Quick reference
Writing a literature review takes eight steps: (1) scope a reviewable question, (2) design a reproducible search, (3) chase citations backward and forward from seed papers, (4) screen in two passes against written criteria, (5) build a synthesis matrix, (6) outline by themes — not authors, (7) draft paragraphs that argue and then cite, and (8) close with a precise gap statement that positions your work. Budget 35–70 hours over 4–8 weeks for a typical thesis chapter or journal review.
Why most literature reviews go wrong before the writing starts
A literature review is an argument about a body of research, not a report on it. The drafts that fail — the ones examiners return with "more synthesis needed" and reviewers reject as "descriptive" — almost never fail at the sentence level. They fail upstream: the question was never scoped, so the corpus sprawled; the search was never logged, so the coverage can't be defended; nothing was extracted into a comparable format, so the only available move at the keyboard was summarizing one paper after another. The eight steps below are ordered so that each one manufactures the input the next one needs. Follow them in sequence and the actual writing — the part people dread — becomes the easiest step, because by then the argument already exists in your synthesis matrix and outline.
If you want to see what the finished product looks like in different fields and formats before you start, keep our worked literature review examples open in another tab — the steps below build one.
The 8 steps
- 1Scope the question before you read anything
- 2Design a search strategy you could hand to someone else
- 3Chase citations backward and forward from your seeds
- 4Screen in two passes against written criteria
- 5Build a synthesis matrix before writing a word
- 6Outline by themes, never by authors or years
- 7Draft paragraphs that argue, then cite
- 8Close with a gap statement and position your work
Step 1 · 2–4 hours · output: one sentence + boundary list
Scope the question before you read anything
A literature review fails or succeeds at this step, before a single paper is read. “AI in education” is a topic; “how does adaptive AI tutoring affect math achievement in middle-school students, 2015–2025” is a reviewable question. The difference is boundaries: a defined population, a defined intervention or phenomenon, a defined outcome, and a defined time window. Without them you cannot decide what to include, so you will read forever and synthesize never.
Write the question as one sentence, then write the boundary list underneath it: which fields you will draw from, which years, which study types (empirical only? theory? grey literature?), and which languages. This list becomes your screening criteria in Step 4 — writing it now, before reading biases you, is what separates a defensible review from an ad-hoc one. If you are still choosing between candidate questions, our worked set of research questions by methodology is the fastest calibration.
Concrete example
Weak: “remote work and productivity.” Reviewable: “What does empirical research from 2019–2025 say about how fully-remote arrangements affect individual output and team coordination in knowledge work?” — population (knowledge workers), phenomenon (fully remote), outcomes (output, coordination), window (2019–2025).
The pitfall at this step
Scoping by enthusiasm instead of by answerability. A 12-month thesis cannot review a question that spans four disciplines and forty years. Right-size now: a question you can saturate in 60–150 papers, not 800.
Step 2 · 3–6 hours · output: query log + database list
Design a search strategy you could hand to someone else
Searching is not typing your topic into Google Scholar once. Decompose the question into 2–4 concept blocks, list synonyms for each (authors in different fields name the same thing differently — “remote work”, “telework”, “distributed teams”, “WFH”), and combine the blocks with AND while joining synonyms with OR. Run the strings across at least two databases with different coverage — Google Scholar for breadth, plus a domain database (PubMed, Scopus, ACM DL, EconLit) for precision.
Keep a query log: the exact string, the database, the date, and the rough hit count. This costs five minutes per session and pays off twice — first when a committee member asks “how did you search?”, and second when you return after three months and cannot remember which combinations you already tried. Even a narrative review benefits from a search you could replicate; for a systematic review it is mandatory, which is the core distinction we unpack in our literature review vs systematic review comparison.
Concrete example
Concept blocks for the remote-work question: (“remote work” OR telework OR “work from home” OR “distributed teams”) AND (productivity OR output OR performance OR coordination) AND (2019–2025 filter). Run in Google Scholar + Scopus + an org-behavior database; log all three.
The pitfall at this step
Single-database, single-language searching. Scholar alone over-weights highly-cited English-language work; a domain database surfaces the methodical middle of the field that your synthesis actually needs.
