When your page credits a study it never read
A secondhand citation is a working link attached to the wrong source: your page says a 2020 study in PNAS found something, and links to a blog that quoted PNAS. The link resolves, the study exists, the claim may even be true - and the attribution is false. RankEcho refuses to publish one, because a link has to lead where the words say it leads.
What is a secondhand citation?
It is the sentence that reads like scholarship and is not. The page states a finding, credits a journal or an institution, and links to something else entirely - a roundup, a blog, a summary that mentioned the original. Every visible part checks out. The claim is attributed to a source nobody involved in writing the page ever opened.
Here is one this product produced during development. A draft wrote that a 2020 study in PNAS found brain-training users improved on untrained assessments, and linked to a marketing blog. The blog had said it. PNAS may well have said it too. But the page was now telling readers it had read PNAS, and it had read a blog.
Academic guidance has a name for the underlying error and a rule against it. Brown University's library guidance on generative AI puts it plainly: do not use sources an AI tool cited without reading those sources yourself. The rule is old. What is new is the volume, and that it now happens inside pages published to win citations.
Why doesn't a working link mean the attribution is right?
Because a link answers a different question than an attribution does. The link says where you can go. The attribution says where the claim came from. A page can be correct about the first and wrong about the second, and nothing about the reading experience reveals the gap.
Consider the two failures side by side. A fabricated citation names a paper that does not exist - click it and you find nothing, and the error announces itself. A secondhand citation names a paper that does exist, links to a page that really does discuss it, and carries a claim that may be accurate. Nothing announces anything.
That is why it survives review. A reader checking whether the link works finds that it works. A reader checking whether the study exists finds that it exists. Only a reader who compares what the page claims to have read against what it actually linked will see it - and almost nobody does that, because there is no visible reason to.
How does an AI writing tool produce one?
Mechanically, and without anything going wrong. A tool that researches before writing fetches pages and extracts useful sentences from them. Some of those sentences are the page's own findings. Others are the page quoting somebody else. Both arrive as text with a URL attached.
The writer then uses a fact and attributes it to the URL it came with. If the sentence was the page's own finding, the attribution is right. If the sentence was the page quoting PNAS, the writer has a claim about PNAS and a link to a blog - and joining them produces exactly the error.
No hallucination is involved. Every word was in a real document. The tool did what it was asked. The flaw is that the extraction step lost the distinction between a source and a source's source, and nothing downstream noticed.
Why is this harder to catch than an invented citation?
Fabrication is a solved verification problem. Studies of AI-generated references have found fabrication rates ranging from roughly 18 to 69 percent depending on the model and the field, and every one of those is caught by the same check: look for the paper, and if it is not there, delete the sentence. Tedious, but decidable.
Misattribution is not decidable that way. The paper is there. Deciding whether it supports the claim means reading it - which is precisely the work the tool was used to avoid. Coverage of this problem in research settings has noted that expert reviewers signed off on misattributed claims repeatedly, because the citations did not look like hallucinations.
And the same distortion happens in the reading direction, which suggests it is structural rather than a quirk of one tool. Columbia Journalism Review's study of eight AI search products found them citing syndicated copies rather than original articles, and in the worst case one product credited the wrong source in more than half of its answers.
Does a wrong attribution actually cost you citations?
This is the part that turns an integrity problem into a commercial one. Engines increasingly cross-reference claims against other sources, and a page whose claims do not hold up under that check is a page they can quote incorrectly - or skip.
Verifiability is doing real work in citation selection. Analysis of how AI systems choose sources keeps landing on the same trait: pages presenting claims that can be checked against other reliable sources are cited more often than pages that cannot be checked. A misattributed claim fails that check in a way that looks worse than no claim at all, because the page asserted a provenance it does not have.
So the reason to fix it is not only that it is wrong. You published the page to be cited. The false attribution is a reason not to cite it.
How do you catch it before publishing?
Two checks, and neither needs a model. First, at extraction: a sentence carrying its own citation is marked as secondhand. Phrases like a study in, according to, a named year beside the word study, a DOI, or an et al are all signals that the page you read is quoting somebody else.
Second, at the gate: compare the publication named inside a link's text against the host the link points at. If the words say Nature and the URL says someone's blog, refuse the draft and say which one. This catches the error even when the extraction step missed it, and it costs nothing to run.
RankEcho does both. Secondhand facts reach the writer in a separate section with an explicit instruction - attribute this to the page we read, not to the source it quotes - and any draft whose link text credits a publication the URL does not match is rejected with the mismatch named.
What should the page say instead?
Credit what you read. The honest form of the earlier example is a sentence naming the blog and its claim: as this roundup reports, a 2020 PNAS study found X. The reader now knows exactly how far the provenance goes, and can decide whether to go further.
Or go and read the original, then cite it. That is better writing and it is more work, and the choice between them belongs to a person rather than a tool. What is not acceptable is the third option, where the page implies the first without doing the second.
The general rule is short enough to keep: the link must lead where the words say it leads. If it cannot, change the words.
What this check cannot catch
It cannot tell whether the source supports the claim. It compares the publication named in a link against the link's destination. A claim credited correctly to a page that does not actually support it passes this check and fails the reader.
It cannot catch a paraphrase that drifts. A draft written here said over forty thousand MRI scans where the source said over forty thousand participants - same figure, different claim, no rule violated. Human review exists for exactly that.
And it cannot see a claim with no link at all. That is a different check: a claim ledger listing every factual statement and whether a source sits beside it. Both are needed, and neither replaces reading.
Frequently asked questions
A hallucinated citation names a source that does not exist, so clicking it reveals the error immediately. A secondhand citation names a real source and links to a different real page that quoted it - everything resolves, and the only fault is that the page claims a provenance it does not have.
No. Citing the blog and saying so is fine and often appropriate. What is wrong is crediting the study while linking to the blog, because that tells a reader you consulted the original when you consulted a summary of it.
That check will not find this. A secondhand citation resolves perfectly - the destination exists, loads, and genuinely discusses the source named. The mismatch is between what the link text claims and where the link goes, which requires comparing the two rather than testing one.
It can. Verifiability is a repeated factor in how engines select sources, and engines increasingly cross-check claims against other pages. A claim whose stated provenance does not match its link is a claim that fails that cross-check, which makes the page a riskier source to quote.
In two places. Facts extracted from a page that are themselves quotations are separated and labelled, with an instruction to attribute them to the page we read. And any draft whose link text credits a publication the destination host does not match is refused before it can be stored, with the mismatch named so it can be corrected.