Beyond the Summary

How to Plan, Structure, and Place a High-Impact Review Article

Soorathep Kheawhom

9 September 2026

Review articles earn about three times
the citations of a research paper.

Most reviews are never cited at all.

Both are true. Writing a review is not a strategy. Writing the right review, at the right moment, is.

The two numbers

Why this session exists

Average

2.95×

Citation multiplier of a review over an original research article, across all fields. Range 1.34× to 6.74× by discipline.

Reality

21%

Share of 538 reviews published in one field that had passed ten citations two years later. The other 79% sank without trace.

Miranda & García-Carpintero, J. Informetr. 12, 1015 (2018) · Ketcham & Crawford, Lab. Invest. 87, 1174 (2007)

Five questions

Today

20 min What kinds of review exist, and which one is mine?
6 min What does a review do to my academic profile?
12 min When should I write one, and what counts as a contribution?
15 min What is the process, and where do AI tools fit?
16 min What does the finished paper look like, down to the title?

Two short exercises on paper. Take-away: a one-page proposal template you can fill in this week.

Part 1 · The genre map

Ten different things are called a review. This part ends with a three-question test that tells you which one is yours.

“Review” is a journal label, not a method

Part 1

What the journal calls it

Review · Critical Review · Tutorial Review · Minireview · Focus · Perspective · Survey · Roadmap · Opinion · Viewpoint

Commercial format names. They set word counts and page limits. They tell you almost nothing about how to do the work.

What you actually have to do

Narrative · Systematic · Scoping · Mapping · Critical · Integrative · Meta-analysis · Umbrella · Rapid

Methodological types. They determine your search protocol, your screening burden, and whether your claims are defensible.

Choose the method first, then find a venue whose format fits it. Never the reverse.

What a review article actually is

Part 1

If you have never written one, start by looking at the object, not at the definition.

There are no new experiments in it. Everything already exists somewhere in the literature. What does not exist yet is the arrangement, and the arrangement is what you contribute.
The figures are drawn, not measured. The most reused item in most reviews is a single conceptual figure summarising the framework. Budget weeks for it, not an afternoon.
The section headings are the contribution. The numbered headings of the body are the taxonomy you invented. If they could be lifted from any other review of the topic, you have not contributed anything yet.

Anatomy of the object: Title → Abstract → Introduction and prior-review table → your framework → body organised by that framework → challenges and outlook → 100+ references. The two bold sections are what make it a review rather than a summary.

Three things that get taught wrong

Part 1

Myth

“A systematic review is better than a critical review.”

Systematic describes how the evidence was found. It buys transparency and reproducibility. It does not buy insight.

Myth

“Meta-analysis is the highest level of review.”

It is a statistical step used inside a systematic review when outcomes are commensurable. Not a rung above anything.

Myth

“A critical review is just an unsystematic review.”

Its value is the reasoning: testing assumptions, locating contradictions, separating correlation from mechanism.

Being systematic tells you how reliably the literature was collected. Being critical tells you what new understanding was created from it.

SALSA: the four dials behind every review type

Part 1

Search

How much of the literature did you look for, and can someone repeat it?

Appraisal

Did you judge the quality of what you found, or accept every paper equally?

Synthesis

How did you bring the studies together: narrative, tabular, or statistical?

Analysis

What new understanding did you produce that no single paper contains? This is the dial people skip, and the only one that produces a contribution.

Grant & Booth, Health Info. Libr. J. 26, 91 (2009); extended to 48 review types by Sutton et al., ibid. 36, 202 (2019)

One topic, four completely different papers

Part 1

Same literature, same author. The type is a decision about which question you are answering.

Narrative

“Artificial Interphases for Aqueous Zinc Anodes: Progress and Prospects”

Answers: what has the field been doing?

3–4 months · ~120 refs · one author

Scoping

“Mapping the Evidence for Artificial Interphases on Zinc Anodes: What Is Measured, and What Is Not”

Answers: what evidence exists, and where are the holes?

5–6 months · documented search · two screeners

Critical

“Cycling Stability of Coated Zinc Anodes Is Set by Test Conditions, Not by Coating Chemistry”

Answers: is the field’s central claim actually supported?

