How to Find Research Papers: Reviews, Citation Trails, and AI
Published:
Do not search for papers one by one, in isolation
In Part 2, I explained why a research question should come before a method. The next step is to place that question in the existing literature.
If you cannot find the first relevant paper, it is easy to spend time only looking through search results. You may not know whether there is little prior work or whether you have simply not found it yet.
The key is not to search for papers one by one, in isolation. Start with a review paper to map the field. Follow references and cited-by papers. Then refine your search terms. AI can help you generate possible paths, but it is a research assistant, not the source of the answer.
This article is for people who have just started research and do not know how to find their first papers. It does not assume subject-specific knowledge or a paid AI plan.
What you will learn
- How to connect review papers, references, and cited-by papers into a search path
- How to refine search terms from papers and use English searches to expand the search
- How to use AI to shape a research request and choose papers to inspect, while checking the original sources yourself
First, look for a map of the field, not one perfect answer
At the start, you may want to find a paper that answers your question exactly. But the terms used for the same topic can differ across fields. Researchers may frame the problem, comparison, and evaluation in different ways.
Your first goal is not one perfect paper. It is a map of the field:
- What words do researchers use for this problem?
- What methods and debates are common?
- What is known, and what is still debated?
- Which papers are cited repeatedly?
A review paper is a useful starting point. It gathers and organizes research on a topic. It is not a final authority, but it can reveal search terms, widely cited authors, and useful points of comparison.
Try combining a central term from your question with review, survey, or systematic review. If Japanese searches do not find enough, search in English as well. English search is useful even if you do not write your research in English: it helps you find the terms and papers used internationally.
Use references and cited-by papers to make a reading path
Once you find one relevant paper, do not stop there. Follow it backward and forward to create a path through the literature.
| Direction | What to inspect | What it can clarify |
|---|---|---|
| References | Earlier papers cited by the paper | Assumptions, problem framing, and the origins of a method |
| Cited-by papers | Later papers that cite the paper | Extensions, applications, comparisons, criticism, and newer evaluations |
References can help you see why a problem was framed in a particular way and what a method builds on. Cited-by papers can help you see how later work used, compared, or developed it.
To understand how authors assess their own method, read not only the results but also the discussion, conclusion, and future work. To understand later evaluation, look through cited-by papers. To understand the starting point of a problem, definition, method, or metric, look through references.
A high citation count alone does not make a paper the most important one for your question. A paper can be cited because it is old, convenient as a method, or useful outside its original field. Before opening a paper, one short purpose is enough: “I want to understand where this method began,” or “I want to check this evaluation method.”
Use generative AI to narrow the next references
When a paper has many references, I sometimes give generative AI the paper PDF or its reference list and ask it to help prioritize the next candidates. Asking only for “important papers” is not enough. Include your question and what you want to clarify now.
For example:
This paper uses a particular problem framing and method.
From its references, rank the papers I should read to understand where that framing and method began.
For each paper, separate:
- its relationship to this paper; and
- why I should read it.If you want to examine the authors’ assessment or later comparisons, focus instead on the discussion, conclusion, future work, and cited-by papers.
An AI recommendation is only a proposal for reading order. Check the title, abstract, and relationship to the citing paper yourself. Removing a candidate that does not fit your question is also part of building a useful search path.
Refine your search terms from papers
When your first search terms fail, the question itself may not be wrong. Your words may simply differ from those used in the field.
After finding one paper, collect terms from the following places. Here, an evaluation metric is a criterion defined in advance for comparing conditions numerically. In research that does not center on numerical comparisons, use the names of methods and materials used in that field as search terms instead.
- its title, abstract, and keywords;
- nouns repeated in its problem statement;
- names for methods and, in studies that compare conditions quantitatively, evaluation metrics;
- category headings in review papers; and
- titles in its references and cited-by papers.
You can also give a paper PDF or abstract to generative AI and ask for several search-term candidates. Separate the target, problem, condition, method, and evaluation metric. Do not use every suggestion unchanged. Test the terms in a real search and keep the ones that return relevant papers.
