AI Answer Link Checker - Find Hallucinated URLs & Fake Quotes

Paste an AI answer. This Actor finds every link in it, checks that each URL really exists, and tests whether the page contains the sentence or quote the link was cited for. You get a status and a grounding score per claim, and a Wayback Machine copy when a link is dead. Built for people who build RAG and agent products, run AI content teams or fact-check chatbot output, and who have learned that "the link returns 200" does not mean "the page says that".

Input examples

1. A chatbot answer with links

{ "answer": "Turing \"was an English mathematician and logician\" ([source](https://en.wikipedia.org/wiki/Alan_Turing)). Productivity rose 47% in 2024 [1].\n\n[1] https://example.com/reports/2024-study" }

2. Structured citations from a RAG pipeline

{ "claims": [{ "url": "https://example.com/", "quote": "This domain is for use in documentation examples without needing permission." }] }

3. Links only, show only the broken ones

{ "urls": ["https://example.com/a", "https://example.org/b"], "onlyProblems": true }

Run with no input to see a demo on a short answer that contains a true quote, a fake quote, a dead page and an invented domain.

How a claim is checked

1. Pairing. Each URL is matched to the sentence it appears in (markdown link, bare URL, or a numbered [n] citation) and to any quoted passage in that sentence.
2. Link. The URL is fetched. DNS failures, 404 and 410 are DEAD; a 200 page that says "not found" is SOFT_404; a deep link that lands on the home page is REDIRECTED_TO_HOMEPAGE; sites that refuse the check are BLOCKED (free, never guessed).
3. Quote. The page text is compared with the claim: exact match first, then the best one to three sentence window by overlap of content words and phrases. Numbers in the claim that are missing from the page are listed.
4. Archive. For links that are not live, the Wayback Machine is asked for a copy, and the copy is tested against the claim too.

Input guide

Field	What it does
answer	The text to check. Markdown links, bare URLs and [1] style citations are understood.
claims	Explicit {url, quote} pairs.
urls	Links to check with no claim attached.
supportThreshold	Overlap needed to call a claim supported (default 70).
useWayback	Look for archived copies of failing links (default on).
onlyProblems	Hide verified and working links (they are not charged).

Output

Field	Meaning
status	VERIFIED, QUOTE_WEAK, QUOTE_NOT_SUPPORTED, LINK_OK, LINK_OK_UNCHECKED, DEAD, SOFT_404, REDIRECTED_TO_HOMEPAGE, BLOCKED, UNREACHABLE
groundingScore, groundingLevel	0-100 overlap; EXACT, NEAR_EXACT, SUPPORTED, WEAK or NOT_SUPPORTED
claim, claimType	The text that was tested and whether it was a quote or a sentence
bestMatchOnPage	The closest passage on the page, so you can read it yourself
missingNumbers	Figures in the claim that the page never mentions
archive	Wayback snapshot URL and date, and whether the copy supports the claim
explanation	Plain-English verdict

Sample row:

{ "type": "claim", "url": "https://example.com/", "status": "QUOTE_NOT_SUPPORTED", "httpStatus": 200, "groundingScore": 9, "groundingLevel": "NOT_SUPPORTED",
  "claim": "this domain was used by NASA to calibrate the Voyager probes in 1977", "claimType": "quote",
  "bestMatchOnPage": "This domain is for use in documentation examples without needing permission.",
  "explanation": "The link works, but nothing on the page matches the claim (9% overlap). The quote or fact may be invented or taken from another source." }

Use it with

• A RAG or agent pipeline: send the model's answer and citations here before showing them; drop or flag anything that is not VERIFIED.
• Content QA: run it on AI-drafted articles to catch invented sources before publication.
• Fake Reference Checker: companion Actor for academic references.

FAQ

Is this a fact-checker? No. It checks whether the cited page contains the claim, not whether the claim is true.

Why is a correct paraphrase marked weak? Support is lexical overlap of words and phrases. A heavy paraphrase of a real page can score low. Read bestMatchOnPage and decide.

Why BLOCKED? Many news and publisher sites refuse requests from cloud IP ranges. Those links are reported honestly and are not charged.

Honest limits

• Pages that build their text with JavaScript, sit behind a login, a paywall or a bot wall cannot be read.
• PDFs are not read (the link is checked, the quote is not).
• Overlap is not entailment: a page can contain the same words and mean the opposite.
• Only the first 2 MB of each page is read.

Price

Pay per event: $0.002 per claim checked (claim-checked). An answer with 30 cited links costs $0.06. Blocked and unreachable sites are free.