase uses Emtree. PsycINFO, CINAHL and other databases use their own subject vocabularies and indexing structures.

You cannot simply paste a MeSH-heavy search into another database and assume semantic equivalence.

17. Subject headings and free text should usually work together

Cochrane’s current guidance recommends combining subject headings such as MeSH with free-text terms for comprehensive intervention searches.

Subject headings capture indexed concepts; free text captures new language, unindexed records and terminology the vocabulary does not yet represent well.

18. Newly published records may not yet have subject indexing

Database records can enter discovery systems before all controlled-vocabulary indexing is complete.

A search relying only on subject headings can therefore miss recent material.

19. Controlled vocabularies evolve

New concepts gain headings, old terms are reorganised and hierarchical relationships change.

A search strategy preserved for future updating should be reviewed against the current vocabulary rather than rerun mechanically forever.

20. Exploding a subject heading broadens the concept

Many controlled vocabularies allow a searcher to include narrower terms beneath a broader heading in the hierarchy.

This can improve recall, but only when the narrower branches genuinely belong inside the research question.

21. Major-topic searching increases focus but can lose relevant records

Restricting a search to records where a subject is marked as a major focus can improve precision.

It may also remove studies where the concept is essential to eligibility but not the dominant topic of the article.

22. Free-text searching needs field choices

A term searched across every field behaves differently from the same term restricted to title and abstract.

Field restrictions can improve precision, but aggressive restriction can miss relevant records whose key concept appears only in indexing or other metadata.

23. Boolean OR creates a concept family

("machine learning" OR "artificial intelligence" OR "generative AI" OR "large language model*")

OR broadens retrieval by accepting any of the included terms.

24. Boolean AND combines concepts

(artificial intelligence terms)
AND
(education terms)

AND narrows the result set to records containing evidence of both concept families under the database’s search rules.

25. Boolean NOT is high risk

NOT can remove irrelevant material efficiently, but it can also remove relevant records that happen to contain the excluded term.

Use exclusion logic only when the loss mechanism is understood and tested.

26. Parentheses preserve search logic

Without explicit nesting, complex Boolean expressions can be interpreted differently from what the researcher intended.

(teacher* OR educator* OR instructor*)
AND
("large language model*" OR ChatGPT OR "generative AI")

Search logic should be readable by another human before it is trusted by a machine.

27. Proximity searching captures concepts that travel together

Proximity operators retrieve words that occur within a defined distance of one another, sometimes in any order.

This can improve precision for concepts that are too variable for exact phrases but too ambiguous as independent words.

28. Proximity syntax varies by database

One platform may use NEAR/3, another adj3, another a field-specific tilde notation.

This is one reason systematic searches must be translated rather than copied blindly between databases.

29. Phrase searching increases precision

Quotation marks can ask a database to treat multiple words as a phrase.

But phrase searching can turn off automatic mapping or synonym expansion, depending on the platform, so it can also reduce recall.

30. PubMed makes this trade-off visible

Current PubMed Help explains that quotation marks, search tags, wildcards and proximity searching can alter or disable Automatic Term Mapping.

Syntax is therefore not merely punctuation. It changes the semantic expansion performed by the search engine.

31. Truncation captures word endings

educat*
→ educate
→ educated
→ education
→ educational
→ educator
...

Truncation can improve recall while also retrieving unintended words if the stem is too short or ambiguous.

32. Wildcards can capture spelling variation

Some databases allow symbols inside words to represent optional or varying characters.

Because wildcard behaviour varies by platform, every translated search needs testing.

33. Field tags can turn broad searching into targeted retrieval

PubMed, for example, supports tags for title, title/abstract, author, affiliation, MeSH and many other fields.

Field targeting is especially useful when a term is common in references or metadata but only meaningful when it appears in the substantive article description.

34. Search details are an audit tool

PubMed’s Advanced Search exposes how a query was translated through Automatic Term Mapping.

A careful searcher checks this translation rather than assuming the database interpreted natural language exactly as intended.

35. Automatic Term Mapping is helpful and inspectable

PubMed maps untagged terms through translation tables that can include MeSH concepts, synonyms, spelling variants, journal names and authors.

Automation improves ordinary search, but formal evidence searches should verify what the automation actually did.

