Picking the Right Academic Database: A Teaching Walkthrough

A student raises her hand. "Which database should I start with?" Three things are wrong with the answers that students usually give. They treat the question as a preference call ("I just use Google Scholar"). They default to a single source ("Google Scholar finds everything"). And they assume that once a database returns a long list, the work is done. None of these survive a real systematic search.

This walkthrough turns "which database" into a defensible decision rule you can teach tomorrow. The spine is the HKUST(GZ) Library workshop deck (Sep 2025, by Junjun CHENG) plus seven evidence anchors: Martín-Martín et al. 2021 for cross-database coverage, the official Clarivate and Elsevier pages for the two curated citation indexes, Priem/Piwowar/Orr 2022 and the Semantic Scholar API docs for the open graphs, Rethlefsen et al. 2021 for reporting, Bates 1989 for the iterative search model, and the scite/Elicit/Thelwall 2026 trio for what AI assistants can and cannot do for you.

Why "Which Database" Is Even a Question

Start by making students see what happens when you run the same query in six different tools. The headline numbers from Martín-Martín, Thelwall, Orduna-Malea and Delgado López-Cózar's six-database coverage comparison, drawn from 2,918,105 citations to 2,319 English-language 2006 documents, are 88% for Google Scholar, 60% for Microsoft Academic (now retired), ~52% for Scopus (the matrix first row in Figure 1 reads 57% — the paper's abstract rounds to 52%), 54% for Dimensions, 52% for Web of Science, and 28% for COCI. The first row of the 6×6 matrix captures the entire story in one visual.Martin-Martin 2021

Figure 1 (caption): The first row of the matrix shows what each of the six databases finds out of every citation. Google Scholar finds 88% — the flattest of the six. The largest source is not always the most informative overlap partner, because the off-diagonal cells show that the same paper can be reachable through one database but not another, even when both are "big".

The matrix alone makes the point. The UpSet plot, the second canonical figure from the same paper, gives the empirical warrant for keeping Google Scholar in the workflow even when a curated index is your primary source. The single tallest bar is "Google Scholar only" at 26% of all citations — that 26% is the part no curated citation index has at all, mostly preprints, grey literature, and non-English work.Martin-Martin 2021

Figure 2 (caption): Combinations of which databases find each citation, ranked by frequency. The top bar (26%) is the citations that only Google Scholar finds. Below it, the 21% bar is the citations every one of the six databases finds. Below that, 18% is found by all except COCI. The pattern means no single database catches everything.

The deck's own opening visual is a three-circle nesting of Google Scholar > Library PowerSearch > Databases, with the message that the inner circles are filtered, more reliable, more accessible (per the HKUST(GZ) Library workshop deck, Sep 2025).

Figure 3 (caption): The deck's own framing of the same idea. Google Scholar is the largest pool but the least curated. PowerSearch covers the library's subscribed resources. Databases are the most filtered, scholarly, and advanced. The empirical matrix above adds the numbers this nesting leaves out: the 26% Google-Scholar-only slice means the outermost ring cannot be omitted, even when the inner ring is your primary source.

The combination of the three visuals is the chapter's defensible rule. No single database covers the literature. Field-by-field choices matter — the same paper reports that Web of Science covers 25% of citations in the Humanities but 72% in Chemical & Material Sciences, and Google Scholar's coverage is the flattest of the six (85%–90% across all eight broad fields, per Table 2 and §3.2.1 of Martín-Martín et al. 2021). And Google Scholar's 26% unique slice means it is the only source that surfaces grey literature, preprints, and non-English work at scale. Use more than one database, and document the combination.

The Six Database Families — Side by Side

Once students accept that one database is not enough, the next question is what the major options actually are. The deck lists six families: publisher platforms, cross-disciplinary citation indexes, thematic (discipline-specific) databases, open-access repositories, academic search engines (Google Scholar), and new citation indexes (Semantic Scholar, OpenAlex). The first three are subscription tools you reach through your library; the last three are open or partly-open.

