RESEARCH: FINDING AND EVALUATING SOURCES

Most research papers fail before a single source is read. They fail at the moment a writer drops a topic into a search box, grabs the first five results, and starts hunting for sentences that agree with what they already planned to say. That is not research; that is decoration. Real research is a loop — ask, search, read, revise the question, search again — and it is supposed to change what you think. This chapter walks the loop with you: how to turn an assignment into a question narrow enough to actually answer, where the different kinds of sources live and what each one is good for, how to search a database on purpose instead of pleading with it, how to judge credibility in ninety seconds by leaving the page rather than staring at it, and how to keep notes that turn into an outline instead of a mess.

Section 1From Topic to Researchable Question

A topic is a place; a question is a destination

“Nurse burnout” is a place. You can wander around in it for a week. “Do unpredictable schedule changes predict burnout in medical-surgical nurses more strongly than total hours worked?” is a destination. You can tell when you have arrived, you can tell what evidence would get you there, and you can tell when a source is off the path. The single most common reason a research paper reads like a book report is that the writer never made this conversion.

A research question has three properties. It is answerable with evidence you can actually reach in the time you have. It is arguable, meaning a reasonable, informed person could answer it differently than you do. And it is bounded, meaning it names who, where, when, and what about it. Drop any one property and the paper drifts. A question that is answerable but not arguable produces a summary. A question that is arguable but not bounded produces a survey of everything and a conclusion about nothing.

Notice that “arguable” is not the same as “controversial.” You do not need a culture-war topic. You need a question where the evidence genuinely pulls in more than one direction, which is true of almost any real question once you get close enough to it.

Four levers for narrowing

When a question is too big, you almost never need a new topic. You need to pull one or more of four levers: population (who), setting (where), period (when), and angle (what about it — cause, effect, comparison, evaluation, or definition). Pull them one at a time and watch the question tighten.

Here is the chain, written out the way it actually happens on a legal pad. Start: telehealth. Pull the angle lever: does telehealth improve chronic disease management? Pull population: … for adults with type 2 diabetes? Pull setting: … in rural primary care? Pull period: … since 2019? Add a comparison, which is the single most useful angle in college writing: Does asynchronous telehealth follow-up improve blood-sugar control for rural adults with type 2 diabetes compared with in-person visits, in studies published since 2019?

That final question took ninety seconds and it now tells you exactly what to search for, exactly what to exclude, and exactly what your paper has to deliver. It also tells you what your paper is not about, which is the part that saves you from a 2 a.m. rewrite.

Pressure-testing before you spend three hours

Run five checks before you open a database. Can you state the question in one sentence containing a verb like does, how, why, or to what extent? Could an informed person answer it differently than you would? Can you name the kind of evidence that would settle it — a study, a data table, a memo, a court record? If you cannot picture the evidence, you cannot search for it. Does the question fit the assigned length, remembering that fifteen hundred words answers one question and not four? And, honestly: are you willing to be wrong about it?

That last check is not a motivational poster. If you already know your conclusion, you will unconsciously build searches that confirm it, and your reader will feel the thumb on the scale even if they cannot name why.

Section 2The Source Landscape

Scholarly, trade, popular — the ten-second test

You do not need a course in publishing to sort sources. You need six visual cues. A scholarly source usually opens with a structured abstract, names authors with institutional affiliations, uses the vocabulary of a discipline without stopping to define it, ends with a long reference list, carries a DOI, and contains no advertising. A popular source has a byline without credentials, short paragraphs, no reference list, hyperlinks instead of citations, and images chosen for feeling rather than data. A trade source sits between them: written by practitioners for practitioners, light on citations, heavy on practical detail, and often the fastest place to learn what a field currently worries about.

Ten seconds, six cues. If you are still unsure, scroll to the bottom. The reference list is the tell. Twelve citations to other journals means scholarly. Zero citations and a newsletter signup means popular.

Primary, secondary, tertiary — and why the labels move

A primary source is direct evidence produced by a participant or an original investigator. A secondary source interprets, analyzes, or synthesizes primary evidence. A tertiary source compiles and summarizes the secondary literature. Simple enough — until you notice that the labels shift by discipline, which trips up students constantly.

