Blog / AI & search /

Fact-checking AI drafts: a checklist

The draft lands. Confident dates, neat statistics, and a quote that sounds exactly like something a CEO would say.

Lovely. Except a confident sentence isn’t evidence that anything happened.

Fact-check an AI draft like any high-stakes piece. Pick out the claims, find reliable sources, and make sure the wording matches what the source says. I look hardest at numbers, dates, quotes, product details and first-hand experience. A polished paragraph can still be wrong.

Make a claim list

Read the draft once for meaning, then mark every statement a reader could take as fact. That includes headings, tables, captions, examples and summaries, not only the main paragraphs.

Separate claims from opinion. “This workflow is awkward” is an assessment; “the update was released on a particular date” is a factual statement. Opinions still need to be fair and clearly presented, but they require a different kind of review.

Watch for sentences that bundle several claims together. One line may state a date, describe a product feature and imply a result. Split it into individual claims so each can be checked against evidence.

Verify dates and current details

Go to the primary source: an official release note, documentation, policy page or statement from whoever is responsible. Check its publication date and which version it describes. A help page that’s current today may have said something different on the date your post is set.

For a post written at a particular point in time, ask what the author could have known then. Do not import a feature, policy or product name from a later update. A timeline note or dated announcement is often more useful than a search result snippet.

When a claim cannot be confirmed, remove it, qualify it or ask the person who owns the information. Don’t keep a precise date just because it makes the paragraph sound authoritative.

Check numbers and comparisons

Find the original data source and inspect what was measured, who was included and over what period. A percentage without a denominator or baseline can be misleading. A survey result may describe respondents, not an entire market. A benchmark may apply only to a particular device, region or test setup.

Recalculate simple percentages when the source provides the underlying values. Check units, currencies, date ranges and whether the number is an estimate. If the source says “up to” or “approximately”, do not turn that into a guaranteed result.

Never present a made-up example as a real outcome. If a hypothetical helps explain something, label it and don’t give it invented numbers. And if a story needs real experience, I’d rather leave it out than let a model supply one.

Check quotations and attribution

Open the source and read the surrounding context. Confirm the person, wording and date. AI systems may combine fragments, attribute a general statement to a specific person or produce a sentence that appears in no source at all.

If you cannot find the exact quotation, do not present it as a quote. You may be able to paraphrase the source accurately, but cite the source and preserve its meaning. Avoid quotation marks around a phrase that is merely a summary.

Make sure the source supports the sentence, not just the general topic. A page about an algorithm update does not necessarily support a claim about what caused one particular site to lose traffic.

Verify product and technical instructions

For software, check the official documentation for the correct version and interface. Test steps in a safe environment when the instructions could change a live site, delete data or affect tracking. Confirm menu names, settings and prerequisites rather than relying on a screenshot from a different version.

For product descriptions, compare specifications with the manufacturer’s information and the store’s current catalogue. Check dimensions, compatibility, warranty terms, availability and exclusions. Ask the product owner when the source material conflicts.

For SEO recommendations, distinguish documented guidance from interpretation. Google’s public documentation may state what a system does, while a particular ranking explanation remains a hypothesis. Label it accordingly.

Check personal experience with the person

A model will happily write “I tested”, “I saw” or “in my client work” because those phrases sound credible. They stay only if the author really did it. No invented client stories, no before-and-after results, no made-up operational detail.

If an experience would strengthen the post, get the real detail from the person who had it. If there isn’t one, use a clearly hypothetical case or keep it general.

Finish with an editorial decision

For each claim, record the source, whether it supports the statement and what change you made. Remove unnecessary claims that are hard to verify. The goal is not to preserve every generated sentence; it is to publish only what the reader can trust.

After fact-checking, do a separate read for usefulness and voice. A draft can be accurate and still generic, repetitive or irrelevant. Check that the opening addresses a real problem, that each section adds something and that the conclusion gives the reader a next step.

AI-assisted keyword research raises similar judgement questions; see AI-assisted keyword research: where it goes wrong and AI content detectors: why I don’t rely on them.

Take the draft’s most convincing factual sentence and trace it to a primary source. Which other claim needs the same treatment before you publish?