6–9 months · ~90 refs · a thesis and a tier scheme

Systematic + meta

“Do Artificial Interphases Extend Zinc Anode Life? A Systematic Review and Conditional Meta-Analysis”

Answers: how large is the effect, and under which conditions?

9–12 months · protocol first · two or more people

None of these is more advanced than the others. They answer different questions and they fail in different ways.

Five types built around evidence

Part 1

Narrative — “I have read this field for years, and here is what I think matters.” Low cost. But you may never claim that the evidence shows anything.
Systematic — “I searched by a written protocol, and here is everything that met my criteria.” Six to twelve months, two people minimum.
Scoping — “Here is the shape of the evidence, and here are the holes in it.” High search, low appraisal. A realistic first review.
Mapping — “Here is a classification of the field, usually with a database others can search.” The output is the map itself.
Meta-analysis, a method and not a type — “Inside a systematic review: I pooled the numbers into one effect size.” Only possible when outcomes are commensurable.

Narrative versus systematic is not a ranking of quality. It is a ranking of what you are allowed to claim.

Five types built around an argument

Part 1

Critical — “The field believes X. I will show you why that is only partly right, and what to believe instead.” The argument is the work.
State-of-the-art — “Here is what is actually known now, covering the last five to ten years completely.”
Tutorial — “Here is how this method really works, explained for someone who has never used it.” Clarity outranks completeness.
Perspective / Opinion — “Here is where I think this field should go next.” Little writing, but it trades on reputation. Almost always invited.
Umbrella — “Six reviews on this topic disagree. Here is what they collectively show.”

Realistic first review, with a senior co-author: a scoping review, a critical review, or a tutorial review.

Two axes, not a ladder

Part 1

Low systematic · high critical

Critical review

Insight, no documented search

High systematic · high critical

Systematic critical review

Documented search, then real synthesis. The strongest combination available to a junior author.

Low · low

Narrative review

Expert commentary, no protocol

High systematic · low critical

Descriptive systematic review

Rigorous search, thin synthesis

Where does meta-analysis go? Nowhere on this map. It attaches to the right-hand column when the outcomes are commensurable. It sharpens a quantitative claim; it does not move a review upward. And the bottom right is not a failure: a scoping or mapping review belongs there by design.

Do not import the clinical evidence hierarchy

Part 1

Clinical / medical

“Does treatment A improve outcome B compared with C?”
Population and dose vary, inside a largely shared protocol
The unit compared is a pre-defined clinical outcome
Synthesis narrows the uncertainty around an effect
Pooling is often meaningful

Materials / engineering

“Why does this system behave this way, when does it work, and when does it fail?”
Loading, current density, cell format, electrolyte, cut-off criterion. Effectively everything varies
The unit compared is a number whose test conditions differ study by study
Synthesis produces a mechanism that explains why results diverge
Pooling is sometimes precise and physically meaningless

The right review methodology is set by the scientific question, not by a hierarchy borrowed from another discipline.

So which one should you write?

Part 1

1 Can you state a claim about your field that an expert would argue with? No, not yet → go to 2. This is the normal starting position, not a failure. Yes → go to 3.

2 Is your goal to map what exists and find the gaps? Yes → scoping or mapping review. The right first review: defensible, useful, and it tells you whether a critical review is possible later. No → narrative review, fine as a thesis chapter or grant background.

3 Are the outcomes comparable enough to pool or normalise? No → critical review. Add a documented search and it becomes a systematic critical review. Yes → systematic review, with a meta-analysis only if pooling survives scrutiny.

Three questions, four honest answers. None of them is “write the most comprehensive review you can”.

Part 2 · Your academic profile

The honest version: what it buys you, what it costs the field, and what it will not do for your promotion case.

Reviews are over-represented at the top

Part 2

Among the top 0.05–0.1% most-cited papers, the review share rose from roughly 17% in 1990 to about 40% in 2010.
Reviews count as citable items in the impact-factor denominator, exactly like research papers, so a review-heavy portfolio raises a journal’s IF mechanically. Seven of the ten highest-IF journals in the 2025 JCR are review journals.
The multiplier is largest in fields where reviews are rarest. In review-saturated fields the premium collapses.

The premium is real, and it is strategic: it is largest exactly where nobody has reviewed yet.