In literature, the novel is primary and the critical essay about it is secondary. In history, a 1943 ration book, a diary, a census schedule, and a newspaper printed that year are all primary; the historian’s monograph is secondary. In the health sciences, the report of an original study is primary and a review article is secondary. Scholars genuinely disagree about where systematic reviews and meta-analyses belong: many health-science librarians classify them as secondary because they analyze primary studies, while others call them tertiary because, like a textbook, they digest a whole body of work rather than report one investigation. The label matters less than the question you should ask, which is always the same — how many hands has this evidence passed through since it was generated?

Worked example. Your question concerns the Tuskegee syphilis study. The 1972 Associated Press wire story that broke the news is primary. The federal advisory panel’s 1973 report is primary. A 1981 book analyzing the study is secondary. A bioethics textbook chapter summarizing that book is tertiary. If your paper argues about how the story was framed when it broke, only the primary layer will do; a textbook paraphrase cannot support a claim about framing.

Gray literature, government data, and the open web

Gray literature is material produced outside commercial publishing: dissertations, conference proceedings, agency reports, clinical practice guidelines, technical standards, position statements, preprints. It is not lesser. It is often the newest and the most specific thing available, and in fields like nursing, public health, and education, guidelines and agency reports carry real authority. What gray literature lacks is a consistent quality filter, so you supply the filter yourself: who produced it, who funded it, what method is disclosed, and has anyone else engaged with it.

Preprints deserve a sentence of their own. A preprint is a manuscript posted before peer review. It is legitimate to read and sometimes legitimate to cite, but you must label it as a preprint in your prose — “in a preprint that has not yet been peer reviewed, Okonkwo and colleagues report…” — because your reader is entitled to weight it accordingly.

As for the open web: your library’s databases are not a moralistic hoop. They exist because a general search engine ranks by popularity and link structure, while a database ranks by subject indexing within a curated set. Google Scholar sits in between and is excellent for one job in particular, which Section 3 covers. Use all three, but know what each one is optimizing for.

Section 3Searching on Purpose

Build the keyword grid before you type

Databases do not understand sentences. They match strings. So the first move is to break your question into its two to four core concepts, then list every word a different author might have used for each one. That grid is your keyword grid, and building it takes four minutes and saves two hours.

Take the question from Section 1: Does asynchronous telehealth follow-up improve blood-sugar control for rural adults with type 2 diabetes? Four concepts fall out. Concept one, telehealth: telehealth, telemedicine, remote monitoring, e-health, virtual visit, store-and-forward. Concept two, rural: rural, nonmetropolitan, underserved, medically underserved area. Concept three, type 2 diabetes: type 2 diabetes, T2DM, non-insulin-dependent diabetes, NIDDM. Concept four, the outcome: glycemic control, blood glucose, hemoglobin A1c, HbA1c.

Notice what the grid just did. You now know that a paper written in 2014 might say “telemedicine” where a 2024 paper says “virtual care,” and that a search for only one of those terms silently loses half the literature. The grid is where you catch that before it costs you.

From a bad search to a working one

The rewrite below is the whole skill. Read the middle column carefully — the failure is almost never the database’s fault.

Controlled vocabulary: let the index do the work

Big databases assign every article a set of standardized subject terms from a fixed list — controlled vocabulary. PubMed uses MeSH; CINAHL, ERIC, and PsycINFO each have their own thesaurus. When you search the subject heading “Myocardial Infarction,” you retrieve every indexed article on the topic regardless of whether its authors wrote “heart attack,” “MI,” or “coronary thrombosis.” A human indexer already did your synonym work.

Two practical notes. First, find the headings by opening one perfect article and reading its subject-terms field; those are the exact strings to reuse. Second, controlled vocabulary lags. A term coined eighteen months ago has no heading yet, and very recent articles may not be indexed at all. So the mature strategy is both: subject headings to sweep the established literature, free-text keywords to catch what the index has not absorbed. Combine them with OR inside the same concept.

Citation chaining: the fastest route to the good stuff

Citation chaining means using one strong source as a doorway. Backward chaining is mining its reference list: the author already spent a year identifying what matters in that conversation, and the works cited most heavily in the introduction are usually the field’s landmarks. Forward chaining is the opposite direction — finding everything published since that cites it, using the “Cited by” link in Google Scholar or the times-cited count in Scopus or Web of Science.

Worked example. You find a 2019 systematic review on rural telehealth. Backward: its reference list gives you eighty-odd studies, pre-screened for relevance, with the strongest ones discussed by name in the text. Forward: its “Cited by” list gives you everything published since 2019 that engaged with it, including the papers that disagreed — and disagreement is exactly what makes a paper arguable. One good source, two directions, an afternoon of searching replaced.

