Structuring a PhD Literature Review with Elicit and Semantic Scholar

PhD literature reviews are different

A PhD literature review is not a quick search. It needs to be comprehensive, defensible, and demonstrate command of the field. You need to be able to explain not just what you found, but how you searched, what you excluded and why, and that you didn’t miss important work.

AI tools help with the volume problem — finding and screening hundreds of papers — but they don’t replace the scholarly judgment that makes a PhD review credible. This tutorial shows how to use Elicit and Semantic Scholar efficiently while maintaining the rigor your committee will expect.

Phase 1: Define your search before you start

The most common mistake is starting to search before you have a clear question. Vague searches produce vague literature reviews.

Write out your review question in a single sentence:

“What does the evidence show about the effect of [intervention] on [outcome] in [population]?”

Or for a conceptual review:

“How has [concept] been defined, operationalized, and theorized across [field] since [date]?”

Then identify:

Key concepts — what are the 2–4 central terms you’ll search?
Synonyms and related terms — what else would relevant papers call these things?
Inclusion criteria — what study types, populations, date ranges, and languages qualify?
Exclusion criteria — what will you leave out, and why is that defensible?

Document these decisions before you search. Your methods section will need them.

Phase 2: Initial discovery with Semantic Scholar

Semantic Scholar is the right starting point for breadth. Its key advantage is semantic search — it understands meaning, not just keywords — and its citation graph, which lets you find papers connected to ones you’ve already found.

Step 1: Run your primary searches

Search each of your key concepts and combinations. For each promising paper you find:

Click Cite → note the citation count and the citing papers
Click References → skim for foundational papers in the area
Click Cited By → find recent papers that build on this work

Step 2: Use the “Highly Influential Citations” filter

Semantic Scholar flags citations where a paper is considered highly influential based on how it’s cited. Filter by this to quickly identify the most important papers in a subfield.

Step 3: Save to Zotero

Install the Zotero browser connector. Every time you find a paper to include or screen, save it to Zotero with one click. Create collection folders: To Screen, Included, Excluded, Key Papers.

Target: 100–300 papers to screen for a typical PhD chapter. More than 500 suggests your question is too broad.

Phase 3: Screening and extraction with Elicit

Elicit takes your paper set and does the tedious part: screening abstracts against your criteria and extracting structured data.

Step 1: Abstract screening

Upload your list of papers to Elicit or search directly within it. Ask Elicit to screen based on your inclusion criteria:

“For each paper, tell me: (1) Does the study population include [your population]? (2) Does it measure [your outcome]? (3) Is it a [your study design]?”

Elicit returns a table with yes/no/unclear for each criterion per paper. Review the table and make final inclusion decisions — don’t let Elicit make them for you. It will miss nuances and occasionally misread an abstract.

Step 2: Data extraction

For included papers, set up an extraction table in Elicit:

“For each paper, extract: (1) Study design, (2) Sample size and population, (3) Key intervention or exposure, (4) Primary outcome and how it was measured, (5) Main finding and direction of effect, (6) Key limitations noted by the authors.”

Elicit populates the table. Review every cell — errors cluster around:

Papers with complex designs (extraction oversimplifies)
Papers reporting multiple outcomes (Elicit may pick the wrong one)
Non-standard terminology (a measure called something unusual in one paper)

Step 3: Export the extraction table

Export as CSV. This becomes the working document for your synthesis and your evidence table for the appendix.

Phase 4: Synthesis with NotebookLM

Once you have your included papers extracted, use NotebookLM for synthesis.

Upload the PDFs of your included papers. Then query across them:

“Across these papers, what are the most consistent findings? What do studies disagree on, and what might explain the disagreement?”

“What methodological limitations are most commonly acknowledged?”

“Which papers represent each major theoretical position on [contested concept]?”

Use NotebookLM’s notes feature to build themed synthesis sections: Background, Core Findings, Methodological Approaches, Gaps and Contradictions. These become the scaffolding for your writing.

Phase 5: Gap analysis and snowballing

Before finalizing your literature set, do two more searches:

Backward snowballing: Take your 10–15 most important included papers. Look at their reference lists. Are there papers you missed that appear frequently? Add them to your screening set.

Forward snowballing: For the same key papers, check who cites them (use Semantic Scholar’s “Cited By”). Recent papers citing foundational work you’ve identified may be important and weren’t in your original search results.

Gap documentation: Write one paragraph on what the literature doesn’t cover — the gap your research addresses. This is often the most important sentence in your introduction. NotebookLM’s synthesis should surface this naturally if you ask it directly.

Documenting your process (essential)

Your PhD committee — and any journal reviewers — will want to see your search process. Keep a search log with:

Date	Database	Search string	Results returned	After deduplication	After screening
2026-07-01	Semantic Scholar	“federated learning” AND “clinical”	347	302	41

Also document: inclusion/exclusion criteria, how disagreements between screeners were resolved (if applicable), and any post-hoc additions via snowballing.

This documentation is your PRISMA flow diagram if you’re writing a systematic review, or your methods section if you’re writing a narrative review.

What AI tools don’t replace

Your judgment on quality. Elicit can tell you a paper’s stated methods and findings. It can’t tell you whether the methods are appropriate, the analysis is sound, or the conclusions are warranted. Critical appraisal is still your job.
Grey literature. Elicit and Semantic Scholar cover published peer-reviewed research well. For policy reviews or topics where unpublished reports matter, you’ll need to search manually.
Non-English literature. Coverage of non-English papers is uneven in most AI search tools. If your topic has significant literature in other languages, plan a supplementary search.
Field-specific databases. For clinical research, PubMed; for psychology, PsycINFO; for education, ERIC. AI tools don’t replace field-specific database searches for comprehensive reviews.