AI content
How to fact-check an AI-written article
List every claim in the draft that could be checked: numbers, dates, names, prices, quotations, anything that begins "studies show". For each one, find the original source, open it, and confirm it says what the draft says. Rewrite what is off and cut what you cannot trace. It is slower than producing the draft was, and it is the part of publishing with AI that cannot be skipped.
What it caught in our own drafts
The guides on this site were drafted by an AI model. Before publishing, we sent 16 sets of claims to be checked against the primary source: each company's own documentation, Google's policy pages, the research paper itself.
Five came back confirmed as written. One was wrong. One had no source we could find. The other nine were right in outline and off in a detail.
The three worth describing:
"ChatGPT's page fetcher obeys robots.txt." The draft said each OpenAI crawler follows its own robots.txt rules. OpenAI's documentation says that for visits a person asks for, robots.txt rules "may not apply". Our crawler checker's explanation had to change.
"Cloudflare blocks AI crawlers by default on new sites." Cloudflare announced something like that in July 2025, and it was widely reported. Its current documentation says new domains allow them by default. The draft was repeating a headline that had since been overtaken.
"Google has said it does not use llms.txt." This is repeated all over the web. We could not find Google saying it. The closest primary source is a line in Google's guidance that you do not need "AI text files" to appear in its AI features, which is a narrower statement. We now quote that line and attribute nothing more.
All three sounded right. All three are things you will find stated confidently on other sites, which is exactly why a model reproduces them. The lesson we took is that the dangerous errors are the plausible ones that everybody repeats.
The nine smaller fixes were of a kind: a figure quoted from a paper's abstract that the abstract does not contain, a policy paraphrased more broadly than its text, a two-year-old study presented as if it were current.
The method
1. Pull out the checkable claims. Go through the draft and mark every number, date, name, price, quotation and product fact, and every sentence that leans on "research shows" or "experts say". Opinions and advice are not on the list.
2. Find the primary source for each. The primary source is where the fact was first established: the company's own pricing page, the regulator's document, the study itself. A blog post that mentions the study is not it. If the draft gives a link, start there, and check whether it is the original or a page quoting the original.
3. Open it and read the sentence. Not the search-result snippet. Check that the page loads, that it says the thing, and that the number is the same number measured the same way. "Up to 40%" and "40%" are different claims.
4. Check the date. Prices, product features, policies and anything about AI change within months. If the source is more than a year old, look for a newer one or say when it is from.
5. Check claims about your own product against the product. Models describe features you do not have. Nobody outside your company can catch these.
6. Fix, soften or cut. If the source says something narrower, write the narrower thing. If you cannot find a source, the claim goes, however good it sounded. A sentence without the statistic is nearly always fine.
What this does not catch
A claim can be accurate and misleading: a real number from a study of 40 people, a real quote with its qualification removed. Checking the source catches the first if you read past the headline and will miss the second unless you read around the quote.
It does not catch what is missing either. A draft can be correct in every sentence and leave out the one fact that changes the conclusion. That takes someone who knows the subject.
And it does not make advice good. "Publish three times a week" is not a factual claim, so no source will confirm or refute it.
How our product handles it
Each article is researched with live web search before it is written. A link is kept only if the research step read that page or it is on your own site, and dead links are removed. After drafting, every claim is listed next to the page it came from, and claims with no source are flagged on the review screen for you to fix or delete.
That removes most of steps 1 to 3. Steps 4 to 6 still need a person now and then, particularly step 5.
Questions
- Can I ask the AI to fact-check itself?
- It helps a little and it is not enough. A model asked to verify its own claim often agrees with itself. It works better when the checker has to fetch the source and quote the passage, which is how we ran the check described above.
- What if a claim is true but I cannot find a source?
- If you know it first-hand, say that: "in our experience" or "when we tested this". If you do not, leave it out. An unsourced fact and an invented one look the same to a reader.
- How do I know a source is the primary one?
- Ask who measured or decided the thing. If the page you are reading says "according to", follow that until you reach a page that does not.
Sources
Each one was opened and read on 14 November 2026. Figures in the text say what was counted, how many and when.
Content Autopilot team
We make Content Autopilot, an SEO and AI visibility product
Our guides are drafted with AI from the sources listed above, and every figure is checked against its source before it is published. Where we sell something a guide discusses, the guide says so.