The 15-minute AI claim audit: verify a draft before you publish

A claim audit gives you a repeatable way to separate supported facts from fluent guesses before a draft goes live.

What this audit catches

AI drafts often fail in small, expensive ways. A date is right but the product tier is wrong. A study supports correlation while the draft claims causation. A company announcement gets repeated as an independent result. The prose can sound confident enough that these errors survive a normal copy edit. This audit separates sentence quality from claim quality. You review what the sentence asserts, where that assertion came from, and whether the source supports the exact wording.

The six-column claim ledger

• Claim: copy the smallest sentence fragment that can be checked.
• Claim type: label it fact, number, quote, comparison, prediction, or advice.
• Source: paste the direct URL, document title, or internal record.
• Support: record the exact passage, table, screenshot, or observation that backs the claim.
• Risk: mark what could make the claim misleading, such as an old date, a limited sample, or a vendor source.
• Decision: keep, narrow, replace, attribute, or remove.

Use a stoplight test

Green means the source directly supports the wording and is current enough for the topic. Yellow means the source is relevant but the sentence needs a limit, date, attribution, or softer verb. Red means you cannot find direct support, the source contradicts the sentence, or the source only repeats another page without showing its evidence. A red claim does not become green because three low-quality pages repeat it. Trace the statement back to the original document or cut it.

Minute 0 to 3: isolate the claims

Read the draft once and mark only checkable assertions. Do not spend these minutes polishing introductions or transitions. Numbers, dates, named features, quotes, performance claims, legal statements, medical statements, and claims about what a person or company said always enter the ledger. Advice based on your own method can stay, but label it as advice. The label prevents an editorial opinion from masquerading as a measured result.

Minute 3 to 8: inspect the source in context

Open the primary source when one exists. Search for the exact number, feature name, or phrase. Then read the paragraph before and after it. Context is where most errors show up. A feature may be limited to one plan, a result may apply to one country, or a quotation may answer a narrower question than the draft suggests. NotebookLM can jump from a citation to the relevant source passage, but you still need to read that passage and decide whether your sentence matches it.

Minute 8 to 12: repair the wording

Choose the smallest honest repair. Add a date when the fact can change. Name the company when the evidence comes from its own announcement. Replace "proves" with "found" or "reported" when the source does not establish causation. Split a compound claim if the source supports only half. Remove adjectives such as "best," "leading," and "revolutionary" unless the page explains the comparison and evidence. If the repair makes the sentence clumsy, the claim is probably trying to carry too much.

Minute 12 to 15: run the publication gate

• Every number has a source and a date or measurement period.
• Every quote matches the speaker and surrounding context.
• Every changing product fact has been checked against current documentation.
• Vendor evidence is attributed to the vendor rather than presented as independent proof.
• Advice is identified as a method or recommendation, not a universal result.
• Red claims are removed. Yellow claims have a visible limitation or attribution.

Worked example

Draft claim: "NotebookLM guarantees accurate answers because every response includes citations." The citation feature is real, but the guarantee is not. Google says NotebookLM uses material from selected sources as citations and lets users open the cited passage in context. That supports a narrower sentence: "NotebookLM links answers to passages in the selected sources, so you can open the citation and check the wording in context." The revised claim describes the feature without promising that the answer is always correct.

Copy this audit prompt

• Review the draft only for factual claims. Do not rewrite the prose.
• Return a table with claim, type, source needed, support found, risk, and decision.
• Quote the exact sentence under review. Do not invent a source or complete a missing citation from memory.
• Mark a claim red when direct support is missing. Mark it yellow when it needs attribution, a date, narrower scope, or a limitation.
• For each yellow claim, propose the smallest wording change that the cited evidence supports.

Keep an evidence trail

Save the ledger beside the published draft. Add the date you checked each changing fact. When the article needs an update, review the yellow rows first because they carry the most conditions. This turns maintenance into a short evidence check instead of a full rewrite. It also gives an editor, client, or collaborator a clear reason for each correction.

What not to automate

Do not let a model give itself a confidence score and treat that as verification. Confidence is not evidence. Do not accept a search snippet when the underlying page is available. Do not ask the same model that wrote the claim to approve it without showing sources. AI is useful for extracting claims and organizing the ledger. The publication decision still belongs to the person who can open the evidence and take responsibility for the wording.

Sources checked

• Google Search Central: creating helpful, reliable, people-first content
• NotebookLM Help: use chat and inspect citations

Continue with a related guide

• Turn AI Search Console Insights Into a Content Refresh Queue
• Turn One ChatGPT Record Mode Session Into a Creator SOP, Research Memo, and Gumroad Product Update