Six Lessons Beyond Google Scholar

A literature search is not a one-shot Google query. The first three hits on Google Scholar are not a literature review; they are a snapshot of what the algorithm thought was closest to your words. Behind that snapshot sits a layered stack of search platforms, citation indexes, call-number systems, journal metrics, and now a small number of open knowledge graphs that every "AI research tool" quietly sits on top of. The six chapters that follow walk through that stack in order, from the physical book spine to the LLM summariser, and rebuild the mental model most researchers carry into their first literature review.

By the end, the expectation is that you can pick the right search tool for a given question, decode a Library of Congress call number on a book spine, translate a sentence into a Boolean query, read a Journal Impact Factor and a citation count with the right caveats, recognise what an AI research assistant is actually doing under the hood, and treat the whole thing as a loop you return to rather than a single search box you walk away from.

What is a library actually asking you to do, and how do you find one book on a shelf?

A library is not asking you to memorise a stack of databases. The Association of College and Research Libraries, the body that sets information-literacy standards for North American universities, frames what researchers are actually doing in two short sentences that bear repeating verbatim. On Research as Inquiry the ACRL writes: "Research is iterative and depends upon asking increasingly complex or new questions whose answers in turn develop additional questions or lines of inquiry in any field." On Searching as Strategic Exploration it adds: "Searching for information is often nonlinear and iterative, requiring the evaluation of a range of information sources and the mental flexibility to pursue alternate avenues as new understanding develops" ACRL Framework. Both lines are operational, not aspirational: the library expects you to come back, change your question, and pick a different tool when the first one fails. The rest of this chapter is about the smallest unit of that iteration, which is finding one specific book on one specific shelf.

The most direct answer a library gives to "where is this book?" is the call number printed on the spine. The Library of Congress, whose classification system most research libraries in North America use, defines a call number in its instruction sheet G 55 as "a classification number and a book number," with a publication date added as a third element on monograph call numbers since 1982 LoC G 55. The phrase "parsed address" captures the intent: each segment tells the shelver, and you, something specific.

The example the workshop decodes throughout is QA 274.2 R423 1997, printed on the spine of Stochastic Analysis on a mathematics shelf. Each segment parses as follows.

Caption: A call number is a parsed address. Reading QA 274.2 R423 1997 left to right: QA puts the book in subclass Mathematics inside class Q (Science); 274.2 is the classification number (the LoC places QA 273–280 in Probabilities and Mathematical Statistics, with QA 274.2 inside Stochastic analysis); R423 is a Cutter number standing in for the main entry, normally the author; 1997 is the publication year, included as a chronological tie-breaker since 1982. The other QA 274.2 books on the same shelf above show that the classification number stays constant while only the Cutter and the year move, which is what lets a shelver find one specific book in a hundred on the same subject.

Two things follow once you accept the call number as a parsed address. First, the library catalogue is not a search engine that returns a list; it is a discovery layer that fronts both the print collection (organised by these parsed addresses) and the electronic collection (organised by metadata the catalogue harvests from publishers). A modern library discovery layer illustrates the point: a single search box can be configured to search both the catalogue (for physical items and e-books) and an articles index (for journal articles and conference papers), so a query for "tesla inventor" can return both a print copy of W. Bernard Carlson's biography from the library's print collection and the same title's electronic edition through a linked publisher package. The exact default tabs and labels vary by institution and may change over time, so treat the institutional UI as illustrative rather than authoritative.

Caption: The same book — Tesla: inventor of the electrical age by W. Bernard Carlson — appears twice on the results page: once with a call number anchored on the left (TK140.T4 C37 2015) that maps it to a physical shelf, and once with an access link on the right that points to the electronic edition. The left anchor is the catalogue's parsed address; the right link is the catalogue's pointer to the publisher. Both records resolve to the same book, and the discovery-layer front end lets you pick the one your context demands.

The practical takeaway: when a single search box can return both physical and electronic formats of the same work, the question stops being "where is this book" and becomes "which format do I need right now." That question is the smallest unit of the iterative loop the ACRL frames at the top of the chapter. Everything in chapters 2 through 6 is a larger version of the same loop applied to articles, metrics, and AI-generated summaries rather than to a single shelf address.

If Google Scholar finds 88% of citations, why is that not enough?

