How to add citations to AI-generated content

1. Step 1

   Decide which statements need a source

   Statistics, dates, names, quotes, comparisons, "best", "first" or "only" statements and anything a reader could check need a source. Opinions and your own first-hand experience do not, but they should read as yours.

2. Step 2

   Gather your sources before the draft is written

   Collect attributable sources first, primary or official ones where you can, and note the author or organisation, title, publisher, publication date and URL of each. Finding links afterwards makes it easy to attach them to claims they do not support.

3. Step 3

   Give the AI only those sources and tell it to cite only them

   Paste or attach the source text, number the sources, and ask the model to cite by number and to say so when a statement has no source. A model asked to cite from memory can produce a reference that looks real and points to nothing.

4. Step 4

   Attach each claim to the specific passage that supports it

   One claim, one source, and the place in that source where it says so. A source that is merely on the same topic does not support the claim, and a link at the end of a paragraph does not say which sentence it backs.

5. Step 5

   Pick one citation format and use it throughout

   An inline link, a numbered reference or a footnote all work. What matters is that every citation names the source clearly enough that a reader can find the same passage, and that the format does not change halfway through the article.

6. Step 6

   Decide what to do when no source supports a claim

   Find a source, reword the claim to say only what your sources say, present it as your own view or experience, or cut it. Do not leave a factual statement uncited, and do not swap in a link that is only loosely related.

7. Step 7

   Keep the source list with the draft

   Store the sources and the claim-to-source mapping next to the article, with the date you accessed each source. The reviewer needs them to check the draft, and whoever updates the article later needs them to re-check what has changed.

This page is about sourcing. Once a draft exists, checking that each claim matches its source is a separate job: how to fact-check AI content.

• Who: the author, or the organisation when there is no named author.
• What: the title of the page, report or paper.
• Where: the publisher or site, and a URL that opens that exact page.
• When: the publication date, plus the date you accessed it for anything that changes.
• Which part: the section, page or passage when the source is long.

1. Option 1

   Look for a better source

   Search for the original publication rather than a page that repeats it. If an official or primary source states it, cite that one.

2. Option 2

   Reword the claim to match what the sources say

   If the sources support a weaker or narrower statement, write that one. Drop the exact figure, the superlative or the "always" the sources do not back.

3. Option 3

   Present it as your own view or experience

   If it is your judgment or something you saw first-hand, say so plainly. Do not dress an opinion up as a sourced fact.

4. Option 4

   Cut it

   If it adds little and you cannot support it, remove it. An article is stronger without a claim it cannot back.

• Decide which statements need a source. The Evidence Pack records the search intent, the reader problem, the required sections and the required examples next to the sources, all produced before drafting.
• Gather your sources before the draft is written. Research runs before drafting and produces an Evidence Pack. On the article’s Evidence tab each source shows its publisher, author, publication and fetch dates, a credibility score and a primary-source flag, wherever they are known.
• Give the AI only those sources and tell it to cite only them. The draft is written from the Evidence Pack. Its sources are numbered, and the draft cites them by number. The drafting instructions say to ground every factual statement in those sources and never to fabricate statistics, quotes or citations.
• Attach each claim to the specific passage that supports it. After drafting, claims are extracted from the draft and each extracted claim is checked against the numbered source excerpts and marked directly supported, partly supported, contradicted or unsupported. The Claims tab lists each extracted claim with the sources that support or contradict it, contradicted claims first, and a claim no source supports shows no linked sources. Over MCP, the claims resource returns the same claims with their supporting sources.
• Pick one citation format and use it throughout. A citation is its own block in the article: a label and, when it points to a source in the pack, a link to that source. In Markdown it is a link and in HTML an anchor. When an article is pushed to Strapi the citations travel in the article body as links; the separate citations field is not filled by the Publishing tab. uprank writes citations as labelled links and does not format footnotes or reference lists.
• Decide what to do when no source supports a claim. A claim no gathered source supports is marked unsupported, even if it happens to be true. An unsupported or contradicted high-risk claim blocks the article, and other unsupported claims pull the Evidence score down. The draft cites pack sources by number, and a citation numbered outside the Evidence Pack counts as invented and fails a hard gate. Gates: High-risk claims are supported; No invented citations.
• Keep the source list with the draft. The Evidence Pack, the claims with their supporting and contradicting sources, the scores and each article version are stored with the article. Every link in a citation block is checked each time the article is scored, and one that returns not found, does not resolve or points to a private address blocks the article. Over MCP, the evidence pack and claims resources are readable with the article:read scope. Gates: No broken required links.

The full list of gates is on quality gates, and the product view is an AI writer that doesn’t hallucinate sources.

• Prove a source right. It checks claims against the sources it gathered, so a wrong source can still support a claim.
• Catch every wrong citation. The invented-citation check applies to citation blocks. A reference written in running text, rather than as a citation block, is not checked against the Evidence Pack, so check those yourself.
• Replace a reviewer. Articles go to your CMS as drafts by default so a person reads them, and the claim check is itself automated.

Can AI add citations to its own writing reliably?

Not from memory. A model can write a reference with a plausible title, publisher and URL that points to nothing. It can cite reliably only when it is given the actual sources to cite and something then checks that every citation exists. uprank has the draft cite the sources in its Evidence Pack, gathered before drafting, by number, and counts a citation numbered outside the pack as invented.

Should I add citations before or after the AI writes the draft?

Before. Gather the sources first and write from them. Links added to a finished draft are easy to attach to claims they do not actually support, which is why that route needs a separate fact-check afterwards.

Is a link at the end of a paragraph enough?

Only if every claim in the paragraph comes from that one source. If a paragraph makes several claims, attach each to its own source so a reader knows which source backs which sentence.

How does uprank add citations to an article?

It researches into an Evidence Pack before drafting, numbers the sources, and has the draft cite them by number. Each citation becomes a block with a label and, when it points to a source in the pack, a link to that source. Every link in a citation block is checked when the article is scored, and a citation numbered outside the pack blocks the article. You can inspect the sources on the Evidence tab and the sources that support or contradict each claim on the Claims tab before anything is published.

• How to fact-check AI content: check each claim against its source before you publish.
• AI writer that doesn’t hallucinate sources
• Quality gates: every dimension, weight and hard gate.
• FAQ: product, agents, tokens and publishing.
• Agent setup guides: connect a coding agent over MCP.