How to Use Elicit for a Systematic Literature Review
If you have ever spent three days skimming 200 abstracts for a systematic review, you will understand why Elicit exists. It is not a search engine for papers. It is closer to a research assistant that reads the papers for you and pulls out exactly the columns you care about: study design, sample size, effect size, limitations, whatever you need. The screening step that used to take days can shrink to a couple of hours.
The catch is that Elicit rewards a bit of upfront thinking. A vague question produces vague columns and noisy results. A specific, well-framed question produces a table you can actually use. So before you open Elicit, the single most important thing you can do is spend five minutes writing a crisp research question.
Framing Your Research Question
Elicit works best with questions that have a defined population, intervention or exposure, and outcome. This is the PICO framing that systematic reviewers already know, but it applies even outside clinical research.
Bad: "What are the effects of social media on teenagers?" Better: "Does social media use (more than 2 hours/day) increase depressive symptoms in adolescents aged 12 to 17?"
The tighter version gives Elicit something to extract consistently across papers. When you ask a broad question, the columns it suggests will be broad too, and you end up with a table full of "varies" and "unclear."
If you genuinely need to start broad because you do not yet know the field, that is fine. Use the first Elicit run as a mapping exercise, read 20 to 30 abstracts, then reformulate the question before doing the real extraction pass.
Running Your First Search
Once you have a question, type it into the Elicit search bar exactly as a question, not as a set of keywords. Elicit's retrieval is semantic, so "Does mindfulness reduce cortisol levels in healthy adults?" works better than "mindfulness cortisol adults."
After a few seconds, you will see a list of papers. Each row is a paper, with the abstract collapsed by default. A few things to notice immediately:
- Relevance score: Elicit gives each paper an internal relevance ranking. Papers at the top are most semantically similar to your question, but relevance is not the same as quality. Scroll down to see if anything important got buried.
- Year and citations: Both are visible in the default view. For a systematic review, you typically want to filter by date range before going further.
- Open access: A small lock icon indicates whether the full text is available. Elicit extracts richer data when it can read the full text, not just the abstract.
Adding and Configuring Extraction Columns
This is the most powerful part of Elicit. Click "Add column" and you will see a set of pre-built column suggestions (sample size, study design, intervention, outcome measures, country) plus a free-text option where you type your own.
For a systematic review, I usually build a table that looks something like this:
| Column | What I'm asking |
|---|---|
| Study design | RCT, cohort, cross-sectional, etc. |
| Sample size (N) | Total participants |
| Population | Age, condition, setting |
| Intervention / exposure | What was done or measured |
| Outcome | Primary outcome and how it was measured |
| Effect size / key finding | Main quantitative result |
| Limitations | As stated by the authors |
Each column runs a targeted extraction pass using AI. The extraction is not perfect, especially for studies with unusual reporting formats, but the accuracy on well-structured papers is genuinely impressive. I would estimate 80 to 90% accuracy on fields like sample size and study design when the full text is available.
One practical tip: add the "Limitations" column from the start. Reviewers often skip it, but it is the fastest way to spot papers that should be excluded for methodological reasons without reading the whole paper.
Screening Papers Efficiently
Once your columns are populated, you can sort and filter the table to do a first-pass screen.
Here is the workflow I use:
- Sort by study design first. If your protocol requires only RCTs, every observational study gets flagged immediately without reading it.
- Sort by year and exclude anything outside your date window.
- Scan the sample-size column for outliers. A study with N=12 in a field where everything else is N=200 warrants extra scrutiny.
- Use Elicit's "chat with paper" feature on any paper where the extracted data looks odd. Click the paper row, then ask a specific question like "What is the exact sample size at follow-up?" This catches cases where Elicit grabbed the screening number instead of the final enrolled number.
- Mark papers as "include" or "exclude" using the checkbox system. Elicit tracks your decisions, and you can export the list at the end for your PRISMA flow diagram.
Elicit does not do the PRISMA flow automatically, but having a labeled include/exclude list makes building it in a spreadsheet straightforward.
Working With the Full-Text Reader
When Elicit can access the full PDF (either through open access or because you uploaded it manually), it reads the entire paper rather than just the abstract. This matters for extraction accuracy on methodology sections, which are often buried in page 4 of a journal article.
You can upload papers manually by clicking the upload button in the paper row. This is useful for papers you have downloaded from a paid database. Elicit will then use the uploaded PDF for column extraction, even if the paper is paywalled.
The "chat with paper" sidebar is also more useful with full text. You can ask questions like "Does this paper control for confounders X and Y?" and get a grounded answer with the relevant paragraph highlighted.
Exporting to a Review Matrix
When you are done with screening and extraction, click Export at the top right. Elicit offers CSV export, which gives you one row per paper and one column per extraction field.
From there, most reviewers drop the CSV into Excel or Google Sheets and:
- Color-code rows by include/exclude decision
- Add a "Notes" column for reviewer disagreements
- Add a "Risk of bias" column for manual entry
The exported CSV is not a finished evidence table. It needs cleanup: some cells will have long quoted excerpts instead of clean numeric values, and a few cells will be blank or wrong. Budget 30 to 60 minutes for cleanup depending on how many papers you included.
For collaborative reviews with two screeners, Elicit does not yet have a native disagreement-resolution workflow. Export each screener's decisions separately and merge in a spreadsheet, flagging rows where decisions differ.
What Elicit Does Well and Where It Falls Short
Elicit is strongest on English-language, open-access literature in well-structured fields like biomedicine, psychology, and economics. It gets weaker on:
- Non-English papers (it can read them, but extraction quality drops)
- Qualitative studies with no standard metrics to extract
- Very recent papers (indexing has a lag of a few weeks to months)
- Grey literature (conference papers, theses, government reports not in its database)
For most academic literature reviews, those gaps are manageable. For systematic reviews in clinical medicine where completeness is a regulatory requirement, Elicit works best as a complement to Cochrane, PubMed, and Embase, not as a standalone search.
The honest summary: Elicit makes the extraction and screening phases of a literature review substantially faster. The research question framing and final quality judgment still need a human. Use it to handle the volume, and spend your saved time on the thinking.