The curated-index pillar is Web of Science. As of August 2026, the Core Collection contains more than 22,994 journals and 101.7 million records, organised into six editions (Science Citation Index Expanded, Social Sciences Citation Index, Arts & Humanities Citation Index, Emerging Sources Citation Index, Conference Proceedings Citation Index, Book Citation Index), across 254 Clarivate-curated subject categories. The acceptance rate for SCIE, SSCI, and AHCI is 10%–12% — by definition most journals do not qualify, and once admitted, every article is indexed cover to cover.Clarivate WoS Coverage

Figure 4 (caption): The Web of Science platform's radial structure. The central purple sphere is the Core Collection; six bold editions sit on top; three coloured rings around the rim — regional indexes, beyond-published-literature collections, and specialty collections — extend the reach without going through the same editorial gate. Use this to teach that "Web of Science" is not one database; it is a Core plus a constellation.

The broad-alternative pillar is Scopus. As of the Content Coverage Guide (June 2026), Scopus carries 31,570 active sources (29,905 active journals plus books, trade journals, and book series), 470,000 stand-alone books (55% in the social sciences and humanities), 13.2 million conference papers, and 105 million records. Its subject taxonomy (the ASJC tree) runs four clusters down to 27 major areas and 300+ minor areas, which Elsevier markets as "330 disciplines." The Content Selection and Advisory Board, with 17 Subject Chairs, accepts roughly 48% of titles that pass the technical screen.Elsevier Scopus Content

Figure 5 (caption): Elsevier's current coverage snapshot for Scopus. 2.6 billion cited references dating back to 1970; left column counts active journals, books, preprints, publishers, and open-access items; right column counts records, author profiles, and affiliation profiles. The infobox anchors the broad-index counterpoint to the curated WoS pillar in Figure 4.

Side by side, the two curated citation indexes look like a quality-versus-breadth trade-off. Web of Science is more selective (10%–12% acceptance) and stronger in sciences; Scopus is broader (over 30,000 active titles), with deeper social-sciences and humanities coverage and a much larger book footprint. The empirical comparison table in Martín-Martín et al. 2021 (the same paper that supplied Figure 1) gives the cross-check: Scopus covers 31% of citations in the Humanities but 75% in Chemical & Material Sciences, and Web of Science 25% versus 72% in the same two fields. Use both, and pair each with a discipline-specific database (IEEE Xplore, PubMed, SciFinder, JSTOR) for depth.

The open graph layer is the third pillar, and it is the one students keep missing. OpenAlex is a fully open index of works, authors, venues, institutions, and concepts, launched on 1 January 2022 by OurResearch as a drop-in replacement for the now-retired Microsoft Academic Graph. As of September 2026 the live API returns 323 million works, 127 million authors, 256,000 sources, and 136,000 institutions. The Priem, Piwowar & Orr 2022 paper itself only commits to "100% open data, open API, open-source code"; the project's later About and Help Center pages explicitly release the dataset under a CC0 public-domain licence.Priem et al. 2022OpenAlex Help

Figure 6 (caption): OpenAlex's data model, not its system architecture. Six coloured nodes for the entity types (works, authors, venues, institutions, concepts) plus the authorships join object that links authors and institutions to works. The arrows show the direction of "is about" relations, not data flow through software. The visible sixth node, authorships, is what makes affiliation queries non-trivial — the link always goes through this join.

Semantic Scholar is the other open graph, run by the Allen Institute for AI. The live product page reports 214 million papers, 2.49 billion citations, and 79 million authors, with four first-class endpoint groups on the Academic Graph API: paper search (relevance and bulk), paper detail with TLDR summaries and SPECTER2 embeddings, paper citations, and paper references. The "Highly Influential Citation" feature, inherited from Valenzuela, Ha and Etzioni's 2015 AAAI workshop paper, is a feature-based heuristic rather than a deep classifier; a citation counts as influential only if the citing sentence has no overlapping authors, appears at least three times in a single-citation sentence, contains phrases like "build upon" or "following", or references tables or figures in the cited work.Semantic Scholar API

Figure 7 (caption): A schematic of the four endpoint groups on the Semantic Scholar Academic Graph API. Search endpoints return ranked or matched papers; paper-detail returns metadata, TLDR, embeddings, and citation counts; citations and references return the surrounding context and the intents and isInfluential flags. Tools such as Connected Papers, Litmaps, Sourcely, and Elicit call these endpoints; you can call them too.

