How to Build a Source Comparison Matrix That Survives Fact-Checking

A matrix becomes clutter when it tries to compare everything. Start with the output it must support: choosing a method, reconciling two explanations, preparing an interview, or checking claims for a script. Write three to seven questions whose answers would change that output. Those questions become row families.

Record inclusion rules for sources. A first-party policy page may answer current product terms; a tutorial video may show workflow; a research paper may answer a measured outcome. Their authority is claim-dependent. The matrix should record what each source can reasonably establish rather than assigning one universal credibility score.

For each source, store identity, date, version, format, and access path. For each claim cell, store a concise source position, page or timestamp, quotation only when exact wording matters, scope conditions, and verification status. Keep your interpretation in a separate column so readers can distinguish what the source says from what you conclude.

Cochrane's data-collection guidance is designed for systematic reviews, a more formal job than most creator research. Its useful transferable principle is to predefine relevant characteristics and retain enough methodological context to compare studies without stripping away design differences. Use the discipline without implying that a lightweight content matrix is a systematic review.

A row called “benefits” invites each source to discuss a different benefit. Replace it with precise propositions: “The method reduces editing time,” “The method improves locator traceability,” and “The method supports two source types.” A source can then support one row, qualify another, and remain silent on the third.

Do not put several conditions into one cell. If a source supports a claim only for a specific population or date, store the condition beside the position and mark the row qualified. This avoids flattening “sometimes under these conditions” into a green checkmark.

Fill one source column at a time to preserve its argument and terminology. Ask an AI system to propose candidate passages only within the selected source, then open every locator used in a material row. Once all columns are checked, switch perspective and read across each row. That second pass reveals whether the sources truly address the same proposition.

If two sources use the same word differently, add a definition row before comparing outcomes. If one reports a measured result and another offers an opinion, record the evidence type rather than averaging them. The point of the matrix is to make non-comparability visible, not to force a winner.

When cells conflict, check dates, versions, populations, definitions, methods, and whether one source is summarizing the other. Add a short conflict note that names the likely reason and the evidence needed to resolve it. Do not ask AI to harmonize positions before those checks; a smooth synthesis can erase a real boundary.

Use “silent” when a source does not address the claim and “not checked” when the review is incomplete. Neither means false. This distinction matters in public writing because “Source B does not support the claim” is different from “Source B contradicts the claim.”

Turn each accepted row into an outline claim. Link the sentence back to the exact cells that support it, and keep qualifications in the same paragraph as the claim. Rows with unresolved conflict can become a limitation or an explicit comparison rather than disappearing from the story.

Before release, sample every high-impact conclusion and follow the chain from draft to row, cell, locator, and original source. A matrix is valuable only if the trail works. Broken links, unexplained color codes, and cells that contain generated prose without locators should block the related sentence from publication.