36. Search engines rank; systematic searches retrieve

Relevance ranking helps a reader see likely useful results first. It does not prove that lower-ranked results are irrelevant.

A systematic search usually exports the full result set rather than screening only the first page.

37. “Best Match” is a discovery aid, not a completeness guarantee

PubMed currently uses machine-learning ranking for Best Match.

This is useful for exploratory retrieval, while systematic work still depends on an explicit query and complete result handling.

38. Precision asks how much retrieved material is relevant

PRECISION = relevant records retrieved / all records retrieved

High precision reduces screening burden.

39. Recall asks how much relevant material was found

RECALL = relevant records retrieved / all relevant records that exist in the searchable universe

The denominator is usually unknowable in real research, so recall is estimated through known-item tests, supplementary discovery and methodological judgement.

40. Precision and recall trade against one another

Adding synonyms and broader subject headings tends to increase recall and noise. Adding restrictive concepts tends to increase precision and risk missed evidence.

The correct balance depends on purpose.

41. Systematic reviews usually prefer sensitivity

Cochrane states that searches should aim for high sensitivity even when precision becomes relatively low.

Screening hundreds of extra records can be less damaging than missing a pivotal eligible study.

42. Exploratory research can optimise differently

If the goal is to understand a new field quickly, the researcher may deliberately prioritise highly relevant reviews, canonical papers and current overviews first.

Completeness can come later after the field vocabulary is understood.

43. Known-item testing is one of the best search checks

If several papers are already known to be directly relevant, the draft search should retrieve them.

A missing known item reveals a vocabulary, database, indexing or syntax problem worth investigating before the search is finalised.

44. One known paper is not enough

A query can accidentally retrieve one seed paper while remaining weak for the broader literature.

Use multiple known papers representing different eras, terminology and study designs where possible.

45. Database selection is part of method

MEDLINE, Embase, PsycINFO, ERIC, Scopus, Web of Science and discipline-specific databases cover different literatures.

The researcher chooses databases according to subject, document type, geography, discipline and review purpose—not simply institutional familiarity.

46. No major database is “the literature”

Every database has collection boundaries.

A search can be technically flawless inside one database and still miss whole bodies of scholarship that database does not index.

47. Coverage bias becomes evidence bias

If an index under-represents a language, region, profession or publication type, a search limited to that index inherits the omission.

Database choice can shape the apparent answer before screening begins.

48. Subject databases provide depth

Specialist databases often provide stronger controlled vocabularies, document typing and discipline-specific coverage.

Broad citation databases provide cross-disciplinary reach. Strong searches often use both kinds of source.

49. Library discovery systems are useful gateways

University library search boxes can aggregate catalogue records, journal articles, books and repository content.

They are excellent for exploration and known-item discovery but may hide which underlying databases were searched, making them less suitable as the sole source for a reproducible systematic strategy.

50. Google Scholar is powerful and methodologically different

Google Scholar has broad scholarly coverage and strong citation-chaining utility, including repositories and document types outside some curated databases.

Its ranking and coverage are not exposed with the same reproducible query controls expected from bibliographic databases, so systematic-review methods should document its use carefully.

51. Search translation is a database-by-database task

A concept strategy developed in MEDLINE must be translated to Embase, Scopus or another platform using each database’s fields, controlled vocabulary and operators.

Literal copy-paste creates false reproducibility: the text looks identical while the database behaviour differs.

52. Keep a concept table before writing syntax

Concept	Controlled terms	Free-text terms	Notes
Population	Database subject headings	Synonyms, spellings, acronyms	Age/geography only if needed
Intervention/exposure	Subject headings	Brand, generic, historical terms	Include emerging names
Study type	Publication types	Validated filter terms	Use only where appropriate

The concept table separates intellectual design from platform syntax.

53. Search filters are pre-built methodological tools

Validated search filters can help identify study designs such as randomised trials or diagnostic studies.

Filters should be selected from reputable methodological sources and used only in databases where they were designed and tested.

54. A convenient filter can still be wrong for the question

Applying a “human” or study-design filter can remove relevant records when the database already handles those limits differently.

Cochrane specifically cautions against unnecessary filters in sources where they are not required.

55. Date limits need substantive justification

A ten-year limit may seem modern but can delete foundational or still-relevant evidence.