Step 3 · 4–8 hours · output: 30–80 candidate papers beyond keyword search
Chase citations backward and forward from your seeds
Keyword search finds papers that share your vocabulary. Citation chasing finds papers that share your problem — which is a bigger, better set. Pick the 3–5 most central papers you have found so far (your seeds). Backward chasing: read each seed’s reference list and pull everything relevant — this recovers the foundations the field takes for granted. Forward chasing: open each seed’s “Cited by” list on Google Scholar and walk it — this recovers everything the field has done since, including the very recent work whose absence gets manuscripts rejected.
Forward chasing has a scale problem: a well-cited seed has thousands of citing papers, and the list view tells you nothing about where that work comes from. This is where a citation map earns its place in the workflow. Generating a geographic map of a seed author’s citations shows you, in one view, every country and institution engaging with that work. Dense clusters are schools of thought — the labs whose output you must cover. Clusters you did not expect (a strong Korean or Brazilian group you have never cited) are exactly the regional literatures English-keyword search systematically misses, and flagging them strengthens both your coverage and your eventual gap statement.
Run the map for each seed author and compare footprints: where the maps overlap is your field’s core; where one map reaches and the others don’t is a candidate sub-literature to investigate. The full workflow is in our citation map for literature review guide — in practice it takes about 20 minutes per topic and routinely surfaces 10–30 relevant papers that keyword search alone never would.
Concrete example
From four seed papers on adaptive tutoring, backward chasing recovers the 1980s mastery-learning foundations; forward chasing surfaces 2024–2025 LLM-tutor evaluations; the geographic map reveals a substantial Singapore/Korea classroom-deployment literature absent from US-centric review articles.
The pitfall at this step
Chasing only backward. Reference lists end at the seed’s publication date, so backward-only reviews are structurally out of date — the “only 3 sources from the past 5 years” reject comment is almost always a forward-chasing failure.
Step 4 · 4–10 hours · output: final include list (typically 40–150 papers)
Screen in two passes against written criteria
You now have a pool of 150–500 candidates from search plus chasing. Screening turns the pool into a corpus. Pass 1 is title-and-abstract triage: for each candidate, ask only “could this plausibly meet my boundary criteria?” and sort into yes / no / maybe — about 30 seconds per paper, so a 300-paper pool is an afternoon. Pass 2 is full-text screening of the yes/maybe pile against the written criteria from Step 1: right population, right phenomenon, right study type, right window.
Write down a one-phrase exclusion reason for every paper you drop in Pass 2 (“wrong population — undergraduates”, “opinion piece, no data”). This feels bureaucratic for a narrative review, but it is the cheapest insurance in the whole process: when a reviewer asks why an apparently relevant paper is missing, “excluded at full-text: outcome not measured” is an answer, and “I don’t remember seeing it” is a revision request. How many papers should survive? Format-dependent — a master’s chapter synthesizes 40–80 sources, a PhD chapter 100–250; the per-format expectations are tabulated in our literature review examples.
Concrete example
Remote-work pool of 340 candidates → Pass 1 keeps 120 → Pass 2 keeps 74, with logged exclusions like “hybrid not fully-remote (n=18)” and “self-report productivity only, excluded by criteria (n=11)”.
The pitfall at this step
Screening by prestige instead of by criteria. Dropping a relevant paper because the venue is small — or keeping an irrelevant one because it is in Nature — quietly biases the corpus and is exactly what written criteria exist to prevent.
Step 5 · 6–12 hours · output: one spreadsheet row per included paper
Build a synthesis matrix before writing a word
The synthesis matrix is the single highest-leverage artifact in the entire process, and the one most first-time reviewers skip. It is just a spreadsheet: one row per included paper, with columns for the analytic dimensions that matter to your question — typically method, sample/context, key finding, theme, and limitation. Filling it forces you to actually extract from each paper rather than re-read it later, and once filled, it does the synthesis work for you: sort by theme and you see which claims are crowded and which are thin; sort by method and you see whether a contested finding splits along methodological lines; sort by year and you watch the field change its mind.