The point of this chapter is structural, not promotional. Database choice is not between Google Scholar and "everything else". It is between curated indexes (quality gate), broad indexes (multidisciplinary breadth), discipline databases (depth), and open graphs (no paywall, CC0 or ODC-BY). For any given research question, at least two of these are non-substitutable.

Matching the Research Question to the Right Database

The next step is the decision rule students actually want. Translate the six families into a routing rule, then anchor the rule in the PRISMA-S reporting standard so that students see why the choice is auditable, not stylistic.

The three-branch routing rule is the most defensible starting point.

For a STEM question, start with a discipline-specific database (IEEE Xplore for engineering and computer science, PubMed for medicine and life sciences, SciFinder for chemistry) and add either Scopus or Web of Science for the cross-disciplinary scan. Coverage in STEM is high across both curated indexes; the discipline database adds subject-controlled vocabulary and depth that the curated index does not have.

For a social-science question, start with Scopus, add a discipline database (Sage for sociology, EconLit for economics, PsycINFO for psychology), and add Google Scholar because it captures book chapters, working papers, and grey literature that the curated citation indexes under-index.

For a humanities question, the curated citation indexes are weakest. Start with Google Scholar and JSTOR, add DOAJ for open-access humanities journals, and add a discipline database (Historical Abstracts, MLA International Bibliography) for the parts the broader tools miss.

The empirical justification for this routing lives in the field-level table of the same Martín-Martín et al. 2021 paper that supplied Figure 1. Web of Science covers 25% of citations in the Humanities versus 72% in Chemical & Material Sciences. Scopus covers 31% in the Humanities versus 75% in Chemical & Material Sciences. Google Scholar is the flattest of the six (85%–90% across all eight broad fields, per Table 3 and §3.2.1 of the same paper). The contrast across rows is the answer to the question "why does the routing differ by field?"

The reporting obligation is what turns the routing into more than a tip. PRISMA-S, an extension to the PRISMA 2020 Statement published in January 2021 by Rethlefsen and colleagues, organises 16 reporting items into four groups: Information Sources and Methods (items 1–7, what you searched), Search Strategies (items 8–13, how you searched), Peer Review (item 14, who checked the strategy), and Managing Records (items 15–16, what you got and how you de-duplicated). Item 1 mandates naming each individual database searched and stating the platform. Item 8 mandates that the full search strategy for each database be included, "copied and pasted exactly as run". The reason this matters is reproducibility: reviews where a librarian is a co-author include reproducible searches only 64% of the time, and an irreproducible search is one that another team cannot re-run.Rethlefsen et al. 2021

Figure 8 (caption): The 16 PRISMA-S items, organised by reporting group. Information Sources and Methods (7 items) — what you searched, including the database name and platform. Search Strategies (6 items) — how you searched, including the copy-pasted full strategy and the dates. Peer Review (1 item) — PRESS or equivalent. Managing Records (2 items) — total records per database and the deduplication process.

The chapter's rule is therefore: the right database depends on the field and the question. STEM leans on curated citation indexes plus a discipline database, social science leans on Scopus plus Google Scholar plus a discipline database, humanities leans on Google Scholar plus DOAJ plus JSTOR. And whichever combination a student picks must be reported (PRISMA-S Item 1), not just remembered.

Building the Search String — Boolean, Keywords, Truncation

A defensible database choice still returns noise unless the search string inside that database is well-built. The deck's worked example, the question "What is the impact of social media on body image among Generation Z?", supplies every example this chapter needs. The student-facing slide is a colour-coded three-column table, one column per concept, with three rows of expansion: synonyms, phrase searching in double quotes, and truncation with an asterisk (per the HKUST(GZ) Library workshop deck, Sep 2025).

