AI citation rate by content format: what checks out

Five precise-sounding statistics about which content format gets cited most by AI search have been circulating across SEO and GEO vendor blogs through 2026. Each one is quoted as if it were a single settled fact. Checking each against the page it is supposedly published on turns up a narrower and messier picture: two real studies with real sample sizes, one real academic paper being passed off as proprietary industry research, and three figures that do not survive contact with their own stated source.

The two figures that hold up#

Wix Studio’s AI Search Lab published the most rigorous of the group. Researcher Tom Wells analyzed 75,000 AI answers and 1,056,727 citations retrieved from three engines, ChatGPT, Google AI Mode and Perplexity, using deliberately non-branded prompts so the results were not skewed toward users already naming a company. Listicles accounted for 21.9% of citations, articles for 16.7%, and product pages for 13.7%. The methodology section discloses the model list, the prompt design choice, and the publication date, 2026-03-23, which is enough to treat the number as a real, checkable finding rather than a repeated slogan.

DeltaV Digital, run by VP of Operations Brandon Kidd, published a second real dataset: 21,075 AI engine responses tracked across ChatGPT, Perplexity, Gemini, Google AI Overviews and Google AI Mode between April 14 and July 13, 2026, yielding 25,337 citations to the 240 most-cited URLs across eight industries. Its own page states that ranked listicles reached 61% of citations specifically in the B2B technology services vertical, not as a figure that applies to every industry, and the report’s central argument is that citation patterns differ by industry rather than following one universal format ranking.

A third source, cited by Search Engine Land, comes from Evertune Research: an analysis of 25,000 cited URLs drawn from the 6,000 most-cited URLs per model across six engines, spanning roughly 400 million citations. It found that 63% of the most-cited URLs pointed to listicles, and that ranked lists made up 71% to 86% of those listicles rather than unranked lists or institutional rankings. This is a different cut of the same underlying pattern as the Wix and DeltaV figures, format matters and listicles lead, but it is a distinct, separately disclosed dataset, not a repeat of the same number under a new name.

The three figures that do not check out#

“Original-data studies get a 71% citation rate, comparison pages 64%, ranked listicles 61%.” This exact triple circulates on multiple vendor blogs, several of which cite DeltaV Digital as the source. DeltaV Digital’s own page does not contain these three numbers as a three-way breakdown; it contains one industry-specific figure, 61% for listicles in B2B technology services, that only partially matches. When the pages actually publishing the 71%/64%/61% claim were checked directly, one (MaxAEO) presented it as proprietary internal monitoring with no external dataset link, and another (Presence AI) presented similar but not identical figures, 67% for comprehensive guides, 71% for FAQ schema, as self-published research with no linked raw data. The claim does not trace to one real, checkable study; it traces to at least two vendors publishing similar-sounding but non-matching numbers, neither externally verifiable.

“Original research earns citations at 20x the rate of thin content.” Search results attribute this figure to a synthesis of “four independent 2026 datasets,” and one of the pages carrying that framing, Digital Applied’s own citation-ranking-factors report, was fetched directly. Its real content covers a meta-analysis of several named datasets (a Zyppy 54-experiment review, an Ahrefs 75,000-brand correlation study, a BuzzStream four-million-citation analysis, and others) but does not state a 20x multiplier anywhere in the text. Ritner Digital, another page repeating the general claim that original research outperforms thin content, also does not state a 20x figure when checked directly. The multiplier could not be traced to any single page that actually publishes it with a method behind it.

“Pages combining text, images, video and structured elements show 156% higher selection rates.” This is the clearest case of a number being recycled. Checking three different pages that use “156%” turned up three different, unrelated claims: one site attaches 156% to multi-modal content selection, citing an unfetchable “AI Mode Boost 2025” report and admitting its own breakdown table is “derived,” not copied from a published figure. A second site attaches the same 156% to adding original data or statistics to a page, citing “Authoritas 2026.” A third, separately worded claim on the same general topic instead cites SEMrush for a completely different multiplier, 1.4x, for multi-format content. Three vendor pages, three different underlying claims, one recycled number, and no single fetchable primary source behind any version of it.

A real academic figure wearing an industry-research costume#

One number in the original list of five does trace to something real: a 17.3% citation-rate improvement from structural optimization, tested across six generative engines. It comes from a genuine, fetchable paper, “Structural Feature Engineering for Generative Engine Optimization,” by Yu, Yang, Ding and Sato, posted to arXiv in March 2026. The abstract states the figure plainly and names the six-engine test set, though it does not disclose page or trial counts in the abstract itself.

What makes this one worth a section of its own is not the number, it is who is presenting it. Machine Relations Research’s synthesis page states outright that its 17.3% figure is drawn from “GEO-SFE (Yu et al., University of Tokyo / University of Tsukuba / Hiroshima University / NII, arXiv 2603.29979, March 2026),” which is honest attribution. But other vendor pages in the same search results repeat the 17.3% figure without naming Yu et al. at all, presenting it as though it were proprietary GEO-tool research rather than an academic paper’s finding. A real number, correctly sourced in one place, is being laundered into unattributed vendor content elsewhere, which is a different failure mode than the three figures above that were not real to begin with.

Why we care#

This specific corner of SEO content, precise percentages about which content format AI search cites most, is currently one of the most citation-unstable areas this site has checked. Two of five widely repeated statistics traced to real, disclosed studies. One traced to a real academic paper being presented, in places, as if it were unattributed vendor research. Two more could not be verified on any page that claims to originate them, despite specific search effort against several named vendor blogs.

None of this means content format is irrelevant to AI citation. It means the specific numbers attached to that claim are, right now, mostly unreliable. A reader who sees “71% citation rate” or “20x more citations” quoted without a named study, a disclosed sample size, and a link that actually contains the number, should treat the figure as unverified until proven otherwise, the same standard this report applied to check it.