Most relevant pages that an AI engine leaves uncited do not fail because they are badly written. They fail because they do not answer the exact question asked, with the exact names and facts it needs. That is the main finding of a study from Virginia Tech and Zhejiang University, which took 949 cases where an AI engine had a relevant page in hand, cited a competitor instead, and asked why.
One of their examples is a question about three folklore sports in the Indian state of Karnataka. The uncited page was about sport in Karnataka. It covered cricket, football, badminton, kabaddi and the state's best-known athletes. The one sentence that named a folklore sport, Kambala buffalo racing, sat buried deep in the page. The page was on topic. It did not answer.
The study's answer to this kind of failure is not to rewrite the page. It is to find which step failed, then change only what that step needs. In its tests, that approach raised citation rates further than generic GEO rewriting rules, while touching far less of the page.
What did the study test?
The study, arXiv:2603.09296, asks why a relevant web page fails to be cited by an AI engine, and tests a system that diagnoses each failure before repairing it. Its authors are Zhihua Tian, Yuhan Chen and Ruoxi Jia of Virginia Tech, with Yao Tang and Jian Liu of Zhejiang University. It was posted on 10 March 2026 as a preprint, which means it has not been through peer review.
The authors start from a choice of target. Earlier GEO research measured contribution: how much of an AI answer can be traced back to a page. This study measures citation, the yes-or-no event of being listed as a source, because a page that is not cited sends no visitor back. In their baseline runs, 43% of topically relevant pages received no citation at all.
The work has three parts. A taxonomy of citation failures, built from 949 pairs of an uncited page and a cited competitor retrieved for the same question. A system called AgentGEO, which compares the uncited page with the winner, names the likely failure, applies one of nine targeted repair tools, tests again and repeats. And a benchmark called MIMIQ: 204 real web pages, each with 60 generated questions, 20 used to optimise the page and 40 held back to test it. The held-back questions check that an improvement works on questions the system never saw, not just on the one it was tuned for.
One point matters before any number. The AI engine in these tests is a simulation the authors built, using gpt-4.1-mini and claude-haiku-4-5 with explicit citation instructions. It is not ChatGPT, Perplexity or Google. The authors chose it so they could diagnose and repeat experiments, which commercial systems do not allow.
Why do relevant pages fail to be cited?
In six cases out of ten, because the page does not match what the question requires: in the study's 949 failures, 62.2% were semantic alignment problems, 27.1% content quality problems, 10.1% technical problems and 0.6% structural exclusion. The authors place each failure at the step of the AI engine's pipeline where it happens: fetching the page, parsing it, or generating the answer.
| Failure family | Share | Main causes |
|---|---|---|
| Semantic alignment | 62.2% | Missing entities or terms the question needs (32.3%); page written for a different intent (29.1%); outdated or wrong-region content (0.5%) |
| Content quality | 27.1% | Too thin to cite (26.1%); fragmented, padded or unstructured text (1.0%) |
| Technical integrity | 10.1% | Content buried in boilerplate (5.2%); access blocked (3.9%); unparseable text (0.6%); content that needs JavaScript (0.4%) |
| Systemic exclusion | 0.6% | A higher-authority source covers the same facts (0.5%); answer buried beyond the context window (0.1%) |
The two largest causes are close cousins. A contextual gap means the page is on topic but lacks the specific names, figures or terms the question calls for, as with the Karnataka page. Intent divergence means the page answers a different question: a page about the 1974 film of The Turn of the Screw retrieved for a question about when the novella's story takes place. The third large cause, information scarcity, covers pages too short or vague to quote, such as a test-question index that never states the answer.
A citation failure is the case where an AI engine retrieves a relevant page but cites another source in its answer. arXiv:2603.09296 classifies these failures by the pipeline step where they occur, fetching, parsing or generation, and by cause.
Does diagnosing a page beat generic GEO rules?