Figure 9 (caption): The deck's worked research question broken into three concepts (Social Media in blue, Body Image in orange, Generation Z in green), each expanded with a synonym list, a list of phrase-search targets in double quotes, and a truncation root. The yellow "Be careful!" box flags lib* (library, libraries, liberty, liberation) and teen* (teenager, teenagers, teens) — over-truncation pulls in semantically unrelated matches.

The Boolean operators are the set-theoretic backbone of any search string. The deck's Venn diagrams label them with the same social-media example that UCL's library guide uses (CC BY-NC-SA 4.0), which is a clean attribution trail for your slides: OR finds either term (the union of two circles), AND finds only records that contain both terms (the intersection), and NOT excludes one set from another (set difference). Use OR to broaden synonyms within a concept, AND to narrow different concepts, and NOT with caution, because NOT drops any record that mentions both the kept and the excluded terms.UCL Library Skills

Figure 10 (caption): The three Boolean operators rendered as Venn diagrams on the same social-media example. OR (left) shades both circles entirely — every record with either term is retrieved. AND (middle) shades only the overlap — only records with both terms are retrieved. NOT (right) shades only the part of the left circle that does not overlap the right.

The same string can be expressed in two interface styles, and the choice of style is platform-specific. Scopus Advanced Search uses a multi-box form, one input row per concept, joined with an AND dropdown. The deck's example has three rows, each searching the title/abstract/keyword field: row 1 "social media" OR facebook OR instagram OR twitter OR snapchat OR tiktok (the Gen Z deck has a stray opening quote around twitter that should be silently corrected when teaching), row 2 "body image" OR "self-perception" OR "self-esteem" OR "mental health", row 3 teenagers OR adolescents OR youth OR "young people". The WoS equivalent, in the single-string Topic field with TS= notation, is the same logic collapsed into one box with parentheses (per the HKUST(GZ) Library workshop deck, Sep 2025).

Figure 11 (caption): The Scopus "Start exploring" form with three input rows joined by AND. The first row uses Article title/Abstract/Keywords and contains "social media" OR facebook OR "twitter OR Instagram (the deck's stray quote around twitter is a screenshot artefact, not a feature). The second row is "mental health" OR "body image". The third row is teenagers OR adolescents OR youth OR "young people".

Figure 12 (caption): The Web of Science single-string interface, with one Topic box taking the entire Boolean expression in parentheses: ("social media" OR facebook OR Instagram OR Twitter) AND ("Body Image" O… (the screenshot is truncated, but the pattern is the same logic in one box). The Topic field defaults to title/abstract/keywords, and the platform collapses parentheses automatically.

Two operator pitfalls are worth flagging up front. Phrase searching with double quotes forces adjacency — without quotes, most databases default to implicit AND, and "social media" with intervening text becomes a different match. Truncation with an asterisk finds word-family matches (Psycholog* → Psychology, Psychologist, Psychological), but a too-shallow root pulls in noise. The deck's yellow "Be careful!" box, on lib* (library, libraries, liberty, liberation) and teen* (teenager, teenagers, teens), is the same warning against over-truncation that UCL's library guide gives — UCL's own example is Psych* (Psychiatry, Psyche, Psychic), so the lesson is shared but the worked examples are not. The second, non-obvious pitfall is that on PubMed and Web of Science, using wildcards or truncation can disable the database's automatic term mapping or lemmatization, which means a thoughtful searcher might be giving up the platform's best features for an asterisk.UCL Library Skills

The bridge from Boolean-by-hand to AI-assisted keyword expansion is the platform's own quick-add chips. In Web of Science, the deck's Sep 2025 screenshot shows the same three-topic query in full TS= notation returning 1,809 results (a snapshot figure — the live WoS count will drift and should be verified at teach time), with a Quick-add keywords panel suggesting terms like PROBLEMATIC SOCIAL MEDIA USE, FACIAL DISSATISFACTION, STRESS AND ASTHMA, SELFIE-POSTING, and SOCIAL MEDIA ADDICTION, each toggleable as INCLUDE (AND) or EXCLUDE (NOT). This is the platform helping the searcher expand the synonym set, and it is the entry point into the AI tools chapter (per the HKUST(GZ) Library workshop deck, Sep 2025).

