Citation & Quote Verifier for AI Output

Check a citation before you ship it. Give this Actor a claim, a URL, and the quote the model used. It fetches the page, tells you whether that quote is actually there, and records the HTTP trail.

It is for editors, researchers, and anyone reviewing AI output who needs a flat, webhook-friendly row per citation: reachable or not, exact or fuzzy match, paywall, PDF, and the closest Wayback Machine snapshot.

Unofficial, and not affiliated with the Internet Archive, the Wikimedia Foundation, the National Archives, the W3C, OpenAI, or Anthropic. A match is evidence from the fetched text. It is not legal advice and it is not a ruling that the source is true.

What you put in

Leave the input empty, or keep the prefilled list, to check eight public citations (real quotes, wrong quotes, a page with no quote, a dead link, and a one-page PDF). That run is meant to finish in under two minutes.

Or send your own list:

{
  "items": [
    {
      "claim": "The preamble says the people ordained the Constitution to form a more perfect Union.",
      "url": "https://www.archives.gov/founding-docs/constitution-transcript",
      "quote": "We the People of the United States, in Order to form a more perfect Union"
    }
  ]
}

inlineText accepts the same citations as JSON (an array, or an object with items), as CSV with claim, url, and quote columns, or as Markdown. A Markdown link becomes a citation. A blockquote next to that link becomes the quote.

url and quote are optional. No URL is recorded as unsupported and is not billed. No quote, on a page that loads, is a reachability check (no_quote_given).

An explicit empty items array does not fall back to the demo. An omitted items field does.

What you get

One dataset row per citation. Offsets refer to the cleaned page text (tags removed, whitespace collapsed, spaces HTML inserts before punctuation removed). The full page is not stored. evidenceSnippet is a short window around the match, and evidenceHash is its SHA-256.

{
  "resultId": "a1b2c3d4e5f678901234",
  "inputIndex": 0,
  "inputSource": "items",
  "claim": "The preamble says the people ordained the Constitution to form a more perfect Union.",
  "url": "https://www.archives.gov/founding-docs/constitution-transcript",
  "quote": "We the People of the United States, in Order to form a more perfect Union",
  "status": "verified",
  "coverage": "complete",
  "confidence": 0.95,
  "reason": "Exact quote found in the extracted page text.",
  "reachable": true,
  "httpStatus": 200,
  "finalUrl": "https://www.archives.gov/founding-docs/constitution-transcript",
  "redirectChain": "https://www.archives.gov/founding-docs/constitution-transcript (200)",
  "redirectCount": 0,
  "pageTitle": "The Constitution of the United States: A Transcription | National Archives",
  "publishedDate": null,
  "paywall": false,
  "loginWall": false,
  "isPdf": false,
  "quoteExact": true,
  "quoteFuzzy": true,
  "quoteMatchKind": "exact",
  "matchStart": 120,
  "matchEnd": 195,
  "matchCount": 1,
  "evidenceSnippet": "We the People of the United States, in Order to form a more perfect Union",
  "waybackAvailable": true,
  "waybackUrl": "https://web.archive.org/web/20261019111022/https://www.archives.gov/founding-docs/constitution-transcript",
  "llmVerdict": null,
  "llmProvider": "none",
  "charged": true,
  "chargeEvent": "claim-checked",
  "observedAt": "2026-10-23T13:00:00Z",
  "fillRate": 1,
  "schemaVersion": "1"
}

status is one of:

Status	Meaning
verified	The quote is on the page, exactly or after normalising case, whitespace, and quotation marks.
quote_not_found	The page loaded and the quote is not in the extracted text.
unreachable	Non-success HTTP status, soft 404, or the host did not answer.
paywalled	HTTP 401/402, or a paywall/login wall, and the quote was not in the visible text.
no_quote_given	The page loaded and there was no quote to search for.
blocked	Challenge, captcha, or HTTP 429. The quote is not searched in that interstitial.
unsupported	No URL, or the URL is not allowed (private, local, file:). Not billed.
no_results	Nothing billable was delivered. Not billed.

The key-value record OUTPUT is the run summary: counts by status, rowsCharged, rowsDropped, fillRate, chargedUnits, and error counters for proxy, rate-limit, blocked, network, and other.

A named store, citation-claim-verifier-baseline, remembers the last good observation per URL and claim. A timeout, proxy error, or other ambiguous failure does not overwrite it. Turn on alertOnlyOnChange to push and bill only rows whose status, match kind, or page-text hash changed.

