Research · Evidence review Sources & citations

What actually gets cited by AI search: a sober review of the controlled evidence

The GEO advice market runs on writing tricks — add statistics, add quotes, sound authoritative, get cited. The controlled experiments tell a different story: retrieval and position dominate, most content edits do nothing measurable, and being cited is not the same event as shaping the answer. Here is what the strongest public evidence supports, with every primary source linked.

Fifty-four edits went in. Three came out.

The most direct controlled stress test of "conversational SEO" tactics we reviewed is C-SEO Bench, a benchmark accepted at NeurIPS 2025 that evaluated 54 style and content edit conditions across two tasks and six domains. Only 3 of the 54 conditions produced a significant positive improvement in citation ranking. Official primary arxiv.org/abs/2506.11097

The same benchmark found something the advice market rarely mentions: moving a retrieved document toward the front of the model's context produced a far larger effect than any of the content edits tested. And when multiple competing documents adopted the same optimization, the gains shrank — a tactic that works because nobody else uses it is not a strategy — it is a window. Official primary arxiv.org/abs/2506.11097

C-SEO Bench: 3 of 54 edit conditions showed a significant positive effect
C-SEO Bench (NeurIPS 2025): 54 style and content edit conditions tested; 3 showed a significant positive effect on citation ranking. Highlighted positions are illustrative — the point is the ratio.

252,000 runs later, the factor ranking is consistent

A second controlled study, "What Gets Cited," ran 252,000 trials across 6 LLMs, injecting two candidate documents at a time and systematically varying 18 factors with brand anonymization and order counterbalancing. The largest citation-selection factors were topical relevance to the query and position in the context. Explicit pricing information and recent dates showed smaller but consistent positive effects across models. Formatting-only changes — the cheapest and most heavily marketed lever — were negligible. Official primary arxiv.org/abs/2605.25517

Read those two results together and a hierarchy emerges. Whether your page is retrieved as a candidate at all, and where it lands, dominates. Concrete, checkable facts on the page — prices, current dates — help at the margin. Cosmetic rewriting is noise. If your optimization budget is allocated in the reverse order, the controlled evidence says you are paying for the weakest lever first.

Cited, absorbed, supported — three different events

Even a citation is not the finish line. An observational analysis of 602 prompts, 21,143 citations, and 18,151 cited pages found that how often a source gets cited and how deeply it actually shapes the answer are decoupled — pages with real influence on the generated text tended to be longer, well structured, and dense with definitions, numbers, comparisons, and procedural evidence. Official primary arxiv.org/abs/2604.25707

And the citation link itself can overstate what the answer got right. A peer-reviewed human evaluation of four commercial generative search systems, published in Findings of EMNLP 2023, found that on average only 51.5% of generated sentences were fully supported by their citations, and 74.5% of citations actually supported the sentence they were attached to. That is a 2023 snapshot of earlier products — it should not be quoted as a current error rate — but it established the structural point: citation presence and statement support are separate measurements. Official primary aclanthology.org/2023.findings-emnlp.467

Evaluating Verifiability (Findings of EMNLP 2023): citation presence vs. statement support Sentences fully supported by their citations 51.5% Citations that supported their linked sentence 74.5% Human evaluation of four commercial generative search systems, Findings of EMNLP 2023 — a 2023 product snapshot, not a current error rate.
Citation presence and statement support diverge — which is why they have to be measured as separate events, not folded into one visibility score.

How to read this evidence like an operator

Every result above is causal only inside its test bed. C-SEO Bench used fixed candidate documents and specific model snapshots; "What Gets Cited" measured first citations between two injected documents, with no live retrieval or indexing in the loop. That is what a controlled benchmark means: strong internal causality, no claim to being a ranking law of any live commercial engine. Anyone quoting these numbers as guaranteed field effects is stretching the evidence — and anyone dismissing them is discarding the only clean causal data the industry has.

The decision-grade takeaway — the strongest controlled evidence puts retrieval eligibility, topical relevance, and context position first; checkable facts like pricing and current dates second; and generic style rewriting last. And it shows that being cited, being absorbed into the answer, and being supported by the answer are three different events that move independently. A single blended "AI visibility score" hides exactly the distinctions the evidence says matter.

One more filter worth copying: this review used controlled benchmarks, peer-reviewed studies, and platform documentation only. Studies built on data collected and published by GEO tool vendors were set aside — not because they are wrong, but because they cannot separate observation from sales motive, and effect claims deserve sources without a horse in the race.

Buy measurement for the chain, not tricks for one link

The practical consequence of this evidence is an ordering discipline. Before spending on content rewrites, you need to know where your brand actually drops out of the chain: are your pages retrieved as candidates at all, are they cited, do the answers absorb your facts, and do the statements about you hold up against your actual pages? Each stage has a different fix, a different cost, and a different payoff — and the controlled studies above show the stages are not interchangeable.

That chain is what CiteAngle measures. We observe real AI-search answers for the questions your buyers ask in your market, separate appearance, brand mention, and source citation into distinct measured events, and hand you the stage-by-stage picture the controlled evidence says you should be managing — so your next dollar goes to the link that is actually broken.

See where your brand stands in the citation chain — which questions surface you, which answers cite you, and which competitors hold the source slots you want. Request your visibility briefing or review the measurement programs.

Ready to turn US visibility evidence into a market plan? Tell us the US buyer, target state and decision date.

Request a working scope