Use date limits when the technology, policy, disease definition or historical question creates a real boundary—not because round numbers are convenient.

56. Language limits can create bias

Restricting to English reduces screening burden but can systematically remove evidence from regions where interventions or conditions are studied differently.

If language restrictions are necessary, document and justify them as a limitation.

57. Full-text availability should not define eligibility

A record should not be excluded merely because the first search interface lacks immediate full text.

Library holdings, interlibrary loan, repositories, author manuscripts and document delivery can provide legitimate access routes.

58. Search result ranking is not an inclusion criterion

Relevance algorithms order results for convenience.

A record at rank 973 can still be eligible evidence.

59. Citation chaining finds what keyword search misses

Keyword and subject searches retrieve records through metadata. Citation searching retrieves through scholarly relationships.

The two methods see the literature from different angles.

60. Backward citation searching moves into the past

Inspect the references of an important paper to find the studies, theories and methods it built upon.

This is especially effective for finding older terminology and foundational work.

61. Forward citation searching moves into the future

Use citation indexes or scholarly search engines to find later works that cite an important paper.

Forward chaining helps identify replications, critiques, extensions and newer methods.

62. Citation chaining is strongest from good seed papers

A weak or peripheral seed can pull the search into a narrow literature neighbourhood.

Use several high-relevance seeds rather than building the whole discovery process around one famous paper.

63. Related-article functions are useful discovery tools

Databases increasingly use textual similarity, shared citations or machine-learning representations to recommend related papers.

These routes are valuable for exploration but should not become invisible substitutes for a documented systematic strategy.

64. Author searching reveals research programmes

Once a key author is identified, their other publications can reveal related datasets, methods and follow-up studies.

Persistent identifiers such as ORCID can help distinguish authors with similar names.

65. Institution searching can reveal hidden continuity

A laboratory may publish under different author combinations while maintaining the same programme of research.

Affiliation searching can reveal that continuity, although institutions themselves change names and structures over time.

66. Trial registers find research that journals may not show

Prospective trial registration creates records of planned or ongoing studies before journal publication.

Cochrane recommends trial-register searching for intervention reviews because unpublished and ongoing studies matter to publication-bias assessment.

67. Preprints expose early research states

Preprints can make current research discoverable before formal journal review.

Searchers should label publication state clearly and check for later peer-reviewed versions, corrections or withdrawals.

68. Repositories widen access and coverage

Institutional and subject repositories may contain accepted manuscripts, theses, preprints, reports, data and materials absent from commercial databases.

Repository search is especially important where open-access routes and local scholarship matter.

69. Grey literature reduces publication bias

Government reports, technical reports, dissertations, conference materials, regulatory documents and organisational research can contain evidence that never becomes a conventional journal article.

The right grey-literature sources depend on the research question.

70. Grey literature is harder to search reproducibly

Websites change, search interfaces are unstable and ranking algorithms are often opaque.

Formal reviews should document source, date, query, pages screened and any practical stopping rule.

71. Web search can be part of research discovery

Search engines are useful for identifying organisations, policy documents, terminology, experts and hard-to-index reports.

They are less suitable as the only source for a reproducible literature search because ranking and coverage are not fully transparent.

72. Website search and site-restricted web search solve different problems

An organisation’s internal search may know its own document metadata. A general search engine may index different pages and PDFs.

For difficult grey-literature sources, both can be useful.

73. Search logs turn exploration into method

A search log records which sources were searched, when, with what query, and what happened next.

SEARCH RECEIPT

Question:
Database / platform:
Coverage or collection:
Date searched:
Exact query:
Filters / limits:
Results returned:
Export format:
Known-item check:
Translation notes:
Problems / warnings:
Reviewer:
Version:

This is one of the simplest ways to make search work auditable.

74. PRISMA-S formalises search reporting

PRISMA-S is the PRISMA extension for reporting literature searches in systematic reviews. It contains sixteen reporting items designed to make search methods more transparent and reproducible.

Reporting standards do not make a poor search good, but they expose enough of the process for readers to evaluate it.

75. Search documentation is evidence, not administrative debris

Exact strategies, dates, database names and result counts are part of the methodological evidence for a review.

If the final paper cannot reconstruct how evidence entered the review, the evidence-selection process becomes opaque.