The intuitive argument for using Google Scholar alone is simple: it returns more results than anything else. The empirical argument against treating it as a complete substitute is equally simple, once the right numbers are visible.

The most-cited recent comparison of academic search platforms is Martín-Martín et al. (2021), published in Scientometrics. The authors queried six major platforms — Google Scholar, Microsoft Academic, Scopus, Dimensions, Web of Science, and OpenCitations' COCI — for the same population of citations and asked what fraction each platform actually returned. The headline numbers across all subject areas combined are Google Scholar 88%, Microsoft Academic 60%, Scopus 57%, Dimensions 54%, Web of Science 52%, and COCI 28% Martín-Martín 2021. Read by itself, that row looks like a victory lap for Google Scholar. The discipline-by-discipline breakdown tells the more useful story.

Caption: Eight subject areas, six platforms, and one number that quietly determines how complete your literature review is. Google Scholar stays remarkably flat at 85–90% across every row, including the Humanities. The commercial citation indexes do not: Scopus and Web of Science drop to 25–31% in the Humanities, Literature & Arts row, and COCI bottoms out at 18%. The colour gradient (green high, red low) compresses a six-fold spread between the best and worst platforms in a single column.

Two patterns matter for a researcher deciding what to trust. First, Google Scholar's coverage does not collapse in the Humanities the way the commercial indexes do. If your topic touches society, policy, or culture, Google Scholar genuinely finds more than Scopus or Web of Science. Second, that flatness does not make Google Scholar complete. In the same study, 6–12% of the citations the commercial indexes found are not in Google Scholar, particularly non-English venues, so an answer that relies on Google Scholar alone is an answer with a known and unexplained hole Martín-Martín 2021.

The cross-source matrix makes the asymmetry visible from a different angle.

Caption: Each pie-with-bite cell shows the share of citations found by the row source that are also found by the column source. Google Scholar's row reads 88–94% against every other source, which is what "near-superset" means in practice: most of what Scopus, Dimensions, or Web of Science can find, Google Scholar can also find. COCI's row, by contrast, reads 90–98% against the other five sources — which is what "near-subset" looks like: almost everything COCI finds is also in the larger databases. From the other direction, only 28–47% of what the commercial indexes find is also in COCI, which is what "finds very little the others do not" means in practice. Between these extremes, the three second-tier commercial databases overlap 83–88% with each other — heavy redundancy if you pay for all three.

The reason "broadest" does not equal "complete" is methodological, and the study is honest about it. The corpus is 2,515 English-language highly-cited documents from 2006, harvested across 252 Google Scholar subject categories, with 3,073,351 citations resolved through fuzzy-matching author + title + year metadata. That cohort skews toward papers that are heavily indexed and toward the English-language scholarly world; a researcher working on a niche topic in 2026 should expect lower coverage from every source, and the gap between Google Scholar and the commercial indexes may be smaller or larger depending on language, recency, and citation density Martín-Martín 2021. The numbers above should be read as a structural claim about the platform stack, not as a measurement of any individual query.

What the structural claim buys you is a six-tier map of the search-platform landscape, which makes it possible to place any tool you encounter on the right shelf. Cross-disciplinary citation indexes (Web of Science, Scopus) bring curated coverage, citation context, and journal metrics at the cost of subscription and a known bias toward indexed journals. Thematic databases (IEEE Xplore, PubMed, IOPscience, SciFinder, Sage) trade breadth for depth in a single field. Academic search engines (Google Scholar) trade curation for coverage and discoverability. Open-access repositories (DOAJ, CORE, Zenodo, PLOS, arXiv) trade journal prestige for free access. New citation indexes (Semantic Scholar, OpenAlex) trade commercial licence walls for open graphs, which Chapter 5 unpacks. PowerSearch, the discovery layer introduced in Chapter 1, sits across this whole map as the front door the HKUST(GZ) library offers for almost any of them.

How do you turn a sentence into a database query?

Once a tool is chosen, the next question is how to phrase the question so the database understands it. Databases do not parse natural language. They match the words you type against the words in their records, so a sentence like "what is the impact of social media on body image among Generation Z" becomes, in the database's view, an unscored collection of words scattered across many fields. The craft that converts the sentence into something a database can score is short and learnable.