Pricing

This Actor uses pay-per-event pricing. Apify also charges a $0.005 start event once per run. There is no per-dataset-item charge.

Event	When it is billed	Price
claim-checked	A citation row passed validation and the fill-rate check	$0.01
llm-verdict	A supports / neutral / contradicts verdict was parsed from your model	$0.01 extra

You're only charged for rows that pass validation. Unsupported URLs, empty runs, ambiguous failures, and rows held back by the charge cap are not billed. The same idempotency key is not billed twice inside a run. A new run checks the pages again and can bill again.

maxChargeUsd stops new billing once the estimated total, using the prices above, reaches the cap. The platform's own max total charge still wins when it is set.

Run	Billable units	Total, including the $0.005 start
Demo, 8 checks, no model	8 × claim-checked	about $0.085
100 checks, no model	100 × claim-checked	about $1.005
100 checks, each with a model verdict	100 × claim-checked + 100 × llm-verdict	about $2.005

1,000 results ≈ $10, plus the start event. 1,000 results that each also receive an LLM verdict ≈ $20, plus the start event.

Scope and limits

• HTTP first. Redirects are followed manually (up to maxRedirects) and each hop is refused if it points at a private, loopback, or link-local address. localhost, .local, metadata hostnames, embedded credentials, and file: URLs are refused.
• Bodies are capped at 5 MB. PDFs use the first 30 pages. Quotes shorter than 8 characters (12 after normalisation) are not treated as matches.
• Fuzzy matching folds case, whitespace, and curly quotes or dashes. It does not do paraphrase or edit-distance matching. The quote needs to be the passage, not a summary of it.
• JavaScript-only pages can come back with almost no text (quote_not_found, coverage empty). browserFallback is off by default. When on, the Actor retries once with a browser-like User-Agent, then imports Playwright only if that package is installed. The published image does not include a browser, and memory stays at 512 MB (1024 MB maximum) so a real browser is not the supported path.
• Paywalls are detected, not bypassed. Captchas are not solved. A datacenter IP can be blocked; that is a blocked row.
• Wayback lookup calls https://archive.org/wayback/available. It is best-effort. A lookup error lowers coverage from complete to partial and does not fail the check.
• Optional model verdicts use your key against OpenAI (gpt-4o-mini) or Anthropic (claude-3-5-haiku-latest), temperature 0. The model sees the claim and a short excerpt, and it is told the excerpt is untrusted data. No key, or provider none, keeps the deterministic quote match and does not call a model. A failed model call still delivers the quote-match row and does not bill llm-verdict.
• At most 200 citations per run. Polite User-Agent, per-host delay, and retries for timeouts, 429, 502, and 503.
• Personal data is not a target. Do not point this at sites whose terms forbid automated access.

Integrations

The dataset is flat and CSV-friendly. Point a webhook at the run, or pull {{links.apiDefaultDatasetUrl}}/items and the OUTPUT record.

• Make / n8n: HTTP module, “run an Actor”, then iterate the dataset items. Map status, url, quoteMatchKind, and evidenceSnippet into a sheet or a review queue. Use alertOnlyOnChange on a schedule if you only want rows that moved.
• Google Sheets: Apify’s Google Sheets integration, or a webhook that appends claim, status, finalUrl, httpStatus, quoteMatchKind, waybackUrl, llmVerdict, and charged.
• Webhook: The default dataset push is one JSON object per citation. charged: false with a dropReason means that row was stored and not billed.

Run locally

pip install -r requirements.txt
python -m src

With no input in the local key-value store, the eight-citation demo runs. Tests:

pip install pytest pytest-asyncio jsonschema referencing
pytest -m "not smoke"
pytest -m smoke

FAQ

Why is a wrong quote still billed? The page was fetched and the absence is the result. You are paying for the check, not only for hits.

Why was a row free? It failed validation, the URL is not allowed, the run had nothing to check, the charge cap was already hit, or alert-only mode suppressed an unchanged citation.

Does the model see the whole page? No. It sees the claim and a short excerpt. Quote matching does not need a model.

Will it match a paraphrase? No. Exact text, or the same text with different case, whitespace, or quote characters.

Can I raise memory and turn on a browser? You can set browserFallback to true, but this image has no Playwright browser. Install one only in a private build, and expect to need more than the 1024 MB cap before that path is reliable.

Contact

Questions and bug reports: plusonecares@gmail.com. Apify username: plusonets.