Miranda & García-Carpintero, J. Informetr. 12, 1015 (2018); the 1990 share is reported there as 16–18% · 2025 Journal Citation Reports

The other side of the ledger

Part 2

38% — median reduction in future citations to a primary paper after a review covers it. About 70% of reviewed papers lose citations. Analysis of ~6 million articles, 1990–2016.

Your review absorbs credit that belonged to other people. Cite the specific primary paper for a specific claim; cite the review only for context.
Referees are increasingly hostile to reviews with no contribution. Editors in energy and materials now publish editorials asking authors to stop writing unnecessary reviews.
Volume has consequences. Systematic reviews grew 2,728% between 1991 and 2014; roughly a third were judged redundant.

McMahan & McFarland, Am. Sociol. Rev. 86, 341 (2021) · Ioannidis, Milbank Q. 94, 485 (2016) · Kamat, Christopher & Jin, ACS Energy Lett. 10, 5020 (2025)

What a review genuinely buys you

Part 2

Agenda setting. You decide which papers become the canon of your subfield. The largest and least-discussed benefit.
A research identity. A review is where you name and defend the framework your group works under.
Invitation flow. Editors commission people who have already written well on the topic. The first review is what makes the second one an invitation.
Reusable infrastructure. Grant proposal backgrounds, a course module, and the fastest way to onboard a new student.
Low-risk collaboration. A multi-institution review often precedes a joint experimental project.

What a review will not do for you

Part 2

Not a substitute for primary research

A record heavy on reviews and thin on experimental papers reads as a group without a working laboratory.

Not a substitute for expertise

Referees notice reviews assembled from abstracts, and abstract-only reading propagates other people’s errors under your name.

Not faster than a research paper

Six to twelve months, 100+ references, plus revision rounds.

The pattern that works for early-career authors: junior first author, senior sponsor. A study of 1,102 reviews found junior researchers were the predominant authors and served mainly as first authors. You are senior enough to be first author. You are probably not senior enough to be sole author.

Chang & Wu, Scientometrics 130, 4995 (2025) — carried out in library and information science, so indicative rather than proven for engineering

Part 3 · Go / No-Go

Five gates, the contribution ladder, and the question everyone asks: what am I contributing if I have no new data?

Gate 1 and Gate 2: is the field ready, and is it taken?

Part 3

Gate 1 · Field maturity

30+ relevant primary papers in the last 2–3 years means the topic is timely
15–20 within the last 5 years is the absolute minimum
< 50 significant papers overall and the topic is probably too immature
> 120 and your scope is too broad. Narrow it, or split the paper

Gate 2 · Review saturation

The editorial rule at Chemical Reviews: the topic must not have been comprehensively reviewed in the past 3 to 4 years. Every RSC proposal form asks you to justify why there is room for another.

Run the review search before the primary search. In Scopus: your terms AND DOCTYPE(re) AND PUBYEAR > 2020, then repeat with DOCTYPE(ar) for the last three years.

Those two counts give you Gate 1 and Gate 2 in fifteen minutes.

Gates 3 to 5: standing, thesis, capacity

Part 3

3 Standing Do you have primary research papers on this topic from the last three years? RSC and ACS proposal forms require them by DOI. If not, your route is a co-authored review with someone who has them.

4 A thesis Can you state, in one sentence, a claim an expert would argue with? “Advances in X” is not a thesis. “The reported cycle-life gains in X are an artefact of test protocol, not of materials chemistry” is a thesis.

5 Capacity Six to twelve months, a co-author, and a student who will not be derailed. A review is a project, not a gap-filler between experiments.

The contribution ladder

Part 3

Every rung is still called a review. Only the lower half of this list is a contribution.

1. Collection “What papers exist?”
2. Classification “How can they be grouped?”
3. Comparison “How do they differ?”
4. Synthesis “What does the evidence collectively say?”
5. Critique “What is unreliable or missing?”
6. Reconciliation “Why do these studies disagree?”
7. Framework or mechanism “Can one concept account for it all?”
8. Design rules “What should we build or test next?”

The referee’s test: remove the reference list. What is left? A review with 500 references can sit on rung 1; a critical review with 80 can sit on rung 8.

What counts as novelty in a review?

Part 3

Novelty in a review is not new data. It is new structure imposed on data that already exists.

A taxonomy or framework

You name the categories the field then uses to talk about itself.