The first move is to extract the key concepts. For the Generation Z body-image question, the concepts are three nouns: Social Media, Body Image, Generation Z. Everything else in the sentence ("what is the impact of", "on", "among") is glue that means nothing to the database. Once the concepts are isolated, the next move is to expand each one into a list of related terms, because no single word covers every paper in a field. Social Media expands to Facebook, Instagram, Twitter, Snapchat, TikTok, and the umbrella phrase "online platforms". Body Image expands to self-perception, self-esteem, and the related outcomes of mental health, anxiety, and depression. Generation Z expands to teenagers, adolescents, youth, and "young people", with the truncation teen* catching teenager, teenagers, and teens in one term.

The Boolean operators AND, OR, and NOT are the joiners that turn three expanded lists into a single query. Their semantics are simplest when seen as Venn diagrams.

Caption: OR is the union: any record that matches either circle is included, which is why OR is used inside a single concept's synonym list. AND is the intersection: a record must match both circles to be included, which is why AND is used between concepts. NOT is left minus right: a record that matches the left circle but also matches the right circle is dropped, which is why NOT is used to remove an unwanted concept (here, TikTok is excluded from a query about Social Media in general).

Following those semantics, the worked example compiles into a three-line Boolean string:

("social media" OR Facebook OR Instagram OR Twitter OR TikTok)
AND
("body image" OR self-perception OR self-esteem OR "mental health")
AND
(Generation Z OR teenagers OR adolescents OR youth OR "young people")

The same string appears in the Scopus three-field search form, where each concept sits in its own row and the database joins them with AND behind the scenes.

Caption: The abstract Boolean string and the concrete database UI are the same query, viewed two ways. Each Scopus row scopes its search to article title, abstract, and keywords (the dropdown on the left) and contains the OR-joined synonym list for one concept; the rows are joined with AND. The dropdown menus and the explicit operator chips are the database rendering of the parentheses the abstract string spells out.

The ACRL's Searching as Strategic Exploration frame is the conceptual backdrop for this craft. It names two dispositions the workshop is teaching: divergent thinking, which is the brainstorming that produces the synonym list for each concept, and convergent thinking, which is the act of picking the best source for the query you have built ACRL Framework. The frame also warns that first attempts at searching rarely produce adequate results, which is the same point Chapter 6 makes about the loop as a whole. The Boolean string is the unit of work that gets re-run, expanded, narrowed, and re-combined as the search returns evidence.

How do you read the numbers without being fooled?

The results page returns numbers, and the numbers lure readers into two common errors. The first is reading the Journal Impact Factor as a paper-quality score. The second is reading a citation count as evidence that a paper was cited for the right reason. Both numbers carry caveats that the metrics themselves do not print on the page.

The Journal Impact Factor is a journal-level ratio, not a paper-level score. The Clarivate JCR Reference Guide (June 2024) defines it as the count of citations received in the JCR year from any material to items the journal published in the prior two years, divided by the count of citable items (Articles and Reviews only) the journal published in those same two prior years Clarivate JCR Reference Guide. The numerator and denominator use different scopes, which is the structural detail that produces most of the confusion. The numerator pulls from every Web of Science Core Collection index — Science Citation Index Expanded, Social Sciences Citation Index, Arts & Humanities Citation Index, Emerging Sources Citation Index, and the Conference Proceedings and Book Citation Indexes — while the denominator counts only Articles and Reviews. Editorials, letters, news items, and meeting abstracts are excluded from the denominator even if they were cited. The asymmetry is structural and intentional: the metric was designed in the 1970s as a librarian's journal-selection aid, not as a paper-quality, author-quality, or institutional-quality score Clarivate JCR Primer.

The official Clarivate formula card makes the construction visible.

Caption: The numerator (in purple) counts every citation received in 2023 from journals, proceedings, and books to anything the journal published in 2021 and 2022. The denominator (in black) counts only the Articles and Reviews the journal published in 2021 and 2022. The "JCR year" label on the formula is one year older than the release label: a journal's "2024 JIF" is computed from 2023 citations to 2021–2022 items. That year-shift is the second source of confusion: the number labelled 2024 is built from 2023 data.