76. PRESS adds peer review to search strategy design

The PRESS guideline provides a structured method for peer reviewing electronic search strategies.

Its domains include translation of the research question, Boolean and proximity operators, subject headings, text words, spelling and syntax, and limits or filters.

77. Search strategies deserve peer review because small syntax errors can be large evidence errors

One misplaced AND, missing synonym or incorrectly translated subject heading can remove thousands of potentially relevant records.

Search peer review is quality assurance for the evidence inlet.

78. Information specialists are methodological collaborators

Cochrane recommends involving experienced librarians or information specialists from the beginning of a review protocol.

Expert search work includes vocabulary analysis, source selection, syntax translation, peer review, documentation and search updating—not merely database access.

79. Reference management begins at export

Search results should be exported with enough metadata to preserve title, authors, year, source, identifiers, abstract and database provenance where possible.

Weak exports create unnecessary identity problems later.

80. Deduplication is a record-linkage problem

The same article may appear in several databases with slightly different metadata.

Deduplication uses DOI, PMID, title, authors, year and other fields to determine whether records describe the same work.

81. Exact duplicate removal is the easy case

Two records with the same DOI can usually be merged confidently.

Records without identifiers require fuzzy matching and human review.

82. Preprint and journal article are not ordinary duplicates

They may represent different publication states of the same study.

Systematic-review workflows should link them at the study level rather than blindly delete one as a duplicate record.

83. One study can produce many reports

A clinical trial may generate a protocol, primary outcome paper, secondary analysis, safety report and follow-up publication.

Cochrane emphasises that reviews include studies, while search systems often retrieve reports about studies. Search management must eventually connect those reports to one underlying study where appropriate.

84. Screening is not searching

Searching decides which records enter the candidate set. Screening decides which candidate records meet eligibility criteria.

Mixing these stages makes it harder to distinguish a weak search from a strict eligibility rule.

85. Search quality cannot be judged only by the final included studies

A poor search may coincidentally retrieve the studies eventually included while still missing relevant evidence that nobody knows is missing.

Method quality depends on the retrieval design, not only on the final count.

86. Stopping rules need to match the search type

A systematic review generally follows a predefined source and update plan. Exploratory work can stop when additional searches produce diminishing conceptual returns.

Grey-literature and web searching especially benefit from explicit practical stop rules because the searchable universe is effectively unbounded.

87. Search saturation is not universal completeness

Finding no new papers after several citation-chaining rounds can suggest local saturation around a topic cluster.

It does not prove that another database, language or discipline contains nothing relevant.

88. Search updates prevent stale evidence

A review search captures the literature up to a date. New studies can appear before publication of the review.

Important reviews rerun or update searches close enough to publication to keep conclusions current.

89. Saved searches create reproducible update routes

Many databases allow a query to be saved and rerun.

Saved strategies reduce transcription error during updates, but the search should still be reviewed when database vocabularies or interfaces change.

90. Alerts convert search from event to monitoring system

Database and citation alerts can notify researchers when new records match a query or cite a key paper.

This is especially useful for living reviews, rapidly moving fields and long-term research programmes.

91. Alert queries may be narrower than systematic-review queries

A very high-recall systematic search can produce too much noise for weekly monitoring.

Researchers may maintain a sensitive archival strategy and a more precise alert strategy for routine surveillance.

92. Living reviews require continuous evidence operations

A living review is not simply a paper that is edited occasionally. It needs scheduled searches, screening, version control, synthesis updates and clear criteria for when conclusions change.

Search becomes an ongoing production system.

93. Discovery has a freshness problem

A database query run today may include records that were absent last month even for older publication dates because indexing and correction continue.

Search date is therefore part of the evidence state.

94. Retractions can change the meaning of an old search result

A paper retrieved as ordinary evidence at the time of a search can later be corrected or retracted.

High-stakes evidence work should verify current publication status, especially for pivotal sources.

95. Correction awareness belongs in discovery

Searchers should follow errata, retraction notices, corrected versions and linked updates rather than treating the first retrieved PDF as permanently authoritative.

Discovery is version-aware when the evidence matters.

96. Reference lists contain secondary discovery routes

Systematic reviews, guidelines and major narrative reviews can serve as maps to earlier studies.

Their reference lists should be mined critically rather than treated as substitutes for independent database searching.

