In the last post I talked about choosing a topic and drafting an outline. This one is about what I actually did once the outline existed: the literature review. Not the idea of a literature review. The tools I opened, the order I used them in, and how I went from a blank Introduction section to paragraphs with references.
A caveat before I start. I am not trained in formal literature review methodology. I’ve never done a systematic review; I don’t do PRISMA anything, and what I’m about to describe is a working process I put together out of necessity. It gets me to a written Introduction that I can defend. If you’re doing a medical meta-analysis, this is not the workflow for you. For my master’s thesis, I hope it’ll be enough.
Here’s the whole thing laid out first, then I’ll walk through the parts.
Start broad, on purpose
I start with Google. Plain Google, not Google Scholar. I type whatever is in my head about the topic; sometimes a phrase, sometimes a question, sometimes just three loose keywords. The point isn’t to find the perfect paper. The point is to see what the internet thinks this topic is about: which terms are used, which adjacent problems come up, and which phrases people who already work on this use when they write.
This step feels unscientific, and it is. But it’s useful because at the start of a literature review, I usually don’t know the vocabulary of the field well enough to search properly. A broad Google search gives me that vocabulary in about fifteen minutes. After that, Scholar searches are much sharper because I’m using the right words.
From there, I move to the three places where actual papers live: Google Scholar, ResearchGate, and academia.edu. Scholar is the main one — it’s the most complete, and most of what I end up using comes from there. ResearchGate and academia.edu are useful mostly because sometimes they have a PDF of a paper I can’t otherwise access, and because authors upload their own preprints and related working papers that don’t always show up on Scholar.
I write down every sub-question I can think of as its own keyword search. Not just the main topic, but everything adjacent. Why does this problem exist? What alternative methods have been used, what datasets are common, and what does the evaluation usually look like? Each question becomes one or two searches. The goal at this stage is to build a pool of 10–20 candidate papers. Not to read them in detail, just put them to the side.
Pick an anchor paper
Once I have that pool, I look for one paper to anchor the review. The criteria I use, roughly in order:
1. Published in a Q1 journal. Q1 means the journal sits in the top quartile of its field by impact factor. It’s not a perfect proxy for quality, but it’s a reasonable filter. Q1 papers undergo stricter peer review, and the field tends to take them seriously.
2. Recent. Ideally, within the last one to two years. The anchor needs to reflect the current state of the problem, not where it was a decade ago.
3. Well-cited, or at least well-received. If it’s very recent, the citation count will be low by default, so this is soft. But a recent paper with an unusually high citation count for its age is often a good sign.
4. Close to my topic, but not identical. The anchor doesn’t have to cover exactly what I’m doing. It needs to be in the same neighborhood. Similar problem, similar methods, similar data; so that its references and the papers around it will be relevant to me.
The anchor paper matters because of what comes next.
Put the anchor in Connected Papers
Connected Papers is a tool that takes a seed paper and builds a visual graph of related papers. Both papers it cites and papers that cite it, plus papers that are bibliographically near it, even if there’s no direct citation link. You put my anchor paper in, and it gives me a visual cluster of work around that topic that I can navigate.
The default graph view is useful but busy. The two tabs I rely on more are Prior Works and Derivative Works. Prior Works show the foundational papers that the anchor is built on. The classics in the field that I probably need to have read, or at least know about. Derivative Works shows recent papers that extend or build on the anchor, which is where current developments live. Together, they give me a before-and-after picture of the topic.
A couple of practical notes on Connected Papers. The free plan gives you enough for a thesis; you shouldn’t need more than a handful of anchor papers. If you hit the limit, the academic plan is around $6 per month on the annual plan with a student discount. The node size reflects citation count, and the color intensity reflects recency, so large dark nodes are roughly “influential and recent,” which is often where I start reading.
If you don’t want to use Connected Papers at all, Research Rabbit does similar work without the monthly limit, and Semantic Scholar‘s “related papers” feature uses the same underlying database. I’ve used Connected Papers the most because I like the visual layout, but any of these would work.
Writing under the outline
By this point, I have: an outline with sub-section headings, and a set of papers clustered around an anchor. Writing is the process of reading and summarizing what these papers say under the relevant heading.
What I actually do is open the outline document, pick a sub-section, and start reading papers that are relevant to it. As I read, I paraphrase the points I want to use directly under the heading, with the reference next to each paraphrase. I don’t write a polished paragraph on the first pass; I write a list of claims with sources. The paragraph comes later, once I have enough material to see what the sub-section is actually arguing.
Sometimes, while writing, I realize I want to make a claim that I think is true, but I’m not sure which paper said it. When that happens, I use Consensus, which lets you search academic literature with a full sentence or claim rather than a keyword. I type my sentence, and it returns papers that express similar claims, along with a rough confidence signal. I then go to the relevant paper, confirm it actually says what I thought it said, and cite it properly. Consensus doesn’t replace reading the paper; it’s a way of finding the paper I already half-remember.
References as you go, not at the end
One decision I’m glad I made early: I format references in APA from the first paragraph. APA (American Psychological Association) style is widely used in social sciences, health sciences, and quantitative fields, including most of data science. It has two parts: in-text citations in the format (Author, Year) and a reference list at the end of the document, alphabetical by author surname. Which style you use depends on what your program or target journal requires, so check the submission guidelines before you start. Every time I add a new source, I add both the in-text citation and the full reference entry in APA format at the bottom of the document.
This sounds obvious. The reason I’m mentioning it is that the alternative, writing first, fixing references later, creates a mess. You end up with inconsistent formats, missing years, half-remembered author names, and a document-wide cleanup task at the worst possible moment. Fixing references as you go is boring but cheap; fixing them all at the end is painful and error-prone.
I also keep a folder on Drive with PDFs of every paper I cite. This is optional, and some people prefer reference managers like Zotero or Mendeley. A flat folder with author-year filenames works for me and means I can pull up any paper I’ve cited in seconds. Whatever system you use, having the paper itself reachable, not just the citation, saves time when you later need to check a specific figure or number.
Grammar and similarity check
When I finish a section, I run it through Grammarly. The free version catches most grammar and clarity issues, which is what I’m mainly after. I’m writing in English as a second language, and I’ve found that letting a tool run a grammar check saves me from small errors that would otherwise make the writing feel less polished than the ideas behind it. Worth checking: many universities have institutional agreements with Grammarly that give students free access to the premium plan with an .edu email. Mine does, and it’s a few minutes of setup.
Grammarly also has a similarity check (what they call plagiarism check). I run it not because I’m worried I copied something, but as a safety net. When you’ve read twenty papers on a narrow topic, it’s surprisingly easy to write a sentence that ends up phrased very close to one you read, unintentionally. The similarity check flags anything that looks too close to published text, and I can rephrase it before the check becomes someone else’s problem.
What the workflow is and isn’t
To be clear about what this is: a working process I put together to get from blank page to written section, with references, in a reasonable time. It’s not a systematic review methodology. I’m not claiming completeness; there are always papers I miss, and someone doing a more rigorous job would use structured databases, inclusion/exclusion criteria, and PRISMA diagrams. This is just what I did, in case any of it is useful.
Leave a Reply