The Psychology & Marketing case the workshop walks through sits cleanly inside this construction. Its 2024 JIF is 9.1, which means 2023 citations to its 2021–2022 Articles and Reviews, divided by the count of its 2021–2022 Articles and Reviews, yields 9.1 Clarivate JCR Reference Guide. The number is meaningful as a journal summary, but the Clarivate primer itself warns against using it "as a proxy measure for any other entity" — that includes individual papers, individual authors, and individual institutions. A 9.1 journal-level signal tells you nothing about whether any specific paper in that journal is strong, weak, or citable.

The second number, the citation count, has a parallel issue. The count tells you how often, not why. Web of Science addressed that gap by launching the Enriched Cited References classifier in April 2021, which reads the full text of citing articles and assigns each in-text mention to one of five buckets: Background, Basis, Discuss, Support, and Differ Clarivate Citation Context whitepaper. The five buckets resolve the why question at the article level. Background cites a paper to orient the current study inside a scholarly area; Basis cites it because the authors are using its data, methods, or ideas; Discuss cites it because the current study engages with it in detail; Support cites it because the authors report similar results; Differ cites it because the authors report conflicting results or methodology. A paper cited 50 times for Background reasons and 2 times for Basis reasons is a different reading-list candidate from a paper cited 50 times for Basis reasons and 2 times for Background reasons, even though the raw count of 52 is identical.

The classifier surfaces in the Web of Science Sort menu, where the Citation class option opens a submenu with one sort per bucket.

Caption: The Sort-by dropdown on a Web of Science results page exposes the same five buckets as separate sort keys: Background: Highest first, Basis: Highest first, Support: Highest first, Differ: Highest first, Discuss: Highest first. Sorting by Differ first surfaces papers whose claims were challenged in the literature, which is the kind of question a literature review benefits from asking explicitly.

The caveat the whitepaper itself does not print is structural: the classifier depends on the full text of the citing article being available to Clarivate's text-mining pipeline. Articles whose full text is not in Web of Science, or whose publishers have not granted text-mining rights, do not produce classification tags, so the five-bucket signal is a property of the citation graph Web of Science can see, not of the citation graph as a whole Clarivate Citation Context whitepaper. Read together with Chapter 2, this completes the picture: Web of Science is selective in what it captures, but the slice it captures is enriched with both a journal-level metric (JIF) and an article-level intent signal (citation class). The selectivity is the cost; the enrichment is the return.

What is actually inside those AI research tools?

The menu of AI research tools is unusually crowded, and the names sound interchangeable: Elicit, Consensus, Connected Papers, Litmaps, ResearchRabbit, the library's own Research Assistant. They feel different in the UI, but under the hood they are largely downstream of the same two open knowledge graphs. Knowing what those graphs are is the difference between treating an AI tool as an authority over the literature and treating it as a fast front end on a publicly auditable substrate.

The first graph is the Semantic Scholar Academic Graph (S2AG), operated by the Allen Institute for AI. The 2023 paper describing it presents S2AG as a four-stage pipeline that turns raw PDFs into a queryable knowledge graph with public APIs.

Caption: Reading the figure left to right. (1) Data Sources: 50+ inputs including Crossref, arXiv, publisher partnerships, and a web crawler. (2) PDF Content Extraction: text extraction via pdfalto, PDFPlumber, PDFMiner; visual region annotation via LayoutParser on PubLayNet; text span annotation via PaperMage. (3) Knowledge Graph Construction: paper deduplication via S2APLER, citation linking, publication-venue normalization via Fatcat, author disambiguation via S2AND, author-affiliation normalization via S2AFF. (4) Semantic Features: TLDR summarization via BART, citation intent via SciCite, highly-influential-citation classification, fields-of-study classification via S2FOS, paper embeddings via SPECTER2. The three public surfaces on the right are the Graph API, the Datasets API, and the Recommendations API.

The scale at the May 2025 release of the paper is 225 million papers, 105 million authors, 195 thousand venues, 2.8 billion citation edges, 650 million paper-author edges, and 65 million paper-venue edges Kinney et al. 2023. The numbers are self-reported by the AI2 team, which is fine for understanding the architecture but worth flagging when comparing magnitudes across graphs. The API served 1.8 billion requests in 2024, with paper-metadata-by-ID as the dominant use case Kinney et al. 2023.

