Editorial guide · prepared with AI assistance · Evidence and methods
How to verify sources and citations, one claim at a time
Check whether a citation supports the exact claim, not just whether its link works. Use a claim-level audit for research, product reviews and AI-generated answers.
To verify a citation, open the original source, find the passage that supports the claim, and compare the claim’s subject, result, date and conditions with that passage. Then assess the source itself. A working link or a relevant title is not enough: the cited evidence may support only part of the sentence or a narrower conclusion.
Separate the statements inside a sentence
A sentence can contain several claims: a product has a feature, the feature improves performance, and the improvement makes the product the best choice. One source may confirm only the feature. If you check the sentence as a single unit, the supported detail can make the unsupported recommendation seem verified.
Copy the statement you need to use and split it into the smallest meaningful factual claims. Keep conditions attached: the product version, population, period and comparison may determine whether a statement is accurate. Then attach evidence to each claim rather than to the paragraph as a whole.
FActScore, a 2023 research evaluation, uses atomic facts to measure how much generated text is supported by a knowledge source. This article borrows the practical idea of separating claims; it does not reproduce that evaluation or claim that a manual checklist yields a FActScore. [S28]
Check existence, support and reliability separately
Research on attribution makes a related distinction: whether generated statements are supported by identified sources is a question that can be assessed. Google Research’s AIS work proposes a human evaluation framework for that task. Support by an identified source should not be confused with proof that the source itself is correct. [S26]
| Check | Question | Possible failure |
|---|---|---|
| Existence | Does the cited document exist and match its bibliographic details? | A plausible reference points to a different paper |
| Local support | Does the identified passage support this exact statement? | The source discusses the topic but reports no such result |
| Scope | Do version, population, setting and date match? | An old laboratory result becomes a current product claim |
| Source quality | Is the method appropriate and inspectable? | A source repeats an unsupported assertion |
| Independence | Are several links genuinely separate evidence? | Many pages repeat the same original press release |
| Completeness | Does omitted evidence materially change the conclusion? | A favorable result is quoted without its limiting condition |
A hypothetical citation that supports only half a claim
Imagine a report states: Tool A exports CSV files and therefore meets every data-portability requirement. Its citation leads to official documentation confirming CSV export. In this hypothetical case the first claim is supported, while the second still needs evidence about required fields, attachments, permissions and the intended receiving system.
Repair the sentence by keeping the supported fact and naming the gap: the documentation lists CSV export, but the review has not established that the export preserves all required data for the buyer’s migration. The source can remain, attached to what it actually establishes.
If a practical test later checks those fields, cite the test record separately and state the tested version and sample. Do not silently upgrade a documentation review into a hands-on finding. A correct reference can accompany an incorrect inference, so inspect the reasoning between the source and the conclusion.
Follow a statistic back to its origin
A newspaper article, company blog and social post may all cite the same study. Find that study and identify the reported metric, denominator, comparison and dates. If you can read only an abstract, stay within what the abstract supports. Do not supply detailed methods, subgroup findings or causal explanations that you have not checked.
For percentages, ask what was counted and what was excluded. A percentage of accepted submissions is different from a percentage of all attempted submissions. Relative improvement is different from a percentage-point change. Save enough information that another reader can reconstruct the calculation instead of relying on the number’s apparent precision.
Check for a newer version, correction or changed product before reusing an old claim. An archive can establish what a page said at a particular time; it does not make the archived statement current or correct. Preserve the observation date and explain why a historical source is still relevant to your question.
Verify an AI answer against its actual sources
For an AI-generated answer, inspect the real links supplied by the product. Do not assume a model’s explanation of its research is a record of the steps it actually performed. Open the cited page, locate the relevant passage and mark any claim that lacks a supporting source as unresolved.
The ALCE research benchmark evaluates cited generation using separate dimensions that include correctness and citation quality. That distinction is useful when reading an answer: fluent wording, a citation and a correct conclusion are different properties. The 2023 benchmark does not establish today’s error rate for any search product. [S17]
Keep a compact audit record: exact claim, source URL, passage location, source version or access date, verdict and limitation. Useful verdicts include supported, partly supported, contradicted and not checked. These are this guide’s editorial labels, not an automated scientific metric.
Use the audit to improve the decision
Begin with the claim that would most change your decision. If its evidence is strong and applicable, move to the next material claim. If it is unsupported, either obtain better evidence, narrow the wording or avoid relying on it. Checking every link while leaving the decisive inference unexamined creates a detailed record of the wrong task.
When publishing the result, keep the source beside the supported statement and carry its material limits into any summary. A reader should be able to follow the same path from the claim to the passage and understand what remains unknown.
Sources and method
NamedProof editorial guide. Prepared with AI-assisted research and synthesis; the cited primary-source passages were checked for this guide. This is not a commissioned expert assessment and has not received human expert review. Examples labelled hypothetical are teaching examples, not client results.
Limitations
- Atomic-claim decomposition here is a practical editorial approach, not an execution of FActScore, AIS or ALCE.
- Research papers are cited for their stated task definitions; no historical benchmark percentage is presented as current model performance.
Original contribution
A six-check citation audit and a hypothetical CSV-export claim showing the difference between a correct source and an unsupported inference.
Sources
- S28
FActScore: Fine-grained Atomic Evaluation of Factual Precision
Scope: Original evaluation, including generated biographies.
Limitations: Precision differs from completeness, expertise and independence. Historical rates are not current model performance.
Material read: Abstract only; full methods/results NOT read
Sources checked:
- S26
Measuring Attribution in Natural Language Generation Models
Scope: Attribution in tested NLG datasets/tasks.
Limitations: Support against supplied sources does not establish those sources contain ultimate truth; abstract-level finding only.
Material read: Official research abstract read
Sources checked:
- S17
Enabling Large Language Models to Generate Text with Citations (ALCE)
Scope: ALCE datasets and tested 2023 systems.
Limitations: Evaluates supplied-corpus citation support under selected datasets and systems, not present-day general web product reliability. Citation support is distinct from ultimate source truth.
Material read: Full-text selected task, citation-quality and limitation sections read; not whole paper
Sources checked: