How to write a literature review, step by step

Decide which kind of review you are writing

• Narrative review — the review chapter of a thesis or the background section of an article. It selects and interprets the literature to motivate a study; it must be fair but need not be exhaustive.
• Systematic review — answers a precise question with a reproducible search, explicit inclusion criteria and a quality appraisal of every study, usually reported with the PRISMA 2020 statement. The Cochrane Handbook is the standard reference for health research.
• Scoping review — maps how much research exists on a broad topic and what kind, without judging study quality in depth.
• Meta-analysis — a systematic review that pools the studies' numbers statistically.

Most students and early-career researchers write a narrative review. The steps below work for it, and the systematic-review additions are marked.

Step 1: turn the topic into a question

A review of “machine learning in agriculture” has no end. A review of “how accurate are smartphone-based models at detecting leaf diseases in field conditions?” tells you what to search, what to include and what to conclude. Write the question down and fix the scope: years, languages, study types and populations.

Step 2: search systematically, even for a narrative review

• List the key concepts of your question and two to five synonyms for each; combine synonyms with OR and concepts with AND.
• Search at least two databases: Scopus or Web of Science for breadth, a subject database such as PubMed or IEEE Xplore for depth, and OpenAlex or Google Scholar to catch what they miss.
• Follow citations both ways from the best papers: their references (backward) and the papers that cite them (forward).
• Record each search — database, date, exact query, number of results. A systematic review must report this; a narrative review benefits from it when a reviewer asks how you searched.

The AI research assistant answers questions with citations to real papers from OpenAlex, which is a quick way to find starting points and search terms — but read the papers it cites before you rely on them.

Step 3: screen what you found

Remove duplicates, then screen titles and abstracts against your criteria, then read the full text of what remains. Note why each excluded paper was excluded. In a systematic review two people screen independently and the numbers at each stage go into a PRISMA flow diagram.

Step 4: extract into a matrix

Make a table with one row per source and the same columns for every one: aim, method, sample or data, main finding with its numbers, limitations, and how it relates to your question. The matrix is where patterns appear — which methods dominate, where results disagree, what nobody has tested. A summarizer can draft a first pass of each row from the paper's text; check every number against the paper.

Step 5: synthesise by theme, not by paper

The commonest weakness of a literature review is a string of paragraphs that each start with an author's name. Instead, organise the sections by theme, method or debate, and open each paragraph with a claim the evidence supports:

• Weak: Smith (2021) used a CNN on lab images. Lee (2022) used transfer learning. Rao (2023) tested phones in the field.
• Strong: Accuracy reported on laboratory images drops sharply in field conditions: models above 95% on curated datasets fell to between 60% and 80% on phone photos (Lee, 2022; Rao, 2023), and only Rao tested lighting variation directly.

The numbers in that example are illustrative; yours come from your matrix. Each paragraph should compare, contrast and evaluate sources, and say what the comparison means for your question.

Step 6: write it up

1. Introduction — the question, why it matters, the scope and how you searched.
2. Body — the thematic sections, from the broad and settled to the specific and disputed.
3. Gap and conclusion — what is known, what is contested, what is missing, and how your study (or future research) addresses it.

Cite as you write and keep references in a manager or a .bib file; the BibTeX cleaner removes duplicates and flags missing fields before submission.

Mistakes to avoid

• Summarising papers one by one with no synthesis.
• Citing only what supports your hypothesis.
• Relying on old sources in a fast-moving field, or on secondary citations you never read.
• Patchwriting — copying sentence structure from the sources with a few words changed. It is plagiarism even when cited; see how to paraphrase without plagiarism.

Before submitting, run the review through the plagiarism checker: literature reviews are the section most often flagged, because they discuss the same sources as everyone else.