Figure 13 (caption): A composite of three WoS screenshots. The top purple box is the full TS=(...) query with the three concepts joined by AND. The lower-left is the WoS result page (1,809 records in the deck's Sep 2025 screenshot — treat the count as a snapshot, not a current number), with the green Add keywords button highlighted. The lower-right is the AI-enabled topic and keyword suggestion panel, with the suggested terms listed as chips the searcher can INCLUDE (AND) or EXCLUDE (NOT) without leaving the page.

The chapter's rule is concrete: a Boolean string lives inside parentheses, three operators (OR broadens synonyms, AND narrows different concepts, NOT excludes), phrase marks keep words adjacent, and truncation finds word-family matches but can disable PubMed auto-mapping and WoS lemmatization. The same string looks different in Scopus 3-box versus WoS single-string, and a defensible search reports the full string as it was run, in the form the platform accepted.

From One Seed Paper to a Citation Network

Topic-based Boolean searches always leave gaps. The reason is in the literature long before digital indexes, and it is named. In 1989, Marcia Bates published "The design of browsing and berrypicking techniques for the online search interface" in Online Review 13(5):407–424, arguing that the query is not static and the retrieval is not a single best set. Real-life searchers change the question itself as they learn from each document they pick up (the "evolving search" model), and useful references come in bit-at-a-time selections scattered across many moves (the "berrypicking" model).Bates 1989

Figure 14 (caption): Bates's canonical schematic of the iterative search. A curving track runs from Q0 (lower-left) through successive query-variation nodes Q1–Q5, with cloud-shaped T (thought) markers and stacked-document icons along the path, terminating at E (upper-right). The figure predates digital citation indexes but describes the model they operationalise. The shape is a curve, not a single best set, and the queries themselves change as the searcher reads.

Bates also listed the six manual search techniques that real searchers combine: footnote chasing (backward chaining), citation searching (forward chaining), journal run, area scanning, subject searches in bibliographies and abstracting/indexing services, and author searching. The empirical warrant for spending effort on chaining is in the paper itself: footnote chasing was found to retrieve more relevant material than the original search in 69% of cases. That figure is the basis for any "start with one good paper, then chain" advice you give students.

The modern operationalisation of berrypicking for systematic reviews is CLUSTER, proposed by Andrew Booth and colleagues in 2013 (BMC Medical Research Methodology 13:118) and reviewed by Tsang and Maden in 2021. The mnemonic stands for seven search-access points (Citations, Lead authors, Unpublished materials, Scholar searches, Theories, Early examples, Related projects), operationalised as a 13-step procedure that starts from a single "key pearl citation". In practice, no review has followed the full 13 steps; the most-followed are step 1 (identify the key citation, 14 of 16 reviews in Tsang and Maden's sample), step 4 (lead-author search, 9 of 16), and step 5 (citation searches, 8 of 16). Realist reviews are the most common use case. CLUSTER is supplementary to topic-based Boolean, never a replacement.Booth 2013 / Tsang 2021

Figure 15 (caption): The CLUSTER "inclusion by affirmation" mechanism. A paper initially marked Unsure on title/abstract review is independently affirmed for inclusion when its sibling study from the same project is itself included. This is the visual idiom for the rule that the search procedure can rescue borderline papers that a topic-based Boolean would have rejected.

In the modern open-graph tools, the same chaining is automated. The deck's citation-chaining example uses a single seed paper (Marengo et al. 2018, "Highly-visual social media and internalizing symptoms in adolescence", Computers in Human Behavior 82:63–69) and shows the Web of Science Citation Network panel: 132 Forward citations in the Core Collection, 62 Backward references, and a "View Related Records" link (per the HKUST(GZ) Library workshop deck, Sep 2025).

