The GEO optimisation conversation has been about getting cited. More citations, higher citation rate, broader platform coverage. That framing treats citation as the endpoint. A May 2026 preprint suggests citation is not the endpoint — citation fidelity is. We think this is the most important framing shift in GEO research since the field started measuring citation volume.
The paper — "Verified Misguidance: Measuring Structural Citation Failures in Search-Augmented LLMs" (arXiv:2605.28565) — introduces a construct the field has been missing: the Fidelity Failure Rate (FFR). Across 761,495 evaluated citation pairs from ten models and five commercial providers, it measured what fraction of citations distort, amplify, reverse, or strip qualifiers from the content they cite. The answer: 30.6%.
That is roughly one citation in three.
Evidence tier: Preprint — not yet peer-reviewed. All figures are from a May 2026 preprint measuring a snapshot of specific model versions. Treat as current best available evidence, not as stable platform properties. These rates will change as models update.
What "verified misguidance" means — and why it is different from hallucination
A hallucinated citation is a source that does not exist. The AI invents a paper, a URL, or an author name. Verified misguidance is the opposite failure mode. The source exists. It is retrievable. The AI retrieved it and placed it in its context window. And then it attributed a claim to that source that the source does not make — or makes differently.
The source is real. The citation is real. The representation is wrong.
For organisations producing high-stakes content — policy analysis, research, legal guidance, financial commentary — this distinction matters. You cannot verify your way to safety by checking that your sources are cited. The paper's name for this phenomenon is precise: Verified Misguidance. A reader following that citation will find a real source. They will find that it does not say what the AI said it says.
The two failure modes the paper identifies
The study separates citation quality problems into two distinct types. They operate differently and have different implications for content producers.
Fidelity failure (FFR) is the central finding. The cited source is real and domain-relevant, but the AI's generated claim distorts the source in one of several ways: amplifying a qualified finding into a definitive one, reversing a direction or qualifier, stripping the conditions under which something is true, or attributing a claim the source does not make at all. The source is appropriate — the representation of it is not.
Source suitability failure (SSF) is a separate failure mode. The cited source is real but domain-inappropriate: the AI cites a source about Topic X to support a claim about Topic Y. The source exists and was retrieved; it simply has no bearing on the claim it is attached to.
These are not the same problem. A source can fail on fidelity without being domain-inappropriate. A source can be domain-appropriate and still be misrepresented. The paper measures both — and the 30.6% FFR is the fidelity failure rate, not a combined figure.
The numbers — and the variance
The 30.6% overall FFR is a mean across the full dataset. The model-level variance is wide — and for our purposes, that variance matters as much as the average.
At the high end: gpt-5 showed an FFR of 42.3% in this dataset. At the low end: claude-haiku showed 12.3%. That is a 30-point spread between the best and worst performers measured here — across a study that covered ten models and five commercial providers.
These figures are from a preprint measuring a snapshot of specific model versions in commercial search-augmented deployments. GPT-5, claude-haiku, and the other models tested here will update. Their fidelity profiles will change. The 42.3% and 12.3% figures are the best available evidence as of this paper — not permanent platform properties.
The practical implication of the variance is direct: which AI platform a reader is using when they encounter a citation to your content is a 30-point-spread factor in whether that citation represents your content accurately. A reader using a platform running a high-FFR model has a materially different chance of being accurately informed than a reader using a low-FFR model.
The provider finding — and what it means for content creators
The paper's most structurally important finding for practitioners is buried in the variance analysis. When the researchers decomposed what drives fidelity failure rates, 88–96.5% of the variance was explained by provider-level effects — that is, by which retrieval backend the platform uses. Not by model capability in isolation. Not by properties of the source content.
By the retrieval backend.
This finding has a direct and uncomfortable implication: content creators cannot directly control their fidelity failure rate. The dominant driver of whether your content is cited accurately is architectural — it lives in the retrieval system the platform has built, not in how you have written or structured your content. Optimisation of your content cannot fix a fidelity failure that originates in the retrieval layer.
This does not make the finding useless. Awareness matters. Understanding that verified misguidance is a real, measurable phenomenon — that an AI citing your content is not the same as an AI representing your content — changes what questions organisations should ask about their AI visibility. Our view: the finding has value as a frame even when the lever is not in the content producer's hands. Knowing the problem exists is the prerequisite for asking the right questions about it.
What this study does not tell us
The paper is a May 2026 preprint. It has not completed peer review. The methodology and conclusions have not been independently validated.
The dataset covers ten models and five commercial providers in search-augmented deployments — not every AI platform, not every deployment context. The FFR figures apply to the specific model versions and retrieval configurations measured. As those change, the rates will change.
The 88–96.5% provider-level variance finding is the most important constraint on how this research should be applied. It means that even with full knowledge of your content's fidelity risk profile, the dominant variable is outside your control. The paper establishes that the problem is real and measurable; it does not supply a content-level solution to it.
A separate preprint (H6, arXiv:2605.06635) finds that factual accuracy reaches only 39–77% in deep research agents — a related but distinct finding about what happens when AI tools run many sequential searches. These two findings are independent lenses on citation unreliability. The FFR measured in H5 and the factual accuracy range in H6 are not the same rate and should not be combined or compared as though they are.
The Psytable tools and this finding
We do not currently have a fidelity audit tool. That is an honest answer and the correct place to start.
The Platform Variance tool surfaces citation consistency across platforms — whether a source is reliably cited by ChatGPT, Perplexity, and Google simultaneously. That is an adjacent question: consistency of citation selection, not fidelity of citation representation. A source cited consistently across platforms could still be misrepresented consistently.
The Absorption Analyser scores content against the structural and linguistic properties associated with deeper absorption — whether cited content contributes to the shape of the generated answer. That is also an adjacent question: absorption depth, not accuracy of attribution.
Both tools address parts of the citation quality picture. Neither addresses fidelity failure directly. A tool that measures whether a generated claim accurately represents the source it cites would require a different architecture — one that compares output claims against source content at the sentence level. That gap exists in the current tool landscape, not just in psytable's suite. The Verified Misguidance paper is, among other things, the first published measurement framework for that problem at commercial scale.
References
"Verified Misguidance: Measuring Structural Citation Failures in Search-Augmented LLMs" (2026). Preprint. arXiv:2605.28565
"Cited but Not Verified: Parsing and Evaluating Source Attribution in LLM Deep Research Agents" (2026). Preprint. arXiv:2605.06635
Check citation consistency across platforms.
The Platform Variance tool shows whether your content is cited by ChatGPT, Perplexity, and Google AI Overviews — or only by some of them. It does not measure fidelity, but cross-platform citation stability is the adjacent question this finding makes worth asking.