Using this paper, suggest English search terms for finding related research.
- Separate the target, problem, condition, method, and evaluation metric.
- For each, list the wording used in the paper and several alternatives.
- Give five search queries that combine those terms.
- Show which wording in the paper supports each candidate.It is easier to manage new terms when you group them by what you want to find.
| What you want to find | How to form the search terms |
|---|---|
| A map of the field | Target, phenomenon, or problem + review / survey |
| An assumption or cause | Target + condition + factor |
| A comparison method in a quantitative study | Target + problem + benchmark / evaluation / comparison |
| Recent work | A central term + year, or cited-by papers for a representative paper |
If a search finds nothing, do not only add more words. Return to broader terms. Separate your question into target, problem, condition, and method. Search first for the target and problem, then add conditions or methods.
Use AI research tools to expand candidates and search paths
Some AI tools research across multiple sources and return a report with citations. ChatGPT and Gemini call this Deep Research; Claude calls it Research. Their names and interfaces differ, but the basic practice is the same: state the question, purpose, and scope, then verify the sources yourself. They can be useful for expanding search terms and paper candidates when the question requires several steps of investigation.
To begin, choose the research feature in the service you use, then state your question, purpose, and scope. In ChatGPT, Deep Research is available from the tools menu. Gemini offers it from the prompt area, and Claude offers Research from the chat interface. Interfaces and access conditions can change, so check the official guidance for ChatGPT, Gemini, and Claude.
I use ChatGPT, Gemini, and Claude, but this article does not compare them or explain their detailed interfaces. The same approach can be used across services.
I also ask generative AI to help draft the request I send to Deep Research. First, I write down why I am investigating something, what decision the result will support, what I want to investigate, and what I want to exclude. I then ask generative AI to organize the request into four parts: research purpose, questions to investigate, expected output format, and exclusions. I read the final request myself and revise it until the research stage, conditions to compare, and out-of-scope areas are right.
Do not choose papers to cite from an AI research report alone. Check whether a paper exists, what it actually claims, and whether a citation relationship is correct on the paper page, DOI page, publisher or conference page, and in the paper itself. OpenAI also notes that Deep Research can hallucinate facts or make incorrect inferences. Official explanation
Ask for research that supports a decision, rather than asking for an answer or “the best method.” A request can take this form:
# Request: (title of the research task)
## Research purpose
(Describe what you are considering and what decision the result will support.)
## Questions to investigate
### 1. (A point to investigate)
- (State the methods, conditions, field, and time period to include.)
- (State the assumptions or properties to compare.)
### 2. (Another point, if needed)
- (State what to investigate.)
## Expected output format
- For each paper, give the title, authors, year, venue, and URL or DOI.
- Separate the method, assumptions, strengths and weaknesses, and relationship to this research purpose.
- Separate facts, inferences, and points that have not been verified.
## Exclusions
- (State excluded fields, methods already examined, time periods, or comparison conditions.)The goal is not a conclusion. It is a starting point for finding prior work, comparison conditions, and questions that need checking. A cited source is not automatically correct; trace promising candidates back to the original source.
I will share what I learn from Perplexity and NotebookLM later
Perplexity and NotebookLM may also be useful for literature work, but I have not used them enough yet. After testing them, I will share what I learn about finding candidates, checking sources and citation relationships, and organizing materials in a separate article.
Summary: Use several paths to find papers
Before concluding that there is little prior work, try several paths: search terms, review papers, references, cited-by papers, and AI-assisted search. Once you have those paths, you can see what you have already checked and what to try next when you get stuck.
Literature Research Guide 1: Check the publication type before reading
Once you find a paper, check its publication type and review status. Paper Types and Reliability: What to Check Before Reading introduces the differences between peer-reviewed papers, preprints, journal articles, and conference papers.
Next, I will cover how to start reading a paper with an AI summary, verify it in the original paper, and connect it to your own research question.