How to Build a Paper Comparison Matrix with Elicit – Toolveyor
How to Build a Paper Comparison Matrix with Elicit
Build a paper comparison matrix with Elicit: define extraction columns, verify source evidence, preserve missing values and check export and plan limits.
Written by Andrew White
•7 min read
A paper comparison matrix is useful when the same question must be answered across a small, chosen set of studies. The hard part is keeping a sample size, outcome or limitation attached to the right paper and the right measurement conditions.
Elicit can help produce that first structured table. This guide explains how to scope the input, define extraction fields and check the result before using it in a literature review or research briefing.
Basis of this guide: Elicit's official documentation and public pricing, checked October 20, 2026. The workflow and prompts below are proposed editorial templates. We did not run an authenticated extraction experiment or measure its accuracy. The cover is a conceptual illustration.
What you need before starting
Choose three to five papers addressing one question, with full texts you are permitted to upload. Keep their titles, identifiers and versions in a separate source list. This small starting batch is our recommendation for making manual checks manageable, not an Elicit limit.
Write a one-sentence comparison question, such as: “Which study designs and outcome measures do these papers use to evaluate remote team collaboration?” Avoid asking which intervention is best until you have checked whether the studies are comparable.
Prepare three things:
A fixed list of papers, with supplements where necessary.
A field dictionary specifying exactly what belongs in each column.
A place to record human corrections, unresolved questions and the date of verification.
For a broader choice of discovery and citation tools, see our AI research tools guide. Here the starting point is a selected paper set, rather than an open-ended literature search.
1. Start in the current Research Agent workflow
Elicit's migration guidance says that Find Papers, Extract Data and Chat with Papers became part of Research Agent on September 30, 2026. Older sessions remain view-only; they are not the place to build a new editable matrix.
Use the Extract Data skill from the homepage, or give Research Agent your papers and request a table. For an existing collection, name it explicitly and specify that the comparison should stay within that collection. Elicit documents collection scoping and paper identification by title, DOI, PMID or filename. An instruction to restrict sources still needs checking against the returned paper list.
Checkpoint: Confirm that every intended paper appears once. Match identifiers and versions; a conference paper and its expanded journal version may represent overlapping evidence. If the agent adds outside sources, remove them from the comparison or explain why they belong before proceeding.
Share:
2. Define the columns before requesting extraction
A useful matrix separates the reported finding from your interpretation. Start with a few fields, check them, then add more. These are editorial field definitions, not a claim about default Elicit column names.
Field
Extraction instruction
Human check
Study identity
Title, year and persistent identifier
Match the supplied version
Design and setting
Reported design and where the study occurred
Distinguish an experiment from an observational study
Sample
Number analyzed, unit and relevant subgroup
Do not substitute the recruited total
Outcome and timing
Exact outcome definition and measurement time
Keep different scales and follow-up periods distinct
Reported result
Value, unit and uncertainty if reported
Preserve the denominator and comparison group
Evidence location
Supporting quotation and section, page or table when available
Open the original and locate the passage
Review status
Leave for the human reviewer
Mark checked, unresolved or excluded
For non-empirical papers, adapt the fields instead of forcing a sample size into the table. A conceptual framework may need assumptions, proposed mechanism and supporting examples. Use “not applicable” only when the field truly does not apply.
Elicit documents an Add column control on agent-created tables; you can also ask the agent to add a field. A column that blends population, outcome and interpretation is difficult to audit. Split it when the answers become ambiguous.
3. Request a bounded first matrix
Use this as a starting prompt, replacing the bracketed text. It describes the requested output; it is not a tested guarantee that every instruction will be followed.
Use only these supplied papers: [titles and identifiers]. Compare them for this question: [question]. Create one row per paper. Include study design and setting, analyzed sample with unit, outcome definition and timing, and reported result with its unit and uncertainty where provided. Attach supporting source citations. Do not introduce additional papers or infer missing values. Distinguish not reported, full text unavailable and not applicable. Flag conflicting values for manual review. Do not rank the studies or combine their results.
