Running a Systematic Literature Review with AI: A Step-by-Step Workflow

Before you start

AI tools can dramatically speed up discovery, screening, and extraction — but they don’t replace your judgment on relevance, quality, or interpretation. Treat every AI-generated summary as a first draft that you verify against the source, not a finished result.

No single tool does the whole job well. The most reliable approach uses a layered workflow: a free discovery tool to find papers, a structured-extraction tool to screen and pull data, and a synthesis tool to pull it all together — with a reference manager underneath the whole thing.

Step 1: Define your research question and inclusion criteria

Before opening any tool, write down:

• Your specific research question (as narrow as you can make it)
• Inclusion/exclusion criteria (study type, population, date range, language)
• The outcome or comparison you’re extracting

Why this matters: every AI extraction tool works from fields you define. A vague question produces vague, inconsistent extraction — the tool can’t fix a fuzzy research question for you.

Step 2: Broad discovery with a free search tool

Start in Semantic Scholar (or ResearchRabbit if you want a visual citation map instead of a list). Search your core question and skim the first 30–50 results for relevance, not depth.

Goal at this stage: build a rough candidate pool, not a final list. Aim for breadth over precision — you’ll narrow later.

Step 3: Structured screening and extraction with Elicit

Import your candidate pool (or re-run your search) in Elicit. Here’s where the real time-saving happens:

1. Define your extraction fields exactly as you wrote them in Step 1 (e.g., sample size, methodology, effect size, population).
2. Let Elicit populate those fields across your paper set automatically.
3. Elicit will link every extracted data point back to the sentence in the source paper it came from — click through and verify each one, especially for any number you plan to report.

Step 4: Synthesize across your verified paper set with NotebookLM

Once your extraction table is verified, upload your final paper set (PDFs) into NotebookLM. Ask it targeted synthesis questions:

• “What do these papers agree on regarding [outcome]?”
• “Where do these papers disagree, and why might that be?”
• “Summarize the methodological differences across this set.”

Because NotebookLM grounds every answer in the specific documents you uploaded, you can click through to verify each synthesized claim — this is a meaningfully lower hallucination risk than asking a general-purpose chatbot the same question with no source constraint.

Step 5: Manage citations and draft with Zotero underneath everything

Keep your verified paper set in Zotero throughout — not just at the end. This gives you:

• Consistent citation formatting when you draft
• A clean audit trail of what was included/excluded and why
• A stable library independent of any single AI tool’s continued access or pricing changes

Common pitfalls across the whole workflow

• Citation fabrication: general-purpose AI chatbots (not the research-focused tools above) can invent plausible-sounding but nonexistent papers. Stick to tools that search live academic indexes and link to real sources.
• Overtrusting relevance rankings: semantic search surfaces topically similar papers, not necessarily the right papers for your specific criteria. Always eyeball results yourself.
• Skipping verification on numbers you’ll actually cite: any statistic that ends up in your manuscript should be checked against the original PDF, not just the extraction table.
• Treating synthesis as complete: an AI synthesis is only as good as the paper set behind it — it can’t warn you about papers it never saw.

Quick reference: tool sequence

Related content

• Elicit — full tool review
• NotebookLM — full tool review
• AI Tools for Literature Review: Elicit vs. Consensus vs. NotebookLM vs. Semantic Scholar