Only 46% of cited sentences are supported by sources

A spot-check of 37 AI-written article sentences with citations found that less than half were clearly supported by a close sentence on the cited page[1]. Over half either did not match or were unreachable, highlighting a gap between citation and direct support[1]. This article breaks down what was checked, what “not supported” really means, and what this kind of checking can—and cannot—tell you about the reliability of sourced answers.

How the spot-check was done: a page-by-page test

The check was run automatically over a one-week span[1]. Each day, it picked a published article with citations, took up to three cited sentences, and opened the cited page for each[1]. The test then looked for a sentence on the page that matched the article’s sentence closely[1]. If a close match was found, it recorded “SUPPORTED”; if not, “NOT SUPPORTED”[1]. If the page could not load, the verdict was “SOURCE UNREACHABLE”[1].

Out of 37 sentences from 8 articles, the results were:

[1]

Notably, only one article had every checked sentence supported; five were mixed, and two had none supported[1].

What “not supported” really covers: paraphrase, mismatch, and misses

A “NOT SUPPORTED” result does not always mean the sentence was wrong or misleading[1]. Of the 19 cases:

• 7 were “no sentence close to this on the page”—often paraphrases or summaries the checker could not match[1].
• 9 were cases where the page’s closest sentence actually said something different[1].
• 3 were counted as not supported because the page was unreachable, even if the citation might have been correct[1].

So, the 51% “not supported” figure is an upper bound: many of these are likely paraphrases or summaries that a strict matcher will flag, rather than true errors[1]. Only a smaller fraction are likely clear mismatches or unsupported claims.

What this kind of check cannot decide

The check only asks whether there is a close sentence on the page, not whether the page as a whole supports the article’s claim[1]. It cannot tell if a summary or paraphrase is faithful to the page, nor can it judge context or intent[1]. For example, if a sentence accurately summarizes a page’s argument but uses different words, the checker will likely flag it as “not supported”[1].

The check also cannot distinguish between a citation that is genuinely unsupported and one that simply paraphrases too loosely to be caught by a sentence matcher[1]. In some cases, the same sentence cited to different sources was found “supported” by one and “not supported” by another, suggesting verdicts may not be stable for every sentence[1].

A worked example: how a supported and an unsupported citation look

Suppose an AI-written article says: “The scaled content abuse policy targets large amounts of unoriginal, low-value content created primarily to manipulate rankings.” The cited page contains the exact sentence. The check records this as “SUPPORTED”[1].

Now, suppose the article says: “Google’s guidance on AI-generated content focuses on content quality and purpose rather than banning a particular production tool.” The cited page discusses content quality but does not have a sentence close to this one; the check records “NOT SUPPORTED”[1]. However, the same sentence, cited to a different Google page, was found supported on another day[1].

This highlights two things:

1. Exact matches are reliably caught.
2. Paraphrases or generalizations, even if accurate, may be flagged as not supported depending on the page and the checker’s strictness[1].

What this means for answer engines and citation reliability

The check was small (37 sentences, one writer, one automatic matcher, no hand-audit) and did not cover answer engines directly[1]. But it suggests that direct, sentence-level support in citations is not guaranteed—even when citations are present and seem plausible[1]. According to Wikipedia policies, citations should directly support the material as presented, not just be generally related[2, 3, 4]. In practice, paraphrase and summary make strict matching difficult, and unreachable pages create a third category of citation failure that is no fault of the writer[1].

So, a citation in an answer or article is only as reliable as the check between the sentence and the source page[1]. A strict check will flag many paraphrases; a looser check may let mismatches through. For practitioners who depend on answer engines or automated citations, this means that a citation marker is not a guarantee of direct support—especially if you need to quote or rely on a specific claim.

Questions readers ask

How big a problem is 46% support?

A 46% support rate means that less than half of cited sentences were directly matched by a close sentence on the cited page in this sample[1]. This is a warning to check citations carefully if your work depends on precise sourcing.

Can a paraphrased summary still be reliable?

Yes, a paraphrased summary can be accurate and reliable, even if a strict checker cannot match it to a specific sentence[2, 3, 4]. But for verifiability, best practice is still to cite sources that directly support the claim, ideally with a clear, attributable sentence or section[2, 3].

Does this mean answer engines are unreliable?

No, but it does mean that answer engines relying only on citation markers may overstate how well their answers are supported by their sources. The check here did not measure answer engine accuracy directly, but it suggests reason for caution when treating citations as proof of support[1].

How the numbers work

These figures were computed for this article, not estimated by its writer. Inputs come from the numbered sources or are assumptions, labelled as such; change an assumption and the results change with it.

Sources

1. Our own data, 5 Nov 2026: Citation spot-check: 37 sentences against their cited pages
2. Wikipedia:Verifiability - Wikipedia
3. Wikipedia:Citing sources - Wikipedia
4. Wikipedia:Reliable sources - Wikipedia