Ask for evidence locations alongside the data rather than accepting a detached narrative summary. Elicit's Research Agent documentation describes clickable citations and iterative changes to outputs. A citation is a route to verification, not proof that the extracted interpretation is correct.
Checkpoint: Inspect one row in full before extending the matrix. If the sample definition is wrong, repair that column's instruction and revisit every affected row. Adding ten more columns first makes the same ambiguity harder to find.
4. Check values against the original paper
Review all values that will support your final comparison. Keep a separate correction log so an overwritten cell does not erase the reason for a change.
Use these review states consistently:
State
Meaning
Next action
Checked
A reviewer located matching evidence
Record the location and review date
Not reported
The consulted material does not state the requested value
Check relevant supplements before treating it as absent
Full text unavailable
The needed source could not be inspected
Obtain permitted access or leave the field unresolved
Conflicting
Two passages or versions disagree
Record both and resolve the difference explicitly
Not comparable
Values use incompatible definitions or conditions
Keep them separate; do not compute a pooled result
For numerical fields, compare units, denominators, time points and subgroup labels. A percentage without its population is not a complete result. A blank cell is not zero, and a missing significance statement is not evidence of no effect.
Elicit's table and figure guidance acknowledges that complex tables can be missed. Inspect the original table when an important result is absent or looks implausible. Premium PDF parsing is available above Basic and consumes more usage across PDFs while enabled. Figure extraction is a separate Scale/Enterprise capability within Systematic Reviews, with additional settings; do not assume that better parsing gives a Plus account figure extraction.
5. Export a checked handoff, with unresolved fields visible
The expected result is a comparison table with traceable values and an accompanying review log. It is not automatically a systematic review, risk-of-bias assessment or defensible pooled analysis.
According to Elicit's export documentation, Research Agent table export requires Plus or higher. CSV and Excel are supported. Export availability for a Research Report is different from table export, so a free report download does not establish free matrix export.
Use the table's export control, then reopen the downloaded file. Check the row count, identifiers, units, multiline text and missing-value labels. Verify what happened to citation links; retain a separate source list and evidence log if the exported format does not preserve the evidence path you need. Keep the raw export and a dated reviewed copy so later corrections remain understandable.
Which plan is enough for this workflow?
For evaluation, Basic is a free plan with limited agent usage. It is not a promise of unlimited extraction or downloadable tables. For an academic user who needs a modest comparison table and export, check Plus before considering the dedicated review workflow.
The public pricing page has separate Academic and Industry selectors. These USD prices were checked with the billing controls on October 20:
Plan and audience
Monthly billing
Annual billing
Basic
Free
Free
Plus, Academic
$19 per user/month
$132 per user/year; $11 monthly equivalent
Pro, Academic
$69 per user/month
$468 per user/year; $39 monthly equivalent
Pro, Industry
$75 per user/month
$588 per user/year; $49 monthly equivalent
Annual equivalents require annual payment; they are not month-to-month offers. Confirm the applicable audience and checkout total. Plus lists five columns at a time; Pro lists twenty and includes the dedicated Systematic Review workflow. A small matrix alone does not establish a need for Pro.
Elicit's usage guidance now describes a shared monthly pool rather than fixed extraction counts. Larger files, more columns and repeat work consume usage. Check Account Settings for the remaining allowance and reset date. Pro and Scale can enable separately billed Extra usage; leave that disabled unless additional spending is intentional. Do not budget from older claims of a fixed number of PDFs per month.
If you still need to decide which papers belong in the comparison, our Elicit screening review examines eligibility criteria, exclusion checks, paid workflow limits and the difference between a screening record and a cited report.
When to choose a different approach
For two short papers, a manually maintained spreadsheet may be quicker to audit than configuring an agent. If you need formal screening decisions, documented inclusion criteria and a structured review process, examine Elicit's Systematic Review workflow instead; the workflow guide places it on Pro, Scale and Enterprise.
If your actual goal is explaining one paper to an audience, our research paper to presentation workflow addresses that different task. Do not turn an unfinished comparison matrix into slides that conceal unresolved evidence.
Next step: Open the Elicit tool page, choose a small permitted paper set and define three decision-relevant fields. Verify one complete row before expanding the table or paying for a larger workflow.
