Use caseFor AI content platforms
Check every article and report your platform publishes.
When your platform generates articles, market reports, or regulatory explainers for paying customers, one fabricated statistic or broken reference erodes trust in the whole pipeline. Lenz checks each factual statement against primary and authoritative sources or your own reference documents, and checks that cited references exist and support what’s written—returning structured results your code can gate on.
Your pipeline’s draft“Google rolled out AI Overviews to search users in the United States in May 2024. That February, Gartner predicted that traditional search engine volume would drop 50% by 2026 as users shift to AI chatbots and virtual agents. OpenAI said in October 2025 that ChatGPT had 800 million weekly active users.”
Two of the three statements hold. One forecast is wrong.
FalseGartner predicted a 25% decline in traditional search engine volume by 2026.
A generation step conflates two separate Gartner forecasts. Lenz checks each statement against primary sources and returns suggested_rewrite and sources so your pipeline can revise or hold the draft before delivery.
01Example
Three statements in a generated article. One forecast is wrong.
| Claim | Verdict | Lenz Score | Sources |
|---|---|---|---|
| OpenAI said in October 2025 that ChatGPT had 800 million weekly active users. | Truehigh confidence | 9/10 | antaranews.comcdn.openai.comtechcrunch.com+21 more |
| Google rolled out AI Overviews to search users in the United States in May 2024. | Truehigh confidence | 10/10 | blog.googlebusiness.google.combloomberg.com+15 more |
| In February 2024, Gartner predicted that traditional search engine volume would drop 50% by 2026 as users shift to AI chatbots and virtual agents.Suggested rewriteIn February 2024, Gartner predicted that traditional search engine volume would drop 25% by 2026 as users shift to AI chatbots and virtual agents. | Falsehigh confidence | 2/10 | gartner.cominc.comprnewswire.com+19 more |
Full verification · Checked 15 November 2026.
Open a row for the sources it used and the reasoning that weighed them.
A full verification returns the fields your code branches on — verdict, confidence, lenz_score, key_finding, suggested_rewrite, and sources — so the release decision stays in your pipeline: publish, revise, or route to an editor.
02The API
One call between generation and delivery.
An LLM and RAG hallucination detection and fact verification API for AI-generated content platforms before you publish reports and citations.
- 01Review a complete draft in one call
POST /api/v1/review(or/api/v1/review/filefor.docxand.pdf) · passescalate.max_citationsandescalate.suggest_edits. Lenz extracts the claims, runs assessments, escalates disputed statements to full verification, and checks the citations. - 02Or compose the four primitives
POST /extract(0 credits, 1,000 calls/day), triage claims withPOST /assess(~15 seconds per call, 1 credit), check references withPOST /citecheck, and escalate adverse or low-confidence claims toPOST /verify(~90 seconds at standard depth for 10 credits, or ~60 seconds atdepth: "low"for 5 credits). - 03Gate on outcome, issues, and suggested editsCheck
outcome == "clean"(or"issues_found","incomplete","unchecked"). Route flagged claims inissues[]and broken references incitation_issues[]to your revision step withclaims[].suggested_edits. - 04Connect via SDK, webhook, n8n, Zapier, or MCPOfficial Python and TypeScript SDKs, HMAC-signed async webhooks, n8n and Zapier workflow steps, or the MCP server.
# pip install lenz-io from lenz_io import Lenz client = Lenz(api_key="lenz_...") review = client.review_and_wait( text=generated_draft, max_verifications=5, max_citations=10, suggest_edits=True, ) # outcome is "clean", "issues_found", "incomplete", or "unchecked" if review.outcome != "clean": for item in review.issues: print("CLAIM:", item.claim, item.verdict) for cite in review.citation_issues: print("CITATION:", cite.finding, cite.cited_url or cite.doi)
// npm install lenz-io import { Lenz } from "lenz-io"; const client = new Lenz({ apiKey: "lenz_..." }); const review = await client.reviewAndWait({ text: generatedDraft, maxVerifications: 5, maxCitations: 10, suggestEdits: true, }); // outcome is "clean", "issues_found", "incomplete", or "unchecked" if (review.outcome !== "clean") { console.log("HOLD:", review.outcome, review.issues, review.citation_issues); }
curl -X POST https://lenz.io/api/v1/review \ -H "Authorization: Bearer $LENZ_API_KEY" \ -H "Lenz-Version: 2026-10-11" \ -H "Content-Type: application/json" \ -d '{ "text": "Google rolled out AI Overviews in May 2024. ...", "escalate": {"max_verifications": 5, "max_citations": 10, "suggest_edits": true} }' # poll GET /api/v1/reviews/{review_id} or pass webhook_url # gate on outcome != "clean", then inspect issues[] and citation_issues[]
Python SDK: pip install lenz-io. TypeScript SDK: npm install lenz-io. Also available as an n8n node, Zapier app, and MCP server. Full endpoint reference on the developer page.