Expect the theme column to be wrong on the first pass — that is the point. You assign provisional themes as you extract, and after 20–30 rows the real thematic structure of the field announces itself and you re-label. Those final theme labels become the section headings of your review in Step 6, which is why the matrix must come before the outline.
| Study | Method | Sample / context | Key finding (your words) | Theme | Limitation |
|---|---|---|---|---|---|
| Study A (2021) | Natural experiment, diff-in-diff | 16,000 call-center & back-office staff, one firm | Output +8% on independent tasks after full-remote shift | Output effects | Single firm; routine tasks only |
| Study B (2022) | Network analysis of communication metadata | 61,000 knowledge workers, one tech company | Cross-team ties and synchronous communication fell; siloing rose | Coordination costs | Metadata only — no output measure |
| Study C (2023) | RCT, hybrid vs office | 1,600 engineers & marketers | No output difference; attrition down a third in hybrid arm | Moderators / retention | Hybrid, not fully remote |
| Study D (2024) | Panel survey + telemetry | 2,300 workers, multi-industry | Self-reported productivity exceeds telemetry-measured output | Measurement disputes | 18-month horizon |
| Study E (2025) | Qualitative interviews | 45 fully-remote team leads, 3 countries | Leads compensate for coordination loss with heavy process scaffolding | Coordination costs | Small n; leads only, no IC perspective |
Illustrative rows for a review of fully-remote work and productivity. Five rows shown; a real matrix has one row per included paper.
Concrete example
See the sample matrix below — five illustrative rows on the remote-work question. Notice how sorting by theme would immediately group rows 1 and 4 (output effects) against rows 2 and 5 (coordination costs), with row 3 as the moderating-variable thread.
The pitfall at this step
Copying abstracts into the “key finding” column. The finding cell must be in your words, one sentence, comparable across rows — “output up ~8% for solo tasks, coordination latency up for interdependent ones” — or the matrix can’t be sorted into an argument.
Step 6 · 2–4 hours · output: heading skeleton with papers assigned to sections
Outline by themes, never by authors or years
The structural decision that separates a review from a book report: organize by theme, not by paper. An author-by-author outline (“Smith found… Jones found… Lee found…”) produces an annotated bibliography no matter how good the prose is. A theme outline takes the final theme labels from your matrix and promotes them to section headings — better still, headings that make claims (“Remote output gains concentrate in independent tasks” beats “Productivity findings”), because a claim-heading forces every paragraph beneath it to do argumentative work.
Three to six themes is the workable range for most formats. Within a theme, chronology is fine — fields do develop — but the chronology serves the theme, never replaces it. Assign every matrix row to at least one section before drafting; papers that fit nowhere are either a missing theme or a screening mistake, and finding that out now costs nothing. For full heading skeletons by format — dissertation chapter, journal article, conference related-work — see our literature review outline guide.
Concrete example
Remote-work outline: (1) Output effects are real but task-dependent. (2) Coordination and innovation costs accrue with team interdependence. (3) Moderators: tenure, home environment, managerial practice. (4) Measurement disputes: self-report vs telemetry. (5) The gap: longitudinal and non-Western evidence.
The pitfall at this step
Letting one blockbuster paper become its own section. If a section heading is effectively a paper title, the theme is underdeveloped — fold the paper into the claims it supports and contests.
Step 7 · 10–25 hours · output: full draft at your format’s word count
Draft paragraphs that argue, then cite
With the matrix and outline in place, drafting is assembly rather than invention. The unit of work is the synthesis paragraph: a topic sentence that makes a claim about the literature (“Evidence on output is consistent for independent tasks but reverses for interdependent ones”), followed by grouped citations that support, qualify, or contest the claim, and a closing sentence that says what is still unknown. The mechanical test: if you deleted every citation, the remaining sentences should still read as a connected argument. If what remains is disconnected summaries, you are writing the annotated bibliography again.
Calibrate length to format before you start, not after: conference related-work sections run 500–1,200 words; a master’s chapter 4,000–8,000; a standalone journal review 6,000–10,000; a PhD chapter 8,000–15,000. Most of the draft’s word budget belongs in the 2–3 themes closest to your own contribution — a common drafting error is spending 40% of the budget on background a reader could get from any textbook. Write the introduction last; you do not know what you are introducing until the themes are drafted.