Figure 16 (caption): A composite from the deck's "Citation Chaining" slide. The seed paper A (Marengo et al. 2018) is shown with two flanking document stacks: B for "Backward Searching" (62 references A cites) and C for "Forward Searching" (132 papers that cite A). The right-side Web of Science Citation Network panel marks Forward in green (132 citations) and Backward in red (62 references). The example shows that topic-based Boolean alone never returns the full cluster.

The chapter's rule is concrete. Topic-based Boolean searches always leave gaps. For complex reviews, run CLUSTER from a known seed paper using forward (cited-by) and backward (references) chaining. Berrypicking is the historical name, CLUSTER is the named procedure, and Semantic Scholar and OpenAlex are the open graphs that automate it.

AI Tools — Use Them, Don't Trust Them Blindly

The set of tools students now meet in any literature-search workshop is large: Elicit, scite, Semantic Scholar, OpenAlex, Connected Papers, Litmaps, Consensus, ResearchRabbit, SciSpace, scite Assistant, ChatGPT via Poe. The teaching rule is the same for all of them: AI tools propose, humans verify.

Elicit is the most prominent current AI research assistant for systematic reviews, drawing on a 138-million-paper internal corpus alongside PubMed and ClinicalTrials.gov. Its systematic-review product follows a five-step PRISMA-2020-aligned workflow (refine, gather, screen, extract, synthesise) and pairs semantic vector search with traditional Boolean keyword search. The May 6, 2026 evaluation blog reports 95% search recall on 994 open-access Cochrane reviews (semantic only), 96.9% sensitivity on abstract screening, 99.5% paper-level recall on full-text screening, and 95.6% accuracy on extraction. The current pricing page (Sep 2026) lists Pro screening at 5,000 papers and Enterprise at 40,000, with synthesis reports drawing from up to 200 sources on the Scale plan. Every claim Elicit generates is paired with a sentence-level citation and a per-criterion quote, which is the differentiator that makes the "AI proposes, human verifies" rule practical rather than aspirational.Elicit

Figure 17 (caption): Elicit's dynamic screening interface. The left panel is a criteria editor (here, "Only include studies where the population age is between 18 and 29"), and the right panel is the live "821 papers found" with a coloured-dot yes/maybe/no count for each of four attributes (Population age, Intervention Type, Study Duration, Outcome Measures). The screenshot is the product UI; the 821 papers is the example dataset, not a product cap.

Figure 18 (caption): Elicit's sentence-level citation feature. A paper card (Held et al. 2002 on magnesium supplementation) carries an AI-generated inclusion rationale and a popover window with the supporting quote ("a placebo group was essential for …") and the inline (Held et al., 2002) marker. The feature is the answer to the "AI proposes, human verifies" rule: every claim points to the exact sentence the model used.

scite adds a different layer. Its Smart Citations feature classifies every citation statement as supporting, contrasting, or mentioning, with a real-world class distribution of 92.6% mentioning, 6.5% supporting, and 0.8% contrasting, which means any classifier that simply defaults to "mentioning" would already score 92.6% accuracy. scite's own SciBERT production model reports an F-score of 0.5897 on the disputing class, 0.6480 on supporting, and 0.9728 on mentioning, with precision tuned above 80% per class. The 2023 independent evaluation by Bakker, Theis-Mahon and Brown, on 98 classified citations of retracted pharmacy publications, found scite labelled 2 supporting, 96 mentioning, and 0 contrasting against human labels of 42 supporting, 39 mentioning, and 17 contrasting, yielding F-measures of 0.096, 0.58, and 0.0 respectively. Coverage is also a problem: only 31% of the 324 original citations were classified at all. The takeaway is that scite's class labels are a triage signal, not a verdict.Nicholson et al. 2021 / Bakker et al. 2023

Figure 19 (caption): scite's side-by-side comparison. The smart citation report (left) shows the citing sentence, the surrounding context, and the citation type label (supporting / mentioning / disputing). The traditional citation report (right) shows only article metadata (title, authors, journal, year, DOI). The differentiator is the contextual label and the in-context citing sentence.

