5 min read

How to fact-check AI-written content before you publish

A claim-by-claim process for checking AI drafts: list every claim, confirm it against a real source, cut what you can't confirm, and keep a record of what you removed.

Fact-check AI content one claim at a time. List every checkable statement, confirm each against a source you've opened yourself, rewrite what the source only partly supports, and cut what you can't confirm. Then log what you removed. Google says to manually fact-check all AI-generated content before publishing, because generative models predict likely words rather than retrieve facts. Google's own Search Central guidance on AI-generated content is the place to start.

Why a read-through isn't enough

A fluent paragraph reads as true whether or not it is. Skimming catches awkward phrasing, not a wrong date or an invented statistic. A claim-by-claim pass forces each statement to earn its place.

The review also has to reach beyond the visible article. Titles, meta descriptions, structured data, and image alt text can appear in Search results, so an error there shows up under your site's name. Structured data is the easiest to miss: a price, rating, or date there sits in code, so an editor reading the rendered page never sees it.

The steps

  1. List every claim. Go through the draft and mark each statistic, figure, name of a person, place, or title, date, product detail, and any sentence that says something is true. This gives you a finite checklist instead of a vague instruction to "review it." Include the headline, meta description, alt text, and structured data in the list.

  2. Find the source for each claim. If the draft names where the information came from, you can validate it faster. If it gave no source, search for the statistic or statement yourself. This step tells you which claims have support and which are unsupported on arrival.

  3. Open the source and match it to the exact sentence. A link only helps if the page behind it backs your exact statement, so read it and check that it says what your draft says, not merely that it covers the same topic. This catches the most common AI failure: a plausible link that doesn't back the claim.

  4. Rank the sources. Prefer government bodies, non-partisan research institutions, and primary documents over summaries and reposts. Check that the details are current, especially for fast-changing topics like technology or news. This keeps an outdated figure from passing just because it was once true.

  5. Match the strength of the wording. If the source says "may," don't upgrade it to "will" because certainty sounds better in a headline. Rewrite the sentence to say exactly what the source says. This removes overstatement, which is often the error that survives step 3.

  6. Confirm company facts with the person who owns them. Services, prices, hours, policies, people, locations, and product details should be checked with the current internal owner of that information. A model can't know what changed last week.

  7. Send specialized claims to an expert. Medical, legal, financial, or technical details need someone qualified to read them. Minor errors can carry significant consequences in specialized areas.

  8. Remove what you can't confirm. If you still can't validate a claim after a real search, reconsider using it. Delete it or replace it with something you can support. Don't ask the model that wrote the draft to vouch for itself; checking its work without independent sources is the first mistake to avoid.

  9. Cite what remains. Link to credible sources wherever the article quotes data, so readers can check it too. Link only for numbers, studies, rules, and quotes, not for common knowledge. Prefer the original source over a news write-up about it.

  10. Record the review. Note who checked the page, which sources were used, and when. Add a short list of the claims you removed and why. That creates accountability and makes the next update faster, and the removal list shows which kinds of claims your drafting process tends to invent.

A worked example

Say a draft contains: "Most buyers now ask AI assistants for vendor recommendations, and 80% trust the answer." Run it through the steps.

  • The statement contains a figure and a broad claim, so both go on the list.
  • No source is attached. You search the 80% figure and find nothing that matches it.
  • You find a survey on a related topic, but it measures something different. A matching topic isn't a matching claim.
  • You can't validate the number, so it comes out. The remaining sentence gets softened to what you can support, or removed if you have nothing.
  • The log records: "80% trust figure, no source found, removed."

The final sentence is shorter and weaker, and it is true. That trade is the point of the process.

Doing this at scale

Manual checking works for a few posts a month. It gets slow once a team publishes weekly. At RootSmith, we run this same loop for every page: we research, draft in your brand voice, check each claim against its source, remove claims that can't be verified, and report what was removed. A person on your team still approves every draft before it publishes to your own domain. The result is a draft that arrives with its removals already listed, so your reviewers spend their time on judgment calls instead of hunting for sources.

Make it routine

Pick one owner for the final read, keep the removal log in the same place as the draft, and don't publish until every item on the claim list is confirmed, rewritten, or cut. A short checklist you run the same way every time will catch more than a careful read done differently each time.

If you'd rather not run this loop by hand, see how RootSmith works.

See what AI says about you

Start free. Pick a plan when you're ready for tracking and content.