Most citation failures are relevance failures

A citation failure is the specific reason an answer engine retrieved your page and then declined to cite it, and the reason is usually not technical.

That last part is the useful finding. Most GEO tooling, ours included, starts at the technical layer because that layer is cheap to measure. A March 2026 paper puts a number on how much of the problem actually lives there, and the number is small.

#The study

Zhihua Tian and co-authors built a diagnostic framework for citation failure rather than another rewriting recipe (arXiv 2603.09296). Their method is simple to describe. For each query where a retrieved webpage went uncited, they paired it with a competitor that was cited on the same query, which isolates the marginal difference between the two. That produced 949 contrastive pairs (arXiv 2603.09296).

The evaluation ran on a benchmark of 204 webpages with 12,240 queries, 60 per page, split into training and held-out test queries so the measurement rewards generalization to unseen questions rather than fitting one prompt (arXiv 2603.09296).

#Where citation failures actually happen

Sorting those 949 pairs into failure dimensions produced a lopsided distribution (arXiv 2603.09296):

Failure dimension	Share	What it covers
Semantic alignment	62.2%	Intent divergence, contextual gaps, outdated information, localization mismatch
Content quality	27.1%	Information scarcity, fragmentation, excessive verbosity, unstructured layout
Technical integrity	10.1%	Access blocking, JavaScript rendering failures, connection errors, unparseable content, boilerplate noise
Systemic exclusion	0.6%	Competitive redundancy, context window truncation

Read the technical row carefully before you act on it. The small share is not an argument for ignoring your crawler. Technical integrity is a gate, not a lever: a page behind a blocked user agent or rendered entirely in client-side JavaScript scores zero on every other dimension, and no amount of rewriting reaches it. The reason the category accounts for only 10.1% of failures is that most pages in the corpus already clear it (arXiv 2603.09296). Clearing a gate you have already cleared buys you nothing further.

The two middle-of-the-pipeline categories are where the work is (arXiv 2603.09296). Semantic alignment at 62.2% means the page loaded fine, parsed fine, and simply answered a different question than the one asked. Content quality at 27.1% means it answered the right question badly.

#Why generic rewrites backfire

The comparison in the paper is against AutoGEO, an ICLR '26 system that prompts frontier models to explain what generative engines prefer, extracts those explanations into reusable preference rules, and rewrites pages against them (arXiv 2510.11438). AutoGEO reports an average improvement of 35.99% on GEO metrics (arXiv 2510.11438). It is a serious system, and it is the closest published analogue to what a GEO checklist does: apply the aggregate best practice to every page.

The diagnostic paper reports that AutoGEO performs worse than the unoptimized baseline in several topic categories, and that the degradation concentrates where the original citation rate was already high (arXiv 2603.09296). Health content is the worked example. Its starting citation rate was already around 80%, leaving little headroom, and generic rewriting removed domain-specific information that was doing the work (arXiv 2603.09296).

That is the mechanism worth internalizing. Aggregate rules are derived from the average page. Your best-performing pages are, by definition, not the average page. Applying the average to them regresses them toward it.

The targeted alternative is cheaper as well as better. The diagnostic system reached a 79.52% citation rate against 56.58% for the unoptimized documents, beating AutoGEO by 10.72 points, while modifying about 5% of content on average against 25% for the baselines (arXiv 2603.09296).

#Some pages cannot be fixed by editing them

The paper is honest about its own ceiling. For certain webpages, diagnostic optimization also failed to lift citation. The case study is a university machine learning course page that accepted every content improvement and still lost to well-known online education platforms, which the authors attribute to domain-level bias that content editing cannot reach (arXiv 2603.09296).

This matches what the July 2026 critical survey of 45 GEO studies describes as a stochastic, partially observable pipeline running from crawling through retrieval, reranking, citation, and absorption, where competition erodes individual gains (arXiv 2607.14035). If the deciding factor sits at the domain level for a given query, the page is not the unit you can move. Recognizing that early saves a quarter of rewriting effort on content that was never in contention.

#Caveats

Three, and they matter for how hard you lean on the percentages.

The failure taxonomy was derived by analyzing the contrastive pairs with language models. The paper does not report human annotation, inter-annotator agreement, or a validation procedure for the category assignments, so treat the split as an ordering of magnitudes rather than a precise measurement.

The results come from a research benchmark evaluated with two models, a GPT and a Claude variant (arXiv 2603.09296). That is not the same as ChatGPT, Perplexity, or Google's AI Mode running against the live web with their own retrieval stacks and their own commercial incentives.

The technical share is conditional on the corpus (arXiv 2603.09296). A benchmark of research-selected webpages skews toward pages that already render and already permit crawling. On a real site with an aggressive CDN, a client-side-rendered app shell, or a robots.txt that has not been reviewed since the AI crawler wave, the technical share of your own failures will run higher than a tenth.

#What this changes

Stop treating GEO as a checklist applied uniformly and start treating it as triage.

Clear the technical gate once, per site, and then stop optimizing it. Confirm the page is fetchable by the crawlers you want, that its substance is in the server-rendered HTML, and that the main content is not buried under boilerplate. Those are pass or fail conditions, not a scale.

Then diagnose per page and per query. For a page that is not getting cited, the first question is which failure dimension applies, not which best practice is missing. Intent divergence is fixed by answering the question the query actually asks, which sometimes means a different page rather than an edited one. Outdated information is fixed by updating and restamping. Information scarcity and verbosity pull in opposite directions, and applying the wrong one makes things worse.

And leave your winners alone. A page already earning citations for its query set has a demonstrated fit that no aggregate rule knows about. The measurable risk in rewriting it is not that the rewrite does nothing. It is that the rewrite deletes the specific detail the engine was citing.