Concrete example
Synthesis paragraph in miniature: “Across natural experiments, fully-remote output holds or improves for individual tasks (A 2021; B 2023), while patent- and review-based measures of collaborative work decline (C 2022; D 2024). The two findings conflict only if productivity is treated as one construct — no included study measures both margins in the same workforce, which Section 5 takes up as the central gap.”
The pitfall at this step
Quote-stitching. Direct quotation is rare in strong reviews outside the humanities; paraphrase forces comprehension and keeps the argumentative voice yours. More than one quote per page is a warning sign.
Step 8 · 2–3 hours · output: gap paragraph + positioning paragraph
Close with a gap statement and position your work
The review ends where your contribution begins, and the join is two paragraphs. First the gap statement: one or two sentences naming exactly what the assembled literature has not done — a population not studied, a mechanism not tested, a context not reached, two literatures that never cite each other. Precision is everything: “more research is needed” is the most-flagged sentence in peer review, while “no included study measures both individual output and team coordination in the same workforce over more than one year” is auditable against your matrix and therefore defensible.
Geographic and structural gaps deserve explicit evidence, and this is the second place a citation map does concrete work: if the maps you generated in Step 3 show a literature clustered in North America and Western Europe, you can write “empirical evidence is concentrated in high-income North Atlantic settings” as a verifiable observation rather than a hand-wave. Then the positioning paragraph: name the two or three closest existing studies, state precisely how your work differs — population, method, scale, theoretical lens — and why that difference matters. Committees and reviewers read for this paragraph specifically; without it, even a genuinely novel contribution reads as incremental.
Concrete example
“The closest studies are C (2022) and D (2024), both single-firm and ≤18 months. We differ on scale (multi-industry panel), horizon (3 years), and measurement (telemetry plus output, not self-report) — which lets us test whether coordination costs decay as teams adapt, a question no included study can answer.”
The pitfall at this step
Claiming a gap you never searched for. A “gap” that is actually a hole in your own search strategy is the most embarrassing reviewer catch there is — which is why the gap statement must trace back to the logged search (Step 2) and the chased citations (Step 3).
Where citation maps fit in the process
Two of the eight steps have a geographic blind spot that no keyword search fixes. In Step 3, forward citation chasing at scale is unmanageable as a list — a well-cited seed has thousands of citing papers — but as a map it is legible in seconds: paste a seed author's Google Scholar profile into CitationMap and you see every country and institution citing that work, which clusters define the field's schools of thought, and which regional literatures your English-keyword search never touched. In Step 8, the same maps convert "the literature seems Western-centric" from a hunch into an auditable observation you can write into your gap statement. The full workflow — seeds, maps, cluster reading lists, gap verification — is in the citation map for literature review guide.
Keep going
- Literature review examples — five worked cases by field plus seven format samples, to see what the finished product looks like.
- Literature review outline — heading skeletons for dissertation chapters, journal reviews, and related-work sections.
- Literature review vs systematic review — which form your project actually calls for, and what changes if you go systematic.
- Research question examples — 25 worked questions by methodology, for calibrating Step 1.
Frequently asked questions
What are the 8 steps to writing a literature review?+
(1) Scope the question — define population, phenomenon, outcomes, and time window before reading. (2) Design a reproducible search strategy with synonym blocks across 2–3 databases. (3) Chase citations backward (reference lists) and forward (“Cited by” chains) from 3–5 seed papers. (4) Screen in two passes — title/abstract, then full text — against written criteria, logging exclusion reasons. (5) Build a synthesis matrix: one row per paper, columns for method, sample, finding, theme, limitation. (6) Outline by themes drawn from the matrix, never author-by-author. (7) Draft synthesis paragraphs that open with claims about the literature and group citations beneath them. (8) Close with a precise gap statement and a paragraph positioning your work against the nearest existing studies.