The deepest realism check is the most recent. In a study submitted to arXiv on 31 August 2026, Mike Thelwall analysed 73,489 articles from 15 health and life-sciences journals scored by two LLMs (GPT-OSS-120B and Gemma 3 27B). The headline finding is that GPT-OSS-120B favours viruses, genes, and cells in its scoring and disfavours surveys, patients, and students. Gemma 3 27B gave relatively higher scores to machine-learning papers, a proof by contradiction that at least one of the two LLMs is biased. Both models also showed input-type-dependent differences, comparing full-text scoring against title-and-abstract scoring. The closing sentence of the abstract is the chapter's rule: consider LLM biases when deciding whether to use them for research evaluation tasks.Thelwall 2026

Figure 20 (caption): Each dot-and-whisker is a cluster of paper titles scored by GPT-OSS-120B. The dashed vertical line is the overall mean (~2.46 on the 2.3–2.7 axis). Topic 7 (virus, viral, viruses) sits highest, Topic 10 (health, covid, covid-19) sits lowest. The virus-versus-patients gap is roughly the same width as the kind of effect a screening tool would treat as a real signal, and that is the point. The findings are from 15 health and life-sciences journals only, so they are a caution, not a universal indictment.

The chapter's rule is concrete. Elicit screens 95% reliably on its own benchmark. scite's classification F-score is 0.0–0.58 on independent samples. LLM topic scoring is biased by topic. Use these tools as triage signals, not as verdicts.

Pitfalls, Reproducibility, and the Iterative Loop

The closing chapter ties the previous six together. The empirical reminder is that no single source catches everything; the 26% Google-Scholar-only slice from the UpSet plot in Chapter 1 still applies, and the open-graph layer (OpenAlex, Semantic Scholar) is now a real alternative to subscription indexes for institutions that cannot afford Scopus or Web of Science.

The reporting reminder is that PRISMA-S makes the database choice and the search string auditable. Item 1 mandates naming each individual database plus the platform (for example, "MEDLINE via Ovid" rather than just "MEDLINE"). Item 8 mandates that the full search strategy for each database be copy-pasted exactly as run. Item 13 mandates the date each strategy was last run. Item 15 mandates the total records identified per database. These four items are the ones that close the loop on the database choice.Rethlefsen et al. 2021

The methodological reminder is that berrypicking and PRISMA-S are not in conflict. The deck's own closing flow diagram shows a loop from "Developing & Executing Your Search Strategy" back to "Selecting Databases to search" and to itself, with a "How Much Literature is Enough?" sidebar reminding students that "Searching for, reading and understanding the existing literature is more of an art than a science" and that "new literature is being published daily" so the searcher must stop at some point. The loop is fine. The rule is that every loop must end with a search another team can re-run, which is the bridge between Bates's open model and PRISMA-S's reporting model (per the HKUST(GZ) Library workshop deck, Sep 2025).

Figure 21 (caption): The deck's own closing schematic. A horizontal flow with Start → Selecting Databases to search → Developing & Executing Your Search Strategy → End, plus two feedback arrows from the executing step back to itself and back to the database step. The pink-bordered sidebar "How Much Literature is Enough?" lists three reminders: searching is an art, new literature appears daily so you must stop, and the volume of research on the topic sets the search budget. The loop is the chapter's last image: pick a database, build a string, run it, refine it, report it.

The practitioner rule that closes the chapter is: use more than one database; field-by-field choices matter; document the search string and dates; peer-review the strategy before submission (PRESS 5/2015 is the canonical checklist for that peer review). The skill you are teaching is not "find the right database" but "pick a defensible combination, build a defensible string, run it on the right platform, and write the combination and the string down so that someone else can re-run it."