97. Review-of-reviews searching can reveal vocabulary quickly

Existing systematic and scoping reviews often expose subject headings, key authors, landmark studies and contested terminology.

This is excellent reconnaissance before a new search strategy is finalised.

98. Search provenance should survive into the publication

A final article should preserve enough of the search route that readers can reconstruct how evidence was found.

This includes databases, platforms, exact strategies where appropriate, dates and supplementary search methods.

99. Database platform and database name are not always identical

MEDLINE can be searched through PubMed and other interfaces. Embase can be accessed through different platforms. Interface syntax and features can differ.

Formal reporting should identify the database and platform when that affects reproducibility.

100. Search strings are executable methods

A search strategy is closer to code than prose: small character changes can materially change output.

Versioning, peer review and test cases are therefore sensible quality controls.

101. Preserve the original query before editing it

If a search is revised after inspection, retain the earlier version and note why the change was made.

Silent query replacement erases methodological history.

102. Search versioning is especially important after peer review

If a PRESS reviewer recommends adding a synonym or correcting a Boolean operator, the final strategy should record the resulting version.

The change becomes part of the evidence-production audit trail.

103. Query portability is a translation problem, not a copying problem

A portable search preserves concepts and logic while adapting syntax and controlled vocabulary to each database.

The intellectual search remains stable even when the code changes.

104. Search validation can use recall sets

Researchers can assemble a set of known eligible or highly relevant records and test whether the search retrieves them.

This does not prove completeness but reveals obvious retrieval failures.

105. Search validation can use text mining cautiously

Term-frequency analysis and automated synonym suggestions can reveal vocabulary missing from a draft query.

Suggested terms still require domain judgement because frequent words are not always conceptually relevant.

106. AI can accelerate concept expansion

Language models can propose synonyms, spelling variants, related phrases and candidate controlled-vocabulary terms.

The safe role is proposal generation followed by verification against the actual database vocabulary and known literature.

107. AI can translate search syntax—imperfectly

An AI system can rewrite a PubMed strategy for Scopus or Embase, but platform operators, fields, truncation and vocabulary rules differ enough that the translated query must be tested manually.

A syntactically plausible query can still be semantically wrong.

108. AI can discover papers through semantic similarity

Embedding-based retrieval can identify conceptually related papers even when exact keywords differ.

This is valuable for exploration and supplementary discovery, particularly in terminologically unstable fields.

109. Semantic search has a reproducibility problem

Model versions, proprietary indexes and ranking algorithms can change without exposing the exact retrieval logic.

For formal evidence synthesis, semantic search should supplement rather than silently replace auditable database strategies unless the retrieval system can provide a reproducible record.

110. AI can hide the source universe

An assistant may answer “the research shows” without saying whether it searched PubMed, Crossref, a proprietary index, the open web or a small internal corpus.

High-standard research discovery requires the source universe to be explicit.

111. AI retrieval should separate finding from judging

RETRIEVE
  → this record appears relevant

VERIFY
  → this is the correct article/version

APPRAISE
  → this study has these strengths/limitations

SYNTHESISE
  → this evidence changes the answer by this amount

Collapsing these stages makes fluent retrieval look like evidence appraisal.

112. AI-generated citations require resolution

Any citation suggested by a generative model should be resolved against a real database record, DOI, PMID or authoritative source before it enters a formal search log.

Plausibility is not provenance.

113. AI can assist deduplication

Similarity models can help detect records that represent the same article despite punctuation, transliteration or metadata differences.

Borderline merges should remain reviewable because false deduplication can delete unique evidence.

114. AI can assist screening, but screening is not the search

Machine-learning tools can prioritise likely relevant records after retrieval.

This can reduce human workload, but it does not repair evidence that was never retrieved because the database or search strategy missed it.

115. Search systems create feedback loops

Frequently cited and highly ranked papers are easier to discover, so they are more likely to be read and cited again.

Research discovery therefore participates in the bibliometric system it later measures.

116. Ranking bias can become knowledge bias

If a discovery system consistently promotes already prominent institutions, languages or journals, less visible but relevant work can remain buried.

Comprehensive retrieval should use methods that can escape the top-ranked neighbourhood.

117. Citation chaining is one escape route

Following references and citing papers moves through the scholarly graph rather than depending entirely on keyword ranking.

