Guide
How to fact-check AI content before you publish
To fact-check AI content, collect your sources before the draft is written, list every factual claim, and check each one against the exact passage of its cited source, high-risk claims first. Then check that time-sensitive claims are still current and that every citation exists and its link resolves, and hold publishing until the draft passes and a person has reviewed it. uprank automates most of this method: it gathers an Evidence Pack before drafting, checks each claim against it, and blocks publishing when any of its 13 hard gates fails. Reading the draft stays with you.
Updated
The method, step by step
Works on any AI draft, whichever tool wrote it.
Step 1
Collect your sources before the draft is written
Gather attributable sources first, primary or official ones where you can, and write from them. Sourcing after the fact makes it easy to bolt links onto claims they do not actually support.
Step 2
List every factual claim in the draft
Pull out each statistic, date, name, quote, comparison and "best", "first" or "only" statement. Opinions can stay as opinions; anything a reader could check cannot.
Step 3
Check each claim against the exact passage of its source
Open the source and find the passage. Mark the claim as supported, partly supported, contradicted or unsupported. A source that is merely on the same topic does not support the claim.
Step 4
Check the high-risk claims first
Medical, financial, legal, safety and reputation claims cost the most when they are wrong. Fix or cut any of them that is unsupported or contradicted before you spend time on the rest.
Step 5
Check time-sensitive claims for freshness
Prices, dates, versions and anything described as current can expire. Give each one a window, and re-check anything that is past it.
Step 6
Confirm every citation exists and every link resolves
Open each link. Check the source is real and that its title, publisher and date match what the draft says. A citation to something that cannot be found counts as an invented one.
Step 7
Hold publishing until it passes, with a person reviewing the draft
Save the article as a draft, have a human read it, and treat any edit as a reason to re-check. Publish only when every check above has passed.
What uprank automates at each step
Each step of the method maps to a stage of the pipeline or a gate that blocks publishing.
- Collect your sources before the draft is written. Research runs before drafting and produces an Evidence Pack of attributable sources. The draft is then written from that pack, with research and drafting as separate pipeline stages.
- List every factual claim in the draft. After the draft, the pipeline extracts the checkable claims from it, each with a claim type and a risk level.
- Check each claim against the exact passage of its source. Each claim is checked against the numbered source excerpts in the Evidence Pack and marked directly supported, partly supported, contradicted or unsupported. Contradicting sources are recorded, not dropped. The Evidence score, which weighs these results, has a minimum (≥ 85 by default).
- Check the high-risk claims first. Medical, financial, legal, safety and reputation claims are rated at least high risk. Any high-risk claim that is unsupported or contradicted blocks the article, and the Risk score has a maximum (≤ 20 by default). Gates: High-risk claims are supported.
- Check time-sensitive claims for freshness. A time-sensitive claim carries a freshness window in days. A claim outside its window blocks the article. Gates: Time-sensitive claims are fresh.
- Confirm every citation exists and every link resolves. Every citation must point to a source in the Evidence Pack; one that does not is an invented citation and blocks the article. uprank also checks that each cited link resolves: one that returns not found, does not resolve or points to a private address blocks the article. Sites that refuse automated checks (401, 403, 429) or return server errors are counted as unknown, not broken, so open those yourself in review. Gates: No invented citations; No broken required links.
- Hold publishing until it passes, with a person reviewing the draft. Drafts go to your CMS as drafts by default. Editing an article makes its earlier score stale, so it must be re-scored before it can be published. Live publishing needs an explicit confirmation and is audited, and over MCP, publication_publish refuses a version that has not passed the gates.
- Also enforced before publishing. Beyond fact-checking: At least one information-gain element; Licensed images only; Unique title. These integrity gates apply to every article.
Pre-publish checklist
One question per hard gate, 13 in all. Publish only when the answer to every one is yes.
Evidence. Is every factual claim backed by a source you have opened and read?
Information gain. Does the article tell the reader something the top results for this topic do not already say?
Practical value. Could a reader act on it after one read, with steps, examples and specifics rather than general advice?
Brand fit. Does it sound like your brand and stay within your positioning and proof points?
Readability. Can someone in your audience follow it in a single pass, with plain words and clear headings?
Risk. Have you ruled out anything that could mislead readers, cause harm or damage your reputation?
No invented citations. Does every citation point to a source that exists and says what the draft says it does?
No broken required links. Does every cited link open on the page it is meant to?
High-risk claims are supported. Is every medical, financial, legal, safety or reputation claim supported by the passage in its source?
Time-sensitive claims are fresh. Is every price, date, version and "current" claim still true today?
At least one information-gain element. Does it include at least one element of its own, such as a framework, worked example, checklist, comparison or template?
Licensed images only. Does every image that is not yours come with a license you have checked?
Unique title. Is the title different from every other article on your site?
Common questions
- Can I use AI to fact-check AI-written content?
- It can do the legwork, such as matching each claim to a passage in a source, but only if it checks against the actual source text rather than its own memory. A person should still review the high-risk claims. uprank’s claim check works from the numbered source excerpts in the Evidence Pack and does not bring in outside sources, and its drafts go to your CMS as drafts by default so a person reviews them.
- What counts as a high-risk claim?
- Claims where being wrong costs the most: medical, financial, legal, safety and reputation claims. uprank rates those at least high risk, and an unsupported or contradicted high-risk claim blocks the article from being published.
- How do I fact-check a statistic in an AI draft?
- Find the original source rather than a page that repeats it. Confirm that the number, what it measures, who it covers and the date all match how the draft uses it. If you cannot find the source, reword the sentence or remove it.
- Does passing a fact-check mean the article is correct?
- No. It means each claim matches its source. It does not prove the sources themselves are right, so prefer primary and official sources and keep a human review step. uprank checks claims against the sources it gathered; it does not independently verify those sources.
Keep reading
- Quality gates: every dimension, weight and hard gate.
- FAQ: product, agents, tokens and publishing.
- AI writer that doesn’t hallucinate sources
- AI blog writer for Strapi
- Agent setup guides: connect a coding agent over MCP.