A synthesis matrix generator helps you compare papers by topic, method, finding, limit, disagreement, and gap. Each claim in the table should point back to a source. The tool should not write your literature review for you. It should organize checked facts so you can build your own argument.
Use a template when you only need a set of columns. Use a spreadsheet when you want full control. Use a research workspace when you need to compare papers without losing the source behind each claim.
Atlas fits after you choose the papers. Add them to a project and ask one question across the set, such as how the studies measure the same outcome. Open the citations in the answer and move only supported findings into your matrix.
Compare papers for a synthesis matrix
Add paper PDFs to Atlas, ask what differs, and inspect cited passages.
Free synthesis matrix template
You can start with a blank template. Both files have the same six columns: source, research question, method and sample, key finding, limitation, and how the source relates to your question. Write your research question at the top before you fill any rows. Fill each cell from the paper itself, not from the abstract alone.
Word template
You can download the Word template (.docx) and fill it in by hand. The page is turned sideways so the six columns fit. It has a blank table with eight rows, then the worked example from this page.
Excel and Google Sheets
You can also download the Excel template (.xlsx) if you prefer a spreadsheet. The first sheet is blank. The second sheet holds the worked example, so you can see how full a cell should be.
To use the template in Google Sheets, open Sheets and choose File, then Import, then upload the Excel file. Google's help page lists .xlsx as a file type Sheets can import (Google Docs Editors Help).
Synthesis matrix criteria

A synthesis matrix is more than a row of notes for each paper. It helps you read across papers and see what they say about the same idea. Read one topic at a time and compare the sources in that column. The tool should make this easy rather than give you a smooth paragraph with no source trail.
Look for five jobs:
- Name each source so you can see which paper supports each topic.
- Keep the method, finding, author-stated limit, and your own judgment in separate fields.
- Show agreement, disagreement, and gaps across sources.
- Keep a citation or source passage for each key entry.
- Leave room for your judgment. The matrix should help you plan the review, not write its argument.
A matrix tool can provide fixed fields, sorting, and export. A spreadsheet is better when you want to create every column yourself. Atlas is strongest when you already have a set of papers to compare: you can ask a question across them, see which source supports each claim, and open the cited passages before adding a row.
The method comes from library and writing guides at USU, North Shore, Johns Hopkins, and FIU. Examples from SubThesis, UNC, Jenni AI, and a public agency show how people turn that method into tools and templates.
Source-checked matrix fields

Germanna Community College’s example lays out one source per row and separates topic relevance, main points, and evidence into columns. Its filled row is about online-dating safety, unrelated to this article’s subject; it demonstrates matrix structure only. Source: Germanna Academic Center for Excellence.
Start with a few fields and add more only when the question calls for them. Too many columns create busywork. Too few mix source facts with your own judgment.
| Field | What it captures | Verification rule |
|---|---|---|
| Source | Author, year, title, or stable citation label | The source must be one you reviewed. |
| Theme | The concept, variable, population, or problem the row belongs to | Use the same theme label across papers when comparing them. |
| Method | Study type, dataset, sample, intervention, or analytic approach | Do not compare findings without noting different methods. |
| Key finding | The source's relevant result or argument | Keep this as close to the source wording as practical. |
| Evidence passage | Page, paragraph, quote fragment, or citation trail | Open the source before relying on the entry. |
| Limitation | Weakness, boundary, caveat, or missing context | Separate author-stated limits from your critique. |
| Agreement or disagreement | How this source relates to the others | Name the sources it agrees or conflicts with. |
| Gap | What remains unanswered after comparing sources | Turn gaps into future search or writing prompts. |
| Researcher interpretation | Your read of what the evidence means | Label this column as your own judgment. |
Table 1: A strong matrix keeps the source finding, the author's limits, the links to other papers, and your own judgment in separate columns.
Build the matrix
- Write the review question before opening the tool. A matrix about AI tutors needs different fields from one about how people test search systems.
- Select the source set. Use DOI, title, author, or topic search where available, but review each paper before it becomes part of the matrix.
- Draft a short list of topics from your question, repeated ideas in the papers, and headings in your notes.
- Enter source facts before your own judgment. Add the finding, method, and author-stated limits first. Write your notes after you can compare rows.
- Read vertically by theme. For each theme, ask which sources agree, which conflict, which use different methods, and which leave a gap.
- Revise the categories. Merge duplicate themes, split vague themes, and delete columns that do not help you compare sources.
- Turn the checked matrix into an outline. Let the strongest themes and tensions set the literature review section order.

This open-access review’s supplementary table compares characteristics of included studies across consistent columns. Its subject is cardiac rehabilitation; it demonstrates one way to read across sources, not a universal field list. Source: Rashidi et al., PLOS ONE S6 File, licensed CC BY 4.0.
For each possible claim, open the citation and read the passage. Then decide what belongs in the matrix. Do not treat an AI answer, abstract, or generated summary as the evidence itself.
Atlas synthesis workflow
Steps in Atlas
Use this Atlas process after you choose the papers:
- Add the papers to one Atlas project. Use PDFs when you can. You can also find a paper by its DOI, arXiv ID, title, or author.
- Ask: "Compare these papers in a table with columns for topic, claim, source passage, limit, disagreement, gap, and citation."
- If the answer blends sources together, ask Atlas to separate sources by row and name which paper supports each claim.
- Open the citation behind each key claim. Make sure it opens the right paper and that the passage supports the full claim.
- Move only checked entries into the matrix. If the citation is weak or wrong, ask a narrower question or read the paper yourself.
Our test run
We ran this on 7 November 2026 with five open-access papers on one question: how often do AI chatbots make up references? We added the five PDFs to one Atlas project and asked for one row per paper. The columns were models tested, sample, how references were checked, what counted as fabricated, main finding, and main limitation. Atlas returned a five-row table with a citation in every cell.