Section 4Evaluating: Read Laterally, Judge by Stakes

Leave the page: the SIFT moves

Researchers at Stanford compared how undergraduates, professional historians, and professional fact-checkers judged unfamiliar websites. The fact-checkers were both faster and more accurate, and the thing they did differently was almost embarrassingly simple: within seconds they left the page and opened new tabs to read what the rest of the internet said about the site. Students, by contrast, stayed put and graded the page’s design, its domain, and its tone. That is vertical reading. The better habit is lateral reading.

Four moves organize the habit; they are usually called SIFT. Stop: notice that you do not know this source, and notice your own emotional reaction, because content engineered to enrage you is content engineered to be shared unchecked. Investigate the source: spend sixty seconds finding out who publishes it and who funds it, using a search engine and not the site’s own About page. Find better coverage: if the claim matters, look for a stronger source making the same claim, and use that one instead. Trace claims to the original: follow the chain back — blog cites news article cites press release cites the actual study — and read the last link in that chain.

Worked example, ninety seconds. A polished page reports that a supplement “reduces post-shift fatigue by 42 percent.” Stop: you do not know this site, and a single precise-sounding number is doing all the persuading. Investigate: a search on the site’s name plus “funding” shows it is operated by a supplement retailer. Find better coverage: no independent outlet reports the finding. Trace: the page links to a press release, the press release names a study, and the study turns out to be a twelve-person trial with no control group, funded by the same retailer. You did not need to read the page carefully. You needed to leave it.

Interrogating a scholarly source on its merits

Lateral reading is the right tool for unfamiliar web sources. For scholarly sources, where authorship and funding are usually disclosed, a structured interrogation works better. Ask five things. Authority: who wrote it, what is their standing, and does the venue have a review process? Purpose: is this reporting evidence, arguing a position, or selling something — and check the funding and conflict-of-interest statement, which is genuinely informative and takes ten seconds to read. Evidence: what is the method, how large is the sample, and does the discussion section’s confidence match the results section’s numbers? Currency: recent enough for this claim, in this field? Convergence: does anyone else find the same thing?

You may have been taught a checklist like CRAAP — currency, relevance, authority, accuracy, purpose. It is a reasonable scaffold, and specialists genuinely disagree about how far to trust it. The substantive criticism, argued most prominently by digital-literacy researchers, is that a well-funded disinformation site passes every item on the checklist: it is current, relevant, authored, footnoted, and clear about its purpose. Checklists work well on scholarly sources, where the questions have findable answers. They work poorly on the open web, where lateral reading is faster and harder to fool. Use both, and know which situation you are in.

Red flags: predatory journals, content farms, synthetic text

A predatory journal collects a publication fee while performing little or no real peer review. It matters to you because a predatory article can look exactly like a legitimate one in a search result. Warning signs: an unsolicited email praising your “esteemed research” with grammatical errors; a fee disclosed only after acceptance; a promise of peer review in seventy-two hours; a scope that spans medicine, engineering, and management; a title one word away from a well-known journal; editorial-board members who did not know they were listed; and no indexing in PubMed, Scopus, Web of Science, or the Directory of Open Access Journals. Check the DOAJ, check whether your library’s database indexes it, and when in doubt ask a librarian — this is precisely the question they most enjoy answering.

Content farms and machine-generated pages share a different profile: no findable author, or an author with no existence outside that one site; confident prose containing no dates, numbers, names, or links; an answer to your search query in the first sentence followed by four paragraphs restating it; and citations that look impeccable but do not resolve. That last one deserves emphasis. Language models are known to fabricate plausible-looking references — real-sounding authors, real journals, invented titles, invented page numbers, invented DOIs. The defense is absolute and simple: never cite a source you have not personally opened. If you cannot retrieve it, it does not go in the paper.

Match the source to the stakes

Not every claim needs a systematic review. The right question is not “is this a good source?” but “is this a good enough source for this claim?” Background facts that no one disputes can rest on a reference work. Contested numbers need the agency that produced them. Causal claims need the strongest evidence you can reach, and they need you to say out loud what the evidence cannot establish.

