What actually moves AI citations — and what doesn't

The GEO advice ecosystem has produced a confident set of beliefs about what makes AI systems cite your content: structured markup, content freshness, data licensing, brand search volume. We have spent time with the evidence behind each of these. Several are wrong in specific, testable ways.

The factors that actually move AI citations form a short list: domain authority, multi-platform brand presence, and specific content interventions. The factor most practitioners focus on — JSON-LD schema markup — produces approximately zero direct lift, confirmed across four independent sources. And a number of claims circulating to explain this or other citation patterns are refuted outright by the evidence.

Schema Markup Does Not Move AI Citations, and the Studies Cited to Prove It Are Contested

Schema.org and JSON-LD markup produce approximately zero direct lift in AI citations. This is verified across four independent sources. The mechanism is straightforward: language models read and represent text; they do not parse structured metadata the way search crawlers do. There are legitimate uses for schema in structured data pipelines, but improving AI citation rates is not one of them.

The practitioner literature has assembled a body of attribution claims to support this conclusion that itself requires scrutiny. A frequently cited claim holds that an Ahrefs study tracked 1,885 sites adding JSON-LD and found zero measurable lift in AI Overviews citations; that specific version is refuted. The study's existence and its AI Mode findings remain contested. A second attribution — that a February 2026 controlled experiment by Mark Williams-Cook established the ~0% schema finding — is also refuted. The finding about schema's lack of direct lift is real; these specific sources cited to support it are not.

Two further myths worth disposing of: schema is not required for Knowledge Graph indexing — that claim is refuted. And the $130 million data licensing deals attributed to OpenAI and Google in Q4 2025 through Q1 2026, sometimes cited as evidence of behind-the-scenes citation influence, did not occur in the form described — also refuted.

Referring Domains Predict Citation Rate Better Than Anything Else We Can Measure

ZipTie's 2026 analysis found that sites with 420 referring domains achieve a 12 percent AI citation rate; sites with 3,200 or more referring domains achieve 68 percent. Domain authority outweighs schema's contribution to AI citation weighting by a factor of 3.5 to 1. This gap is not subtle, and it is consistent with how language models are trained: on text drawn from trusted, widely-linked sources.

Multi-platform presence compounds the effect. Brands present on four or more platforms are 2.8 times more likely to appear in AI responses, according to Eric Buckley's 2025 analysis. The pattern matches the referring-domain finding: independent corroboration across multiple sources functions as a trustworthiness signal regardless of what structured markup exists on any individual page.

Brand search volume does correlate with AI citations across platforms — confirmed at r=0.334 — but that correlation has not been experimentally validated as causal. The claim that brand search volume is the strongest measurable correlate of AI citations is refuted; the correlation exists, but it is not the dominant predictor, and causal direction is not established.

Citing Sources and Writing Well