Then we checked every cell against the full text of each paper. Most cells held up, but three did not. The Chelli row said two reviewers checked the chatbot references in PubMed and DOI databases. The paper does not say that. Two reviewers did screen the systematic reviews. For the chatbot references, the paper only says the research team reviewed them, and it names no DOI database. The same row called all 471 references chatbot output, but 109 of them were the reviews' own references. The Walters row named a limit the authors do not state. They used one fixed prompt, but they never call that a limit.
That is why step 4 matters. The citation chip showed us where to look, and reading the passage showed that the claim went further than the paper. We wrote the worked matrix below by hand from the full papers and checked each cell against the paper text.
Worked matrix from 5 papers
The matrix
This matrix answers one question: how often do AI chatbots make up references, and how should a student check them? Each cell was checked against the open-access full text on 7 November 2026.
| Study | Method and sample | Key finding | Limit or scope |
|---|---|---|---|
| Walters & Wilder (2023), Scientific Reports | GPT-3.5 and GPT-4 wrote 84 short papers on 42 topics in April 2023, with 636 citations. A citation was real if title and author matched. | 55% of GPT-3.5 citations and 18% of GPT-4 citations were made up. Of the real ones, 43% and 24% had substantive errors. | No limits section; topics were like first-year writing assignments. |
| Bhattacharyya et al. (2023), Cureus | ChatGPT-3.5 wrote 30 short medical papers with 115 references. Two reviewers searched Medline, Google Scholar and DOAJ. | 47% of references were made up, 46% were real but wrong, and 7% were real and correct. | One model; custom prompts. |
| Chelli et al. (2024), JMIR | GPT-3.5, GPT-4 and Bard tried to find the studies in 11 rotator cuff reviews, giving 362 references. A reference was hallucinated if 2 of title, first author and year were wrong. | Hallucination rates were 39.6% for GPT-3.5, 28.6% for GPT-4, and 91.4% for Bard. | One clinical topic. |
| Athaluri et al. (2023), Cureus | ChatGPT wrote 50 medical research protocols with 178 references. Five researchers checked each DOI and title. | 28 of 178 references (15.7%) had no working DOI and no Google result. | Small sample; relied on Google. |
| Agrawal et al. (2024), arXiv | Six models gave 1,000 reference titles each on computer science topics. Each title was checked with an exact Bing search. | Rates ran from 46.8% for GPT-4 to 76.7% for Llama 2 Chat 13B. | Computer science only. |
Table 2: Five studies of made-up AI references, compared by method, finding and limit.
Reading down the columns
Now read down the columns instead of across the rows. The finding column says made-up references are common, but the rate swings from 15.7% to 91.4%. The method column explains much of that. Each study defined "made up" in its own way. Athaluri counted a reference only when both the DOI and a Google search failed. Chelli counted a reference when 2 of 3 details were wrong. Bhattacharyya adds a second point: a real reference is not the same as a correct one, since only 7% were both.
Reading down the columns gives you a claim to write. GPT-4 made up fewer references than GPT-3.5 in all three studies that tested both (Walters, Chelli and Agrawal). No model in these studies was safe to trust without checking. Every study tested models that were current in 2023, so newer models are the gap to search next.
What a generator cannot decide for you
A generator can organize rows faster, but it cannot judge the quality of a study for you. A weak paper should not get the same weight as a strong one. It cannot tell you that your search found every relevant paper or turn a conflict between sources into your final argument.
Be especially careful with four boundaries:
- Study quality: Judge quality apart from copying facts into the matrix unless your method uses a set score.
- Source coverage: a neat matrix can still be built from an incomplete source set.
- Citation strength: A citation gives you a path to check the claim, not proof that the claim is right.
- Writing role: The matrix helps you plan the review. You must write the final argument.
Next steps for matrix work
Choose a tool based on the problem you need to prevent.
| Workflow | Use it when | Watch out for |
|---|---|---|
| Atlas source-grounded workflow | You already have papers and need cited cross-source synthesis before filling the matrix. | You still need to open citations and judge whether each matrix entry is supported. |
| Template | You need a quick class assignment structure or first-pass field list. | Templates do not compare sources for you. |
| Spreadsheet | You want manual control, custom columns, filtering, and export. | Manual entry can turn into source-by-source notes if themes are vague. |
| Single-purpose generator | You want fast structured entry around study fields, findings, and limitations. | Check whether export, storage, and citation behavior match your needs. |
Table 3: Four ways to build a synthesis matrix, with when each fits and what to watch for.
For a short class task with five readings, a template or spreadsheet may be enough. For a thesis chapter or research memo, keep the citation and passage close to each claim. Atlas helps turn a folder of papers into a cited view of the main topics, conflicts, limits, and gaps, which you then check passage by passage.
Use the matrix fields and citation checks in this guide when you need a table you can inspect row by row. If you need to turn that evidence into a thematic, chronological, or method-based argument, continue with the broader guide to synthesizing research papers.
For related work, compare legal document organizers, AI article tools, and AI interview analysis tools.
Compare papers for a synthesis matrix
Add paper PDFs to Atlas, ask what differs, and inspect cited passages.