References

1. Martín-Martín, A., Thelwall, M., Orduna-Malea, E., & Delgado López-Cózar, E. (2021). Google Scholar, Microsoft Academic, Scopus, Dimensions, Web of Science, and OpenCitations' COCI: a multidisciplinary comparison of coverage via citations. Scientometrics 126, 871–906. — arXiv:2004.14329
2. Clarivate (2026). Web of Science Coverage Details. Web of Science Help Center, last updated 18 Aug 2026. — https://webofscience.zendesk.com/hc/en-us/articles/48137951657233-Web-of-Science-Coverage-Details
3. Elsevier (2026). Scopus — Content Coverage and Database Composition. Elsevier Products, Content Coverage Guide Jun 2026. — https://www.elsevier.com/products/scopus/content
4. UCL Library Skills (2026). Search techniques — Searching for information. Last updated 27 Aug 2026 (CC BY-NC-SA 4.0). — https://library-guides.ucl.ac.uk/searching-information/search-techniques
5. Elicit (2026). Systematic Reviews and AI Research Assistant. Product page + May 6 2026 evaluation blog + Sep 3 2026 pricing page. — https://elicit.com/solutions/systematic-review
6. Priem, J., Piwowar, H., & Orr, R. (2022). OpenAlex: A fully-open index of scholarly works, authors, venues, institutions, and concepts. arXiv:2205.01833 (STI 2022). — https://arxiv.org/abs/2205.01833
7. Rethlefsen, M. L., Kirtley, S., Waffenschmidt, S., et al. (2021). PRISMA-S: an extension to the PRISMA Statement for Reporting Literature Searches in Systematic Reviews. Systematic Reviews 10, 39. — https://systematicreviewsjournal.biomedcentral.com/articles/10.1186/s13643-020-01542-z
8. Thelwall, M. (2026). Do Large Language Models Favour Any Research Topics? arXiv:2609.00323 v1, 31 Aug 2026. — https://arxiv.org/abs/2609.00323
9. Cheng, J. (2025). Effective Strategies and Tools for Literature Searching — HKUST(GZ) Library Workshop Slides, Sep 2025 (deck artefact, HKUST(GZ) Library). — HKUST(GZ) Library
10. King's College London (2026). Advanced search techniques — Searching for Systematic Reviews & Evidence Synthesis. LibGuide, Last Updated 7 Aug 2026. — https://libguides.kcl.ac.uk/systematicreview/advanced
11. Bates, M. J. (1989). The design of browsing and berrypicking techniques for the online search interface. Online Review 13(5), 407–424. — UCLA GSeIS PDF mirror
12. Tsang, A., & Maden, M. (2021). CLUSTER searching approach to inform evidence syntheses: A methodological review. Journal of Research Synthesis Methods 12(3), 391–404 (re-presenting Booth et al. 2013, BMC Med Res Methodol 13:118). — https://onlinelibrary.wiley.com/doi/10.1002/jrsm.1502
13. Semantic Scholar (2026). Academic Graph API + Kinney et al. 2025, The Semantic Scholar Open Data Platform (arXiv:2301.10140) + Valenzuela et al. 2015, Identifying Meaningful Citations (AAAI Workshop). — https://www.semanticscholar.org/product/api
14. Nicholson, J. M., Mordaunt, M., Lopez, P., et al. (2021). scite: A smart citation index that displays the context of citations and classifies their intent using deep learning. Quantitative Science Studies 2(3), 882–898; Bakker, C., Theis-Mahon, N., & Brown, S. J. (2023). Evaluating the Accuracy of scite, a Smart Citation Index. Hypothesis 35(2). — QSS open access | bioRxiv preprint
15. OpenAlex — Help Center About + Pricing; Semantic Scholar — homepage; VOSviewer 1.6.21 (12 Jun 2026) — homepage + User Manual PDF. — https://help.openalex.org/hc/en-us | https://www.semanticscholar.org/ | https://www.vosviewer.com/