In the authors' simulation, yes: diagnosed, targeted edits raised the citation rate from 56.6% to 79.5% on the GPT engine, against 68.8% for the best rule-based method. That is the "over 40% relative improvement" in the paper's summary. The gain was smaller elsewhere: from 42.4% to 54.8% on the Claude engine, and from 60.2% to 70.0% when the engine selects passages before writing its answer.
The comparison set includes the eight rewriting methods of the original 2024 GEO paper, such as adding statistics, adding quotations, citing sources or adopting an authoritative tone, and AutoGEO, a 2025 method that learns rewriting rules from many pages. The table shows a selection.
| Method | Citation rate, answer with inline citations | Citation rate, passages selected first | Text kept |
|---|---|---|---|
| Original page | 56.6% | 60.2% | — |
| Add technical terms | 55.8% | 57.3% | 21% |
| Add statistics | 60.5% | 59.5% | 69% |
| Keyword stuffing | 62.4% | 65.6% | 30% |
| Easy-to-understand rewrite | 71.1% | 65.3% | 30% |
| AutoGEO (learned rules) | 68.8% | 66.0% | 18% |
| AgentGEO (diagnosis) | 79.5% | 70.0% | 82% |
The last column is the second half of the result. The authors report that their targeted repairs changed about 5% of a page's content on average, against 25% for the baselines. AutoGEO cut pages by 43% of their length on average, and the easy-to-understand rewrite by 53%. AgentGEO kept the page and edited the parts that failed. In plain terms, most uncited pages needed a precise fix, not a new text.
The repairs themselves are ordinary editorial moves. Insert the missing entity into the sentence where it belongs. Turn a paragraph of figures into a table. Add headings and lists to a wall of text. Rewrite the opening to answer the question directly. Put a one- or two-sentence summary at the top. Move a buried answer up. Separate navigation and footers from the main content.
Can generic GEO rules hurt a page?
Yes, in some topics: the study found categories where rule-based rewriting lowered the citation rate, especially where pages were already well cited. AutoGEO did worse than the untouched pages in several topics. The authors' reading is that rules learned from averages fit average pages, and that specialised or under-represented topics depart from those averages.
Diagnosis is not immune either. On health pages, which were already cited about 80% of the time, AgentGEO itself lost ground. Some of its edits removed domain-specific information from the remaining uncited pages, which made them less likely to be cited. For a specialist site, that is the useful warning: an optimisation that simplifies away technical precision can cost the citation it was meant to win.
One baseline deserves a second look. Keyword stuffing raised the citation rate in this simulation, from 56.6% to 62.4%. That says more about the limits of a simulated engine than about what to do on a real site, where search engines have long penalised the practice. It is a reason to read every GEO benchmark as a test bench, not as a playbook.
Which citation failures can content not fix?
Those where the engine prefers a better-known domain: after optimisation, 163 questions on the 50-page test set were still not cited, and the authors found cases that no edit could win. On the training questions, the citation rate rose from 57.0% to 83.7%. The rest resisted.
Their example is a university course page on machine learning, retrieved for "best online machine learning courses". The repairs worked as intended: clearer descriptions, more detail, better alignment with the question. The engine still cited Coursera and edX. The authors conclude that engines may favour domain-level factors that sit outside the page, and that content optimisation alone has a ceiling. In the taxonomy, this is competitive redundancy: a higher-authority source already covers the same facts.
That ceiling is where a different kind of work begins. A page cannot edit its way into authority. Mentions by sources the engine already trusts, and a reputation built outside the site, are what move a brand past it.
How strong is the evidence?
The diagnostic framework is useful and the gains are consistent across settings, but every result comes from a simulated engine, and the main figure holds for one engine configuration. The authors list the commercial-engine test as future work.
The limits are worth stating plainly. The engine is a controlled simulation built on two small models, not a production answer engine. The 60 questions per page were generated by a language model, and diagnoses were made by a language model. The taxonomy shares come from GEO-Bench, a corpus of mostly informational questions across ten domains; a set of commercial questions about local businesses could shift them. The "over 40%" figure applies to the GPT-based engine with inline citations; the other settings gained 16% to 29%. And the paper is a first preprint version.
