A citation is not a quotation

A citation is a pointer, not a quotation: it tells you an AI answer linked to your page, and it tells you nothing about whether the sentence attached to that link says what your page says. Those are two different measurements, and almost every visibility tool on the market reports only the first one.

The gap between them has now been sized. A May 2026 study called "Verified Misguidance" built a dataset, CITETRACE, that follows the full chain from user query to retrieved source to generated sentence, over 11,200 real queries drawn from 28 Stack Exchange communities, 112,000 responses from ten models across five providers, and 761,495 evaluable citation pairs. Every citation in that set points at a real, reachable source. The failures the paper measures are all downstream of that.

#Three ways a live link still fails

The study scores each citation on three axes, each on a one-to-five scale where two or below counts as a failure.

Answer-source fidelity asks whether the citing sentence matches the source content. 30.6% of citations distort their source. The rubric is deliberately not binary, because the interesting failures are not fabrications. They are claims that exist in your page but were, in the authors' words, stripped of a critical qualifier, reversed in direction, or attached to tangential context.

Source suitability asks whether that type of publication is credible for that domain of question. 27.1% of citations come from domain-inappropriate sources: a blog post carrying a medical dosage question, a vendor page carrying a safety claim.

Intent-purpose alignment asks whether the source's communicative purpose matches the query's intent, and it is the one axis that mostly works, failing on 5.1% of citations.

Roll those up to the response a user actually reads and the exposure is near-universal. Across models, the share of responses containing at least one structurally misleading citation runs from 71.3% to 96.1%.

#The variance is in the engine, not in your page

This is the finding that should reorder your priorities. Provider identity accounts for 88% to 96% of citation-quality variance, with alignment and suitability both at 96% and fidelity at 88%. Which engine is reading you predicts how accurately you get represented far better than anything about the page.

And the engines do not fail in the same direction. claude-haiku posts a 12.3% fidelity failure rate against a 30.1% suitability failure rate; gpt-5 inverts it at 42.3% fidelity and 8.0% suitability. The models that pick the most credible sources quote them the least carefully, and the models that quote carefully reach for weaker sources. No model in the set sits in the good corner on both.

Citation density trades the same way. Responses carry 5.3 to 12.7 citations on average depending on the model, and as density rises fidelity failure falls from about 35% to 28% while suitability failure climbs from about 22% to 31%. More links means more careful attribution to a wider and weaker pool.

That pool is the part worth staring at. Across all models, blogs and social media supply 29.2% of citations, company pages 21.1%, official institutions 17.5%, and research 8.8%. The paper attributes the suitability failures to that structural skew in what retrieval surfaces rather than to deliberate model preference. The engine is choosing from what the index hands it.

#Nobody is checking behind the link

The obvious counterargument is that a misleading citation is self-correcting, because a reader who cares will click through. A CHI 2026 study of 1,517 references across 30 question-answer pairs and nine systems tested that directly. In its user study, hover and click frequencies stayed below 25% for every platform, with ChatGPT highest at roughly 22% hover and 12% click. Satisfaction scores stayed above 4.0 anyway. Readers judged plausibility, not sourcing.

Treat that as directional rather than settled: the user study ran with 12 participants. The reference audit behind it is the larger piece of work, and it found the same spread by system that the fidelity study did, with ChatGPT averaging 9.5 references per response and Claude 10.0 against DeepSeek's 2.1.

The practical consequence is that a distorted citation of your page mostly stands. It is not flagged, not corrected, and not visible in any log you own. Your server saw a normal fetch. Your rank tracker saw a citation. The reader saw a claim you did not make.

#What you can actually move

Start by accepting the split. Fidelity and factual absorption are separate stages in the pipeline, which is how the July 2026 critical survey of 45 GEO studies models it: activation, crawl, retrieval, reranking, citation, prominence, absorption, fidelity, user behavior. Optimizing for the citation stage does nothing for the two stages after it.

Three things are in your control.

Write claims that survive extraction. The dominant fidelity failure is a qualifier separated from its number. If your sample size, date range, or scope lives in a sentence two paragraphs above the figure, an extracted passage will carry the figure without it. Keep the qualifier and the number in the same sentence, every time, even when it reads redundantly.

Fix your domain signal, not your prose. Suitability failure is a judgment about publication type against subject matter. A vendor blog answering a regulated question is scored as a weak source regardless of how good the post is. The fix is publishing the substantive claim where its type belongs, in documentation, in a methods page, in third-party work you can point at, rather than rewriting the blog again.

Measure the sentence, not the link. Add a third column to citation tracking beside "cited" and "position": does the citing sentence state what the page states? That is a manual read today, and it is the only column that catches a misquote.

What you cannot do is control it at the source. Google is explicit that appearing in its AI features requires only that a page be indexed and eligible to show with a snippet, with no additional requirements or special optimizations. The controls it documents are all subtractive: nosnippet, data-nosnippet, max-snippet, or noindex. You can reduce what is quoted. There is no markup that makes a quotation accurate.

So the honest position is that a share of your citations will misrepresent you, the share depends mostly on which engine a reader happened to open, and the correction path runs through writing that does not fall apart when a paragraph is cut in half. Counting citations tells you that you were in the room. It does not tell you what was said about you.