It is one reason strong searches combine textual retrieval and relational retrieval.

118. Database searching and graph searching are complementary

TEXT ROUTE
  → keywords
  → subject headings
  → fields
  → Boolean logic

GRAPH ROUTE
  → references
  → citing papers
  → authors
  → institutions
  → related records

COMBINE BOTH
  → stronger discovery

119. Literature discovery is not finished when the PDF is downloaded

The researcher must still confirm identity, version, publication state, correction status and relevance.

Retrieval returns candidates. Evidence use requires verification.

120. A paper can have multiple legitimate versions

Preprint, accepted manuscript and version of record may coexist.

Cite and evaluate the version actually used, then check whether a later authoritative state changes the interpretation.

121. Search result counts are diagnostic, not prestige numbers

An unexpectedly tiny result set can reveal over-restriction. An enormous set can reveal broad terminology or an ambiguous concept.

Result counts help debug the search; they do not measure research importance.

122. Line-by-line search building improves debugging

Complex strategies are easier to inspect when each concept family is built separately and combined only after its behaviour is understood.

#1 population synonyms
#2 intervention synonyms
#3 controlled vocabulary population
#4 controlled vocabulary intervention
#5 (#1 OR #3)
#6 (#2 OR #4)
#7 #5 AND #6

This makes it easier to identify which line caused an unexpected collapse or explosion in results.

123. Search testing should inspect records, not only counts

A thousand results can still be wrong. Read samples from the top, middle and edge of the result set.

Inspect why false positives entered and why known relevant records might be absent.

124. False positives teach vocabulary

Irrelevant records often reveal ambiguous terms that need field restrictions, proximity control or removal.

Noise is diagnostic information.

125. False negatives teach blind spots

A known relevant paper missing from the results can reveal absent synonyms, unexpected indexing, unusual terminology or platform syntax problems.

Missed evidence is often more informative than extra evidence.

126. Search expertise is partly adversarial thinking

Ask how a relevant paper could escape the query.

Could the author use an older term? Could the title omit the outcome? Could the database lack indexing? Could the paper sit in another discipline? Could it exist only as a report?

127. The strongest searches actively hunt for what they might miss

Comprehensiveness comes from failure analysis, not from adding endless synonyms blindly.

The searcher models escape routes and closes the important ones.

128. Search design should record model limits

No search can guarantee complete retrieval of every relevant work in the world.

The defensible claim is narrower: the team searched these sources, using these strategies, at these dates, with these supplementary routes, under these stated limits.

129. A good search can still miss inaccessible literature

Some research is unpublished, poorly indexed, confidential, inaccessible, lost, language-restricted or absent from digital systems.

Search quality should not be confused with omniscience.

130. Search limitations belong in the final interpretation

If the search excludes certain languages, databases or dates, the evidence synthesis should acknowledge the possible effect on conclusions.

Method limits become inference limits.

131. A practical exploratory-search workflow

Write the question in ordinary language.
Identify two or three core concepts.
Find one reliable overview or known paper.
Extract vocabulary, authors and subject headings.
Search a broad scholarly database.
Open highly relevant records.
Follow references backward.
Follow citations forward.
Search key authors and institutions.
Record emerging terms.
Repeat until the field map stabilises.

132. A practical systematic-search workflow

Define eligibility criteria before searching.
Choose question framework and core concepts.
Build controlled-vocabulary and free-text term families.
Select databases by subject coverage.
Build the primary strategy line by line.
Test against known relevant records.
Peer review the strategy where appropriate.
Translate it database by database.
Run and export each search.
Preserve exact queries and search dates.
Deduplicate without collapsing distinct publication states.
Search trial registers, repositories or grey literature where relevant.
Conduct backward and forward citation searching.
Document supplementary searches.
Update close to publication where required.

133. A practical database-selection checklist

What disciplines does the database cover?
Which years?
Which countries and languages?
Which document types?
Does it include conference proceedings?
Does it include books?
Does it include controlled vocabulary?
Does it support citation searching?
Can exact strategies be exported or saved?
Is the platform syntax stable and documented?
What important evidence universe lies outside it?