How long does it take to write a literature review?+
Budget 35–70 working hours for a typical master’s or journal-article review, spread over 4–8 weeks: roughly 10–20 hours on searching, chasing, and screening (Steps 2–4), 6–12 hours on the synthesis matrix, and 15–30 hours on outlining and drafting. A PhD dissertation chapter usually doubles that because the corpus is larger (100–250 sources) and the positioning burden is higher. The most common scheduling mistake is allocating almost all the time to “reading” and almost none to extraction and synthesis — reading without a matrix produces familiarity, not a review. If you have two weeks, shrink the corpus at Step 1 rather than skipping Steps 5 and 6.
How do I start a literature review when I know nothing about the topic?+
Start from one recent review article or handbook chapter in the area — not from primary studies. It hands you the vocabulary for your search blocks, the canonical citations for your seed list, and a first draft of the field’s thematic structure. From there, run Step 3 immediately: backward-chase the review’s references for foundations, forward-chase its “Cited by” list for everything published since, and map its lead author’s citations geographically to see where the field actually lives. Within two or three days you will know the major labs, the live debates, and the obvious seeds — which is enough orientation to scope your own question properly and begin the systematic search.
How many sources should a literature review include?+
It scales with format. A conference related-work section cites 30–80 papers; a master’s thesis chapter 40–80; a standalone journal review 80–150; a PhD dissertation chapter 100–250. Within those ranges, depth beats breadth: examiners consistently prefer 50 sources that are grouped, compared, and argued over to 200 listed in passing. A useful self-check is the ratio of synthesis sentences (claims about groups of papers) to summary sentences (claims about one paper) — strong reviews run well above 1:1. Recency matters as much as count: in fast-moving fields, reviewers expect a clear majority of citations from the past five years, with foundations cited selectively rather than exhaustively.
What is a synthesis matrix and do I really need one?+
A synthesis matrix is a spreadsheet with one row per included paper and columns for the dimensions your argument needs — typically method, sample or context, key finding in your own words, provisional theme, and limitation. It is the bridge between reading and writing: sorting the matrix by theme reveals which claims are crowded and which are thin, sorting by method reveals whether disagreements split along methodological lines, and the final theme labels become your section headings. You need one for any review above about 25 sources. Skipping it is why drafts collapse into paper-by-paper summary — without extracted, comparable findings, the only thing you can write about each source is what it said.
What is the difference between a literature review and a systematic review?+
A (narrative) literature review is an argued synthesis: the author selects, organizes, and interprets prior work to motivate a position or a study, with judgment applied at every step. A systematic review follows a pre-registered, reproducible protocol — explicit eligibility criteria, exhaustive searches across named databases, dual screening, risk-of-bias assessment, and a flow diagram accounting for every record. Systematic reviews aim to eliminate the author’s selection bias; narrative reviews accept and disclose it. The eight steps on this page serve both — a systematic review essentially formalizes Steps 2 and 4 and registers them in advance. Most thesis chapters and journal introductions call for the narrative form.
Should I write the literature review before or after designing my study?+
Mostly before, finished after. The review has to come first in substance because its gap statement is the justification for your study design — population, method, and measures should all trace to limitations you documented in prior work. But the written chapter is rarely finished until late in the project: new papers appear while you collect data (re-run your forward chase before submission), and your framing sharpens once you know your results. The practical rhythm is: complete Steps 1–6 and a rough draft before committing to a design; freeze the search; then do a final updating pass — new forward-chase, matrix additions, light re-draft — in the month before you submit or defend.
Can I use AI tools to help write my literature review?+
For some steps, carefully; for citations, never without verification. AI assistants are genuinely useful for generating synonym blocks in Step 2, triaging abstracts in Step 4, and suggesting candidate theme labels for your matrix. They remain unreliable for the citation layer — fabricated papers, wrong authors, and plausible-but-fake DOIs persist in 2026, and journals now routinely screen for them. The safe rule: AI can propose, but every citation that lands in your bibliography must be verified against the indexed original, and every synthesis claim must trace to a matrix row you extracted yourself. Tools that operate on real indexed records — Google Scholar, Connected Papers, ResearchRabbit, CitationMap — are safer than open-ended chat for discovery.