The second graph is OpenAlex, operated by the non-profit OurResearch. It is the explicit drop-in successor to the Microsoft Academic Graph, which Microsoft retired on January 1, 2022; OpenAlex launched in January 2022 alongside that closure Priem, Piwowar, Orr 2022. Its data model is a heterogeneous directed graph.

Caption: The six labelled nodes are Works, Authors, Venues, Institutions, Concepts, and the connecting Authorships object. The five directed edges (venues → works, concepts → works, authors → authorships, institutions → authorships, authorships → works) encode the relations the graph can answer queries about. A graph is the right mental model: every entity has a persistent OpenAlex ID, every relation has a direction, and every query walks the edges.

OpenAlex publishes everything under CC0 (Creative Commons Zero, the public-domain dedication), exposes the data through a full REST API, releases a fortnightly data dump, and open-sources its code Priem, Piwowar, Orr 2022. The CC0 dedication is the legal reason downstream tools can build on the data without negotiating licences, which is exactly what Elicit, Consensus, and the library's Research Assistant do. Elicit's own support documentation states that it searches "over 138 million academic papers from Semantic Scholar, PubMed, and OpenAlex"; Consensus's help documentation lists Semantic Scholar and OpenAlex alongside its own web crawl as the three data layers behind its answers Priem, Piwowar, Orr 2022. The taxonomy the deck uses for these tools is a two-by-two map.

Caption: The map splits AI research tools into two top-level branches. Use LLM for generative answers contains two sub-branches: Academic Search + LLM (Elicit, Consensus, Assistant by Scite, SCISPACE) and General Web Search + LLM (Bing Chat, Bard, Perplexity). Literature Mapping contains two sub-branches: Science Mapping (VOSviewer, CiteSpace) and Citation Based (Connected Papers, Gecko, Inciteful, Litmaps, ResearchRabbit). The Academic Search + LLM sub-branch is the one that sits on top of the S2AG and OpenAlex graphs discussed above; the Citation Based sub-branch uses the same citation edges to build visual maps of paper neighbourhoods.

The map makes one distinction worth keeping. Use LLM for generative answers tools answer questions by combining retrieval over a graph (often S2AG or OpenAlex) with a language model that produces prose. Literature Mapping tools answer questions by visualising the citation graph itself: which papers cite which, which papers cluster together, which seeds lead to which frontier. The first family is useful when you want a synthesised answer to a question; the second is useful when you want to see the structure of a field around a seed paper. The graph behind the LLM tools is the same graph behind the mapping tools; what changes is what each tool chooses to render on top.

The honest line to close on is the substrate framing. Connected Papers, Elicit, Consensus, and the library's Research Assistant are not independent authorities over the literature. They are consumers of S2AG and OpenAlex, which are themselves consumers of PDFs and metadata. Knowing this lets you predict what each tool is good at (broad recall over open graphs; instant graph-walks around a seed) and where each tool will fail (closed-access content not yet indexed by upstream sources; citation counts that lag the actual publication record; language-model summaries that hallucinate, because the LLM is downstream of a retrieval layer, not a substitute for it).

How does it all loop back into one inquiry process?

The six chapters end where Chapter 1 began: with a loop. The ACRL's Research as Inquiry frame states that "research is iterative and depends upon asking increasingly complex or new questions whose answers in turn develop additional questions or lines of inquiry in any field" ACRL Framework. The workshop's process diagram puts a single feedback arrow on that loop.

Caption: The flow runs left to right: Start → Selecting Databases to search → Developing & Executing Your Search Strategy → End. The arrow that matters is the one that loops back from the strategy step to the database-selection step. It says: when the search returns too many results, too few results, or the wrong results, the right move is not to refine the query and hope; it is to revisit the database choice. That single arrow is the entire workshop's closing argument.

The diagram is small but the choices behind it are the ones the previous chapters unpacked. Selecting Databases is where Chapter 2 lives: choosing between PowerSearch, Google Scholar, subject databases, and citation indexes against the empirical record that no single source covers everything. Developing & Executing is where Chapters 3 and 4 live: extracting key concepts, expanding with synonyms, joining with AND / OR / NOT, and reading the resulting numbers (JIF, citation counts, citation class) without being fooled. AI tools is where Chapter 5 lives, as an accelerant on top of the open graphs that sit beside the closed citation indexes. Each step in the diagram is also a chance to return to a previous step with new information.