An evidence standard

You define how claims should be graded, and re-score the corpus.

A reconciliation

You explain why two bodies of literature that disagree both appear to be right.

A quantitative synthesis

You normalise incomparable reported conditions into one comparable dataset.

A negative finding about the field

You show that a widely believed claim rests on very thin evidence.

A translation

You connect two communities that work on the same problem and never cite each other.

Not a contribution: a newer time window on its own · more references than the last review · a different material system · a bibliometric map with no question behind it. The naming test: if your contribution cannot be given a name another author could put into their own paper, it is not a contribution yet.

Exercise 1 · Write two lines on paper

Line 1

One sentence stating a claim about your topic that an expert would argue with.

Line 2

The titles of the three most recent reviews on your topic, and one clause each on what they missed.

If you cannot write line 1, you do not yet have a critical review. If you cannot write line 2, you have not done the gap analysis.

Part 4 · The process

Eight steps from idea to submission, and where AI tools legitimately fit into them.

Eight steps, and where the time actually goes

Part 4

1. Scope the question in PCC form, plus what is excluded and which years.
2. Gap analysis the prior-review comparison table, built first.
3. Proposal two to five pages to the editor before you write anything.
4. Search multi-database, documented, validated against a seed set.
5. Screen and extract two screeners, a written rule, one template.
6. Appraise grade the evidence. This is where reviews win.
7. Synthesise concept matrix, not a paper-by-paper walk.
8. Write figures first, then the prose around them.

Steps 2, 3 and 6 are the ones beginners skip, and the three that decide whether the review is publishable.

Step 2 · The prior-review comparison table

Part 4

Build this before you write anything else. It is the artefact that gets a proposal accepted.

Prior review — author, year, journal. One row each.
Window — the years it covered.
Organising logic — materials-by-materials, application-driven, bibliometric, and so on.
What it does not do — the specific omission, not a vague complaint.
What we add — the column that matters.

If you cannot fill column 5 with something specific, do not write the paper. A newer time window is not enough on its own, and referees will say so.

Step 3 · The proposal

Part 4

At Chemical Reviews, a manuscript sent without an invitation or an approved proposal is not reviewed. It is returned.

Chemical Reviews (ACS) · max 5 pages

Title, authors and affiliations · a detailed topical outline of 2–3 pages · a list of previous reviews · five closely related papers by the authors · estimated reference and page count · tentative submission date · at least five suggested referees

Decision usually within two months.

RSC journals · under 2 pages

Proposed title, authors, submission date · why the field matters now · comments on other reviews on a similar topic, justifying why there is room for another · section outline with ten or more core references with DOIs · a CV of three pages or less highlighting your last-three-years papers

Acceptance of a proposal does not guarantee publication: full peer review still applies.

Step 4 · The search: reproducible, or it does not count

Part 4

Databases, in this order

Scopus and Web of Science Core Collection as the two primary systems · Engineering Village (Compendex + Inspec) for the only real controlled vocabulary in engineering · Google Scholar as a supplement only — not reproducible, unsuitable as a principal search system

Grey literature matters in engineering and is invisible to Scopus: standards (ASTM, IEC, IEEE), patents, national-laboratory reports, preprints.

Record this, per database

Platform and interface · the full search string verbatim · the date you ran it and the coverage years · every filter · number of records retrieved · who designed and who checked the string · forward and backward citation chasing as a separate arm · deduplication counts before and after

That is the PRISMA-S checklist, condensed. Put it in the supplementary information.

Validate before you screen: pick 10 to 30 papers you already know must be included. If your string misses any of them, the string is not finished.

PRISMA 2020 and PRISMA-S: prisma-statement.org · Gusenbauer & Haddaway, Res. Synth. Methods 11, 181 (2020)

Step 5 · Screening and extraction

Part 4

1. Rayyan free, fast blinded dual screening. Weak deduplication, opaque AI ranking.
2. Covidence end-to-end: dedupe, screen, extract, PRISMA diagram. Paid.
3. ASReview free; active learning ranks likely records first. Prioritisation is not exclusion.
4. Elicit and similar a second extractor, for factual fields only.
5. Connected Papers, ResearchRabbit orientation. Not reproducible, so not a search arm.
6. Zotero / EndNote references and dedup. Not a screening system of record.