Two findings stand on firmer ground because they hold across settings. Targeted edits kept 82% of a page's wording and still beat every rewriting method. And on held-out questions with user profiles never seen during optimisation, the diagnosed pages kept their lead, at 70.6% against 64.2% for AutoGEO.
Supported: in a simulated AI engine, most relevant pages that go uncited fail on alignment with the question or on thin content, and targeted repairs chosen by diagnosis raise citation more than generic rewriting while changing far less text. Not established: that the same shares and gains hold on ChatGPT, Perplexity or Google. Not recommended by the data: keyword stuffing, or any rule applied to every page regardless of why it fails.
How can a business find out why its page is not cited?
By comparing it, question by question, with the page the AI engine cited instead. What follows is the KAIS reading of the study, turned into a method a business can apply. The study automates the comparison; the questions below can be asked by a person, on real engines.
Start from the questions buyers ask, not from the page. For each question where the page is relevant but absent, note which source the engine cited. Then put the two pages side by side and ask four questions, in order.
First, could the engine read the page at all? Content behind a bot block, a login wall or JavaScript never reaches the answer. That family was 10% of the study's failures, but on business sites it can weigh more: a study published earlier this month found AI agents blocked on a third of journeys to sites that are hard for agents to read, a gap discussed in The AI agent reads your site before it recommends you.
Second, does the page answer the question that was asked, for the intent behind it? A pricing question needs a price. A question about a city needs that city's details. Third, does the page contain the specific names, figures and terms the question uses, written where an engine will find them? Fourth, is there enough substance to quote, presented so the answer can be lifted cleanly: a direct first sentence, a table for figures, one subject per section, as covered in Page structure helps AI find your content.
If the answer to all four is yes and the engine still cites a better-known domain, the problem is no longer on the page. That is the signal to work on authority and corroboration elsewhere.
Four checks for an uncited page, in order
- Access: can an engine fetch and read the content without JavaScript, logins or bot blocks?
- Intent: does the page answer the question asked, not a neighbouring one?
- Entities: are the names, figures and terms the question uses on the page?
- Substance: is there a quotable answer, stated directly and structured to be lifted?
Where does this sit in the KAIS AI Visibility Chain?
The study sits between Be Understood and Be Cited: it explains why a page that was retrieved still does not become a source.
Be Discovered → Be Retrieved → Be Understood → Be Trusted → Be Cited → Influence the Answer → Be Recommended.
Its taxonomy maps onto the chain almost link by link. Technical failures break Be Retrieved. Semantic and quality failures break Be Understood. Competitive redundancy, the failure no edit could fix, is a Be Trusted problem. And being cited is still not the end: a source can be listed without shaping the answer, a gap measured in Cited is not used.
An uncited page is not a page to rewrite. It is a page to diagnose.
The study's most practical lesson is a change of question. Not "what makes content AI-friendly?", but "why did the engine pick that page over mine, for this question?" The answer is usually specific, often small, and different from one page to the next.
The Karnataka page did not need a new text. It needed the names of three folklore sports, written where the question would find them.
Sources
- Tian, Chen, Tang, Liu and Jia — Diagnosing and Repairing Citation Failures in Generative Engine Optimization, arXiv:2603.09296v1 [cs.IR], 10 March 2026, Virginia Tech and Zhejiang University. Preprint, not peer-reviewed. Citation-first objective and 43% baseline non-citation (section 1), taxonomy (section 3.2, Figure 2, Appendix C), AgentGEO and repair tools (section 4, Appendix D), MIMIQ benchmark (section 5, Appendix A), main results (Tables 2 and 3), topics and health case (section 6.3, Figure 4), unrecoverable failures (section 6.4), out-of-distribution and HTML results (Table 4) and length reductions (Appendix B). The 5% and 25% edit shares are the authors' summary figures. All figures come from the paper; none were reproduced by KAIS.
- Aggarwal et al. — GEO: Generative Engine Optimization, KDD 2024. The original GEO paper, whose rewriting methods serve as baselines in the study.