What to Do When Research Sources Disagree

Neutron Says Team · October 21, 2026 · 7 min read · Research, AI research, source evaluation, research methods

Check whether sources describe the same thing, trace their evidence and make unresolved disagreements visible. Includes an evidence card and fictional example.

An AI answer may make two incompatible sources sound as if they agree. Before choosing a conclusion, put the claims next to each other and check whether they describe the same thing. A difference in date, plan, definition or population can explain an apparent conflict. Sometimes the evidence really is inconsistent, and the useful answer is to say what remains unresolved.

This workflow is for ordinary research and documentation checks. It helps you make the reasoning inspectable; it does not turn an AI response into an authority. The note-taking product, documents, dates and limits in the worked example below are entirely fictional.

1. Write down the disputed claim

Start with one question narrow enough to check. “Which source is right?” is usually too broad. “What maximum attachment size do the current public documents specify for the Free plan?” gives you something concrete to compare.

Copy the relevant passage from each source into private research notes, keeping enough surrounding context to preserve qualifications. Record the page title, publisher, URL, publication or revision date if shown, and the date you read it. Quote only the amount you need and respect the source's reuse terms when sharing your work.

Separate the source's statement from your interpretation. A page saying that a feature is available on Pro does not establish that it is absent on every other plan. A successful test on one account does not establish availability for all accounts.

Purdue OWL's guidance on synthesis distinguishes comparing relationships across sources from simply summarizing them one after another. It also cautions against forcing relationships or ignoring outlying evidence. The comparison process here applies that principle to a practical research note.

2. Check whether the claims have the same scope

For each disputed statement, ask:

Subject: Is it the same product, version, plan, place or group?
Measure: Are the units and definitions the same? Is a limit per file, per account or per month?
Time: Is the source describing a past state, a current state or a planned change?
Conditions: Are there eligibility rules, exceptions, staged rollouts or configuration requirements?
Evidence: Is this a direct observation, an official description, a secondhand report or someone's interpretation?

Do not combine numbers just because they share a label. A per-file limit and an account storage allowance answer different questions. Averaging them produces a third number that neither source established.

A newer date is a useful clue, not an automatic verdict. A recent comparison article can repeat an old help page. A revised page may not say which section changed. Keep the date and its meaning visible rather than assuming that the newest-looking page answers your exact question.

3. Trace the evidence back to its origin

Several pages can repeat one underlying source. Follow their references before treating them as independent confirmation. If a blog and a comparison page both cite the same release note, you have two retellings and one original release statement.

Use the source best placed to establish the particular claim. A product's documentation describes what its publisher documents; an independent hands-on test describes what the tester observed under stated conditions. Neither automatically answers the other's question. For example, documentation may describe a staged release that a particular account has not received.

USC Libraries' source-evaluation guide recommends examining authorship, dates, revisions, audience and the evidence behind a claim. Apply those questions to the exact passage, not just to the site's appearance or reputation.

If the publisher is unfamiliar, check who is behind it using sources outside that website. Digital Inquiry Group's lateral-reading guidance describes leaving an unfamiliar site to investigate what other sources say about it. This helps you evaluate provenance; it does not prove every statement on an established site.

4. Work through a fictional example

Suppose you are checking the attachment limit for the Free plan of a fictional service called Example Notebook. You find three invented documents:

Help page A, revised September 1: “Free accounts support attachments up to 100 MB per file.”
Release note B, published October 1: “Pro accounts now support attachments up to 200 MB per file.”
Comparison post C, published October 5: “Example Notebook supports 200 MB attachments.” Its reference leads to release note B.

These passages do not establish a change to the Free plan. A and B describe different plans. C drops B's qualification and supplies no independent evidence for a Free-plan limit.

A supportable note would be: “The supplied Free-plan help page documents 100 MB per file. The 200 MB announcement applies to Pro. The comparison post cites that announcement and does not establish a Free-plan increase.” This describes the documents; it does not claim you tested an account or verified that the help page is complete.

Now imagine a fourth fictional document:

Free-plan FAQ D, revised October 15: “Free accounts can attach files up to 200 MB each.” It gives no version or rollout details.

A and D now disagree on the same apparent plan and measure. Do not silently discard A, average the values to 150 MB, or treat D's date as proof that every account has changed. Record the discrepancy and identify what would settle it: an explicit superseding notice, clarification of plan/version coverage, or a permitted test on the relevant account.

Until then, the conclusion should stay narrow: “The supplied public documents conflict on the current Free-plan limit. The available passages do not explain whether one supersedes the other.”

5. Keep one evidence card per disputed claim

Use a compact record you can hand to another reader:

Claim and question: State exactly what you are trying to establish.
Source A and source B: Give titles, publishers and direct links.
Relevant passages: Preserve the conditions and units, with enough context to understand them.
Dates: Separate publication/revision dates from your access date.
Scope: Record plan, version, location, population or other relevant boundaries.
Relationship: Note whether the sources are independent or one repeats the other.
Status and reason: Choose a plain description such as “different scope,” “explicitly superseded,” or “unresolved,” then explain why.
Next check: Name the specific missing evidence, not just “research more.”

Those status labels organize your notes; they are not a statistical confidence score. A missing date or an unanswered question should remain visible. If the sources use different definitions, preserve each definition instead of merging the statements into a false consensus.

6. Ask AI to expose the conflict before drafting

If you use Neutron Says or another AI assistant to organize public research notes, provide the relevant passages and ask for a comparison before a recommendation. Remove private material that the task does not need.

A useful instruction is:

Compare these supplied passages against this exact question. For each claim, list its subject, scope, date, units, source and qualifications. Identify differences in scope and any genuine contradictions. Separate direct statements from inference. Do not choose a winner, invent a superseding update or fill a missing field. State what evidence would resolve each remaining conflict.

Check the comparison against the original pages yourself. A tool may omit a qualifier, attach a date to the wrong claim or mistake a copied report for independent evidence. Ask for a revised comparison when you find that error; repeating the same conclusion in several AI answers does not create another source.

When the evidence supports a conclusion, explain the deciding distinction in the final note. When it does not, show the competing claims, the practical implication of the uncertainty and the next check. A reader should be able to see how you reached the result without trusting the drafting tool.

Use the research-brief template to define the question and stopping point before searching, and the citation checklist to verify each reference. This comparison step connects the verified sources without pretending they say more than they do.