Design the extraction template from your framework, not from the papers. Decide what every included study must tell you before you open the first PDF.

AI tools: where they legitimately help

Part 4 · Step 5b

1. Scoping and gaps citation-mapping tools surface clusters your string missed. Never a search arm.
2. Search strings litsearchr or an LLM for synonyms. Never skip the seed-set recall check.
3. Screening ASReview or an LLM pre-filter can rank, or be a declared second screener. Never the sole screener.
4. Data extraction Elicit pre-fills factual fields. Never interpretive fields: accuracy collapses.
5. Claim checking Scite shows whether a claim is contested. Never accept it unread.
6. Language clarity of sentences you wrote. Never synthesis prose or citations.

The three-line ethical test

Provenance. A primary source for every claim, because you read it. Accountability. Every sentence defensible as your own. Disclosure. Tool, model version, date and stage in the manuscript.

Two hard prohibitions

Never paste a manuscript you are refereeing into a public AI tool: a confidentiality breach at every publisher. Never submit a citation you have not opened.

Step 6 · Appraisal turns a list into a review

Part 4

Almost no engineering review does this. It is the cheapest available way to be distinctive.

A worked example: evidence tiers

Tier 0 — claim asserted; conditions not reported
Tier 1 — single measurement, idealised conditions
Tier 2 — replicated, conditions fully reported
Tier 3 — independently reproduced under a stated protocol

Publish the tier assignment for every study. It becomes the table people cite.

Clinical appraisal logic, translated

Risk of bias — uncontrolled test conditions, author-run comparisons against untuned baselines
Indirectness — coin cell versus pouch, lab coupon versus field component
Imprecision — n = 1, no replicates, no error bars
Publication bias — the record-efficiency ratchet: only improvements are ever published

From the GRADE downgrade domains. Borrow the logic, drop the clinical vocabulary.

If your review treats a single unreplicated result and a multi-lab reproduction as equal evidence, it is not a review. It is a bibliography.

Step 7 · Concept matrix, not a paper-by-paper walk

Part 4

Author-centric — wrong

“Zhang et al. reported X. Subsequently, Lee et al. found Y. In a related study, Kumar et al. observed Z…”

One paragraph per paper. “Subsequently”, “meanwhile”, “in another study”. The reader learns what happened, in what order, and nothing else.

Concept-centric — right

“Three mechanisms are invoked for the observed capacity fade. The first, supported by Tier 2 evidence in four independent studies, is… The second is asserted widely but rests on a single unreplicated measurement…”

One paragraph per concept, several papers per paragraph, and a verdict at the end of each.

Build a matrix with concepts as rows and studies as columns. Every body paragraph is one row. A row with one entry is not a concept yet; a column in no row is a paper that does not belong.

Part 5 · Structure and craft

What the finished paper looks like, how to title it, what to budget, and the seven ways it goes wrong.

The anatomy of a review article

Part 5

1. Title specific. It carries the framework, not just the topic.
2. Abstract a miniature of the review, not an introduction. It must state the conclusions.
3. Introduction problem, gap, hook. The scope, the years, and the prior-review table.
4. Framework your taxonomy or tier scheme, in its own numbered section. This is the contribution.
5. Body organised by the framework. One concept per section, a verdict at the end of each.
6. Cross-cutting standards, datasets, scale-up, safety, cost, reproducibility.
7. Challenges and outlook three to five named challenges with actionable milestones.
8. Conclusion short. What the field should now believe that it did not before.

Sections 4 and 7 distinguish a review from a summary. Draft them first.

Four ways to organise the body

Part 5

Function-centric

By what the work does: synthesis, characterisation, modelling, deployment. Good when methods differ more than materials do.

System-oriented

By layer: electrode, electrolyte, interface, cell, pack, system. Good when the field has a natural physical hierarchy.

Application-driven

By end use: grid storage, transport, portable, off-grid. Good when the same technology faces very different requirements.

Framework-based

You propose a conceptual framework and organise everything under it. Highest risk, highest reward, and the only one of the four that produces something nameable.

Mutual-exclusivity test: could a reader who has seen only your section headings predict which section any given paper belongs in?

The title is the first test of whether you have a thesis

Part 5

Left: could sit on any review of the topic. Right: could only sit on yours.