Section 5Capturing What You Find

The research log

Keep a research log — a running table with seven columns: date, database, the exact search string you ran, filters applied, number of results, what you kept, and what you would change next time. It takes thirty seconds per search and it pays three ways. It stops you from re-running on Thursday the search you already ran on Tuesday. It lets you reproduce a promising result you half remember. And when an instructor asks how you found a source, you have an answer instead of a shrug.

One entry, written properly: Mar 12 · CINAHL · ("nurse burnout" OR "compassion fatigue") AND ("12-hour shift" OR "shift length") · peer reviewed, 2015–2026, English · 214 results · kept 6, best is the 2022 multi-hospital survey · next time add scheduling predictability as a third concept. That last clause is the valuable one; it is a note from you to you, four hours from now.

Notes in three symbols

Most accidental plagiarism is a note-taking accident, not a moral failure. A student copies a sentence into a document meaning to rework it later, and three days later cannot tell which words are the source’s. Fix it at the point of capture with three labels. Q for quotation: the exact words, inside quotation marks, with the page or paragraph number. P for paraphrase: fully restated in your own sentence structure, with the source and page. M for me: your own thinking, reactions, objections, connections.

Worked example. Sample source sentence, from an illustrative article: “Units that adopted self-scheduling reported a 19 percent drop in unfilled shifts over eighteen months, though the effect narrowed once agency staffing was accounted for.”

Q note: “a 19 percent drop in unfilled shifts over eighteen months” (Rivera 44). P note: Where units moved to self-scheduling, unfilled shifts fell by close to a fifth across eighteen months — but part of that difference traced back to contract staffing rather than to the scheduling change itself (Rivera 44). M note: If agency staffing explains part of the gain, my paper cannot claim self-scheduling alone fixes coverage — check whether the other three studies controlled for it the same way. That M note is a future topic sentence.

Now the failure mode, because seeing it once inoculates you. Patchwriting looks like this: “Units that used self-scheduling reported a 19 percent decrease in unfilled shifts over 18 months, although the effect narrowed once agency staffing was considered (Rivera 44).” A few words have been swapped — adopted to used, drop to decrease, accounted for to considered — while the sentence architecture is untouched. A citation does not rescue it; borrowed structure still requires quotation marks. Chapter 9 takes this apart in detail, but the prevention happens here, in your notes, months before the draft.

The annotated bibliography that earns its keep

An annotated bibliography entry does three jobs in roughly 100 to 150 words: it summarizes what the source claims and how it knows, it evaluates the quality and the limits, and it relates the source to your specific question. Most student annotations do the first job vaguely and skip the other two.

Weak (sample): “This article is about nurse burnout. The authors did a study and found some interesting results. It had a lot of good information and it will be very helpful for my paper.” Nothing here survives contact with a reader. What study? What results? Helpful how?

Strong (sample): “Rivera and Oyelaran surveyed 612 medical-surgical nurses across four hospitals in one state and found that self-reported burnout tracked more closely with unpredictable schedule changes than with total hours worked. The sample is large, but it is cross-sectional and single-region, so the study can show association and not causation, and the burnout measure is self-report. I will use it to complicate the common assumption that shift length is the primary driver, and it is the main reason I narrowed my question from hours worked to schedule stability. It also supplied the search vocabulary I now use — schedule predictability, shift churn.” Still one paragraph. It names the method, the number, the finding, three specific limits, and the exact role the source plays in the paper.

The synthesis matrix, and knowing when to stop

Notes organized by source produce a paper organized by source — the dreaded “Smith says… Then Jones says…” structure that reads like a stack of book reports. A synthesis matrix flips the axis. Sources become rows, your sub-questions become columns, and each cell records what that source says about that sub-question. When the grid is full, read down a column instead of across a row. That column is a paragraph, and the agreements, conflicts, and silences in it are your topic sentence.

Read the predictability column. The two studies that looked at it point the same way, the systematic review never examined it, and the agency report is arguing about something else entirely. You have just written a topic sentence: Where researchers have looked at predictability directly, it outperforms raw shift length — but the policy conversation is still organized around ratios and hours. That sentence came out of the grid, not out of inspiration.

When do you stop searching? Three signals. New searches return names you already have and reference lists you already mined — call it saturation. You can predict what the next article will say before you open it. And you can answer your research question out loud, in a paragraph, without looking anything up. There is also an unromantic fourth signal: the time you budgeted has run out, and a finished paper with six good sources beats an unfinished one with thirty.

Words worth knowing