134. A practical search-quality checklist

Question translated into appropriate concepts.
Synonyms and historical terms considered.
Controlled vocabulary included where appropriate.
Boolean logic checked.
Parentheses checked.
Phrase and proximity behaviour checked.
Truncation checked for unintended words.
Field tags checked.
Limits justified.
Known relevant records retrieved.
Database translation tested.
Search warnings reviewed.
Exact strategy preserved.

135. A practical AI-assisted search checklist

Use AI to propose terms, not certify them.
Verify controlled-vocabulary terms in the actual thesaurus.
Verify every generated citation.
Test translated syntax in the target database.
Record which AI assistance materially changed the strategy.
Keep the source universe explicit.
Do not use semantic ranking as proof of completeness.
Keep search, appraisal and synthesis as separate stages.
Preserve human review of deduplication edge cases.
Return important claims to primary records.

136. Failure modes

Failure	What breaks
Question pasted directly into search box	Concept structure and vocabulary variation disappear.
One preferred keyword only	Synonymous literature is missed.
Controlled vocabulary only	New and unindexed records can disappear.
Free text only	Indexed conceptual relationships are underused.
Every PICO element searched	Recall may collapse unnecessarily.
NOT used aggressively	Relevant records are silently deleted.
One database = whole literature	Coverage bias becomes evidence bias.
Copy-paste search across databases	Syntax and subject-vocabulary differences corrupt retrieval.
Top-ranked results only	Ranking replaces completeness.
No citation chaining	Relational discovery routes are ignored.
No grey-literature route	Publication bias can increase.
Preprint deleted as duplicate of journal paper	Version history and study relationships are lost.
No exact search log	Method cannot be audited.
AI suggestions accepted unverified	Invented terms, citations or syntax can enter the method.
Semantic search treated as reproducible systematic search	Opaque ranking replaces auditable retrieval.

137. The deeper model: search is a measurement instrument

A research search does not merely “look around”. It samples the documentary record according to a designed retrieval instrument.

REAL EVIDENCE UNIVERSE
  ↓
DATABASE COVERAGE
  ↓
INDEXING QUALITY
  ↓
SEARCH STRATEGY
  ↓
RANKING / RETRIEVAL
  ↓
EXPORTED RECORDS
  ↓
SCREENING
  ↓
INCLUDED EVIDENCE

Bias can enter at every layer. Good searching makes those layers more visible.

138. The evidence inlet deserves the same quality control as analysis

Researchers routinely peer review statistical models while treating literature search as clerical work.

That is backwards. If relevant studies never enter the evidence set, downstream analysis cannot recover them.

139. Search quality is non-compensatory

A sophisticated meta-analysis cannot compensate for a search that systematically missed a major evidence class.

Beautiful prose cannot compensate for undocumented source selection. A large result set cannot compensate for the wrong databases.

140. The Wintour rule: every evidence claim needs a retrieval return path

CLAIM
  → INCLUDED SOURCE
  → RECORD IDENTITY
  → DATABASE / REPOSITORY
  → SEARCH OR DISCOVERY ROUTE
  → QUERY / CITATION CHAIN / REGISTER
  → DATE RETRIEVED
  → CURRENT PUBLICATION STATE

If the return path cannot be reconstructed, confidence should fall.

Good research discovery is not finding more papers. It is building a search system whose misses, limits and decisions are visible enough to be challenged.

141. Why literature search matters to civilisation

Human knowledge is now too large to read sequentially. Every serious institution therefore depends on retrieval systems: medicine, engineering, law, education, government, science, libraries, journalism and artificial intelligence.

When search works, evidence travels from the people who created it to the people who need it. When search fails, knowledge can exist and still behave as though it were lost.

The civilisation-scale skill is therefore not memorising database syntax. It is understanding how questions become retrieval systems—and how those systems can be tested, documented, corrected and improved.

Current authority routes

Continue the Archives and Publishing series

Publication control: Wintour House · eduKate Publishing · question, evidence, retrieval, source-boundary, provenance, model-limit, freshness, correction and archive gates.

World Return: Before trusting a literature search, ask what it was designed to find, which evidence worlds it could not see, how its terms were translated, what supplementary routes were used, when it was last run, and whether another researcher can follow the route back to the same documentary landscape.

How Literature Search and Research Discovery Work | From Research Questions, Keywords and Controlled Vocabularies to Database Search, Citation Chaining, Alerts and Evidence Retrieval