Two specific techniques show how the loop pays off in practice.

Backward and forward citation chaining. Pick one paper you trust on the topic — a seed article. Backward searching looks at the seed's reference list and pulls the works that influenced it; this is the right move when you are looking for foundational sources that pre-date the seed. Forward searching looks at the articles that have cited the seed since publication; this is the right move when you are tracking how the field has moved on. The two directions around a seed article together open a much wider literature than any single database query.

Caption: The central article A sits between two clusters. The left cluster holds the references (the B articles that influenced A), reachable through backward searching. The right cluster holds the citing articles (the C articles influenced by A), reachable through forward searching. The same seed article opens two literature networks that go in opposite time directions; using both is how citation-tracking tools like Connected Papers and Litmaps generate their visual maps.

Adjusting the question. When a Boolean search returns too many results, the right move is usually to narrow the question (replace "social media" with "TikTok") or to apply database-side filters (publication year, document type, peer-review status). When it returns too few results, the right move is to broaden the question (replace "Generation Z" with "youth"), to add a synonym ("body image" → "self-perception"), or to drop a concept entirely. Each adjustment is a small experiment, and the loop returns to Selecting Databases only when the experiment shows that the database is the wrong instrument, not just the wrong string.

The closing image the deck leaves on screen is the whole literature-search workflow in one frame, which is the recap of the workshop.

Caption: The diagram integrates every tool category from Chapters 2 through 5. Left: the four database families (Scholarly Databases subscribed by the Library, Open Access Repositories, Emerging Citation Indexes, and Google Scholar). Right: the four search-strategy moves (breaking the question into key concepts, using Boolean operators, phrase quoting, truncation). Middle: the same Start → Selecting Databases → Developing & Executing → End loop from the previous figure, with its feedback arrow. The diagram makes the loop explicit: the database choice and the search strategy are not sequential; they are revisited together as the search returns evidence.

The mental model that survives the workshop has five layers. The call number is a parsed address, not a random label. The Google Scholar 88% number is the broadest single number, but coverage varies by discipline and no single source is complete. A Boolean query is built from key concepts, synonyms, and three joiners (AND, OR, NOT). A Journal Impact Factor is a journal-level ratio with a specific numerator and denominator, and a citation count is a how-often signal that a citation-class taxonomy can convert into a why-was-this-cited signal. AI research tools are downstream of two open graphs (S2AG and OpenAlex), and the graphs are the auditable substrate, not the AI. The loop connects them, because literature searching is iterative in the way the ACRL frame describes, and the loop is the closing argument of every chapter the deck walks through.

References

1. Framework for Information Literacy for Higher Education — ACRL, adopted January 11, 2016 — https://www.ala.org/acrl/standards/ilframework
2. G 55 Call Numbers — Library of Congress, Policy and Standards Division, February 2018 — https://www.loc.gov/aba/publications/FreeCSM/G055.pdf
3. Google Scholar, Microsoft Academic, Scopus, Dimensions, Web of Science, and OpenCitations' COCI: a multidisciplinary comparison of coverage via citations — Martín-Martín, A., Thelwall, M., Orduna-Malea, E., Delgado López-Cózar, E. Scientometrics 126, 871–906 (2021) — https://link.springer.com/article/10.1007/s11192-020-03690-4
4. Citation Context in Web of Science: Visualizing references by article section and reference classification — Clarivate, May 2022 — https://clarivate.com/webofsciencegroup/wp-content/uploads/sites/2/dlm_uploads/2022/05/202205-WoS-citation-context-final.pdf
5. The Semantic Scholar Open Data Platform — Kinney, R. et al. arXiv:2301.10140 v2, April 2025 — https://arxiv.org/abs/2301.10140
6. OpenAlex: A fully-open index of scholarly works, authors, venues, institutions, and concepts — Priem, J., Piwowar, H., Orr, R. arXiv:2205.01833 (2022) — https://arxiv.org/abs/2205.01833
7. Journal Citation Reports: Document Types Included in the Impact Factor Calculation — Clarivate, JCR Reference Guide June 2024 — https://www.ehu.eus/documents/1738125/2981345/JCR-Reference-Guide-2024.pdf