Most AI writers will hand you an article full of confident statistics and zero way to check them. Bylined doesn't ship a claim until the source has been fetched, the passage has been matched, and the number has been verified. Here's exactly how.
If any one fails, it's a marketing claim, not a receipt. Bylined enforces all four before an article ships.
When Bylined fetched the page, it returned a 200, not a 404, paywall error, or rate-limit. We log the response code and timestamp.
The quoted sentence has to actually appear at the URL — string-matched, not "the topic was mentioned." This is what catches phantom citations.
If a stat is "41.3%", the source has to say 41.3%. We re-extract numbers and cross-check — models will sometimes say 41.2% and we'd never catch it without this step.
Every source URL is snapshotted to the Wayback Machine the moment Bylined uses it. When the URL drifts six months later, you can still verify what was originally there.
An article either passes the verification gate at stage 4 or it doesn't get published. There's no "best effort" mode.
Before the writer model sees a single token, Bylined runs your topic through a live web search, fetches the pages that come back, and reads them for usable stats and quotes — so the model writes from real sources, not from memory.
Each fetched page is parsed for stats, named quotes, and dates. Every entry is stored with its origin URL, the exact surrounding sentence, and the timestamp it was retrieved.
The writer model can only cite from the facts library. Every numeric claim must attach a source ID and the exact passage it's drawing from. If no source supports a sentence, the model writes it without a number — or skips the claim entirely.
Every cited claim is re-checked against its source: URL fetched fresh, passage string-matched, number compared. If any check fails, that section gets regenerated — or the claim is stripped. This is the step that distinguishes "with receipts" from marketing-speak.
Each verified source URL gets archived to the Wayback Machine at write time. When the original page changes or 404s — and it will — your receipt still resolves to the version that supported the claim.
Some claims aren't checkable with current tooling. We mark them honestly instead of faking it.
If the source is behind a hard paywall (NYT, WSJ, academic journals), we can attribute but can't string-match the passage. These get tagged unverified-paywall in the receipts panel — visible to you, not buried.
Numbers buried in JS-rendered charts, infographics, or PDF tables aren't reliably parseable. We currently skip these as sources rather than risk citing what we can't verify.
Some sentences don't need a source — "Email marketing platforms send email to your customers" doesn't deserve a citation. Those run as plain prose, not as fake-cited claims with a token receipt.
If you bring internal numbers (CTR, GMV, conversion rates), Bylined will use them and tag them internal. The receipt links to whichever doc you uploaded, not a public URL.
Internally we track the share of generated claims that pass verification on the first pass. We refuse to ship the product if it dips below 95%. If you're going to charge for "with receipts," the receipts have to actually verify.
Early internal benchmark across a small keyword sample.
14 days free. Every article comes with receipts — on every plan.
Start free