03Citations and your sources
Check the references, and check against your customer’s source pack.
Citation checking
POST /citecheck resolves DOIs (Crossref, OpenAlex, PubMed, Semantic Scholar), statutory and case citations (EUR-Lex, US Code, eCFR, UK, German law, CourtListener), and public URLs, then checks whether the source passage supports the sentence.
Customer reference documents
On API version 2026-10-11, pass supplied_sources (up to 50 reference documents, 100,000 characters each, 500,000 characters total) on /review, /review/file, /verify, /verify/batch, or /citecheck so generated content is checked against your customer’s brief, dataset, or filing alongside the public record. Checks using supplied_sources stay private, bypass cache, and do not qualify for the warranty.
A guardrail filters. Lenz investigates.
In-process guardrails
Compare generated output against the prompt context at token time. Built for sub-second chat filtering, not checking whether an external statistic, date, or cited study is true.
LenzPre-publish verification
Searches independent primary sources and registries, runs an adversarial debate across multiple model vendors, and has three independent reviewers score the evidence.
Returns the verdict, score, sources, reasoning, and a suggested rewrite your pipeline can apply or route to an editor.
The open evaluation: where frontier models disagree on real-world fact-checks
| Field | Meaning | /assess | /verify |
|---|---|---|---|
| verdict | True, Mostly True, Mixed, Mostly False, False, or “No Public Record” on /verify | Yes | Yes |
| confidence | Agreement and certainty across the reviewer panel (high, medium, low) | Yes | Yes |
| lenz_score | Integer score from 1 to 10; null for “No Public Record” | No | Yes |
| key_finding | The decisive finding in one sentence | No | Yes |
| suggested_rewrite | Corrected sentence when a factual detail is wrong, guarded for token preservation | No | Yes |
| sources | URL, title, snippet, and publication date for each cited source | No | Yes |
| coverage | Whether the verification qualified for the warranty and its signed certificate | No | Yes |
Warranty
Vendors disclaim accuracy. Lenz warrants the result.
A verdict can qualify when a full verification returns True or False with high panel confidence and enough independent sources behind it. Your code reads that from the coverage block.
On a paid Pro or Scale plan, every full verification that qualifies carries a contractual warranty, at no extra charge.
05Pricing
Try free, no card.
$7.99/ month
500 assessments or 50 full verifications a month, in any combination.
Or $6.59 a month, billed $79 a year
Choose Plus$99/ month
5,000 assessments or 500 full verifications a month, in any combination. The warranty on qualifying verdicts.
Or $83 a month, billed $990 a year
Choose Pro$399/ month
20,000 assessments or 2,000 full verifications a month, for production volumes. The warranty on qualifying verdicts.
Or $333 a month, billed $3,990 a year
Choose ScaleEvery plan includes 1,000 extractions a day, an extraction being one call however long the text.
Compare plans06Questions
Start with /review when you have a complete generated draft and want claim triage, full verification, citation checks, and claims[].suggested_edits in one asynchronous job. Check outcome == “clean” to gate delivery, and inspect issues[] and citation_issues[] when a draft needs attention. Use /extract, /assess, /verify, and /citecheck separately when your pipeline needs custom per-claim routing thresholds.
Yes. On API version 2026-10-11, pass supplied_sources (up to 50 documents, 100,000 characters per document, 500,000 characters total) on /review, /review/file, /verify, /verify/batch, or /citecheck. Lenz checks statements and matched citations against your supplied text as well as public sources and registries. Supplied-source checks are always private, never served from cache, and excluded from the warranty.
/assess is synchronous and takes ~15 seconds per call. /verify takes ~90 seconds at standard depth (~60 seconds at depth: “low”) and supports async webhooks. /review runs asynchronously over a complete draft with polling or webhook delivery. Lenz is built for pre-publish content and report pipelines, not sub-second chat token streaming.
A RAG groundedness check measures whether an answer stayed close to a retrieved prompt chunk, and an AI content detector estimates whether prose was written by a model. Lenz checks whether the factual claims and citations in the text are supported by independent sources and registries.
When no public record settles a statement, Lenz answers “No Public Record” instead of a score. That doesn’t make the statement false.
Yes. The Lenz n8n node, Zapier app, and MCP server call the same endpoints and return the same structured fields.
Not answered here? Talk to us