“Recent Advances in Zinc-Ion Battery Cathodes”

“Why Capacity Fade in Zinc-Ion Cathodes Is a Dissolution Problem, Not a Structural One”

“A Review of Machine Learning for Materials Discovery”

“Machine Learning for Materials Discovery: Five Benchmarks That Do Not Test Generalisation”

“Progress and Perspectives on Solid Electrolytes”

“Reported Critical Current Densities in Solid Electrolytes: A Tier-Based Reassessment”

A title that would fit on any review of the topic is not a title yet. Put the claim, or the framework, into it.

Budgets, and the three tables that get you cited

Part 5

Rules of thumb

~25 references per 1,000 words
max 7 display items in a Nature Reviews-style Review
100–120 references for a comprehensive engineering survey
3–5 named challenges in the outlook
1 framework, defined once, used everywhere

Exceeding a journal’s stated limits is a desk-rejection trigger, not a negotiating position.

The three tables

The prior-review table — what existed, what it missed, what you add. Editors read it first.

The evidence table — every included study, its conditions and its tier. The artefact other authors reuse, and reuse is citation.

The gap table — what is not known, why, and what experiment would settle it.

Seven ways reviews get rejected

Part 5

1. The laundry list one paragraph per paper, joined by “subsequently”.
2. No thesis the reader cannot say what the authors think.
3. Undefined scope no statement of what is included, excluded, or which years.
4. Cherry-picked evidence no criteria, so nobody can tell what was left out.
5. Self-citation bloat your own work singled out. Referees always notice.
6. Stale coverage the newest references are two years old.
7. Generic outlook “Further research is needed.” Never helped anyone.

Every one of these is visible in the first two pages. That is where desk rejections happen.

Exercise 2 · Locate your own review

Task 1

On the two-axis map, mark where the review you are planning sits today. Be honest about the horizontal axis.

Task 2

Apply the referee’s test to it. Remove the reference list. Write down what is left.

If the honest answer to Task 2 is “a well-organised list”, you now know exactly which rung you have to climb, and Part 4 told you how.

Generative AI: what you must disclose in 2026

Part 5

Universal across ACS, RSC, Elsevier, Springer Nature, Wiley, IEEE, ICMJE and COPE

AI cannot be an author
Authors remain fully accountable for every sentence and every citation
Referees must not upload manuscripts to public AI tools
Grammar and spell checking normally needs no declaration

Where the declaration goes — Elsevier: immediately before the reference list. ACS and IEEE: the Acknowledgement. Springer Nature: Introduction or Acknowledgements. ICMJE: also in the cover letter.

Specific to review articles

ACS states that editors may reject or require revision where AI use appears excessive, and names generated literature reviews explicitly. Publishers now diverge on whether AI-assisted literature reviews are permitted at all.

The three highest-risk uses are AI-written synthesis prose, AI-suggested citations, and AI-only screening with no human verification.

Disclose the tool, the model and version, the date, and the exact stage. Then state that every citation was verified against the primary source. Because you did verify them.

If you remember three things

A review is an argument, not a summary. If an expert could not disagree with your central claim, you have not written a review.
The proposal comes before the manuscript. Two pages, sent to an editor, decides whether the next twelve months are worth spending.
Grade your evidence. The cheapest way to be the review people cite instead of the review people skim.

Take-away: the one-page proposal template in your handout. Fill it in this week, while the topic is still hot.

Questions

Sources and further reading

1. Typology Grant & Booth (2009) · Sutton et al. (2019)
2. Reporting prisma-statement.org · synthesismanual.jbi.global · roses-reporting.com
3. Engineering method Kitchenham & Charters, EBSE-2007-01 · Niyato et al., arXiv:2509.25828
4. Searching Gusenbauer & Haddaway (2020) · Gusenbauer & Gauster (2025)
5. Appraisal gradepro.org/handbook · Shea et al., AMSTAR 2 (2017)
6. Craft Pautasso (2013) · Webster & Watson (2002) · Nat. Rev. Bioeng. 2, 907 (2024)
7. The debate Miranda & García-Carpintero (2018) · McMahan & McFarland (2021) · Kamat et al. (2025)

Download the participant handout (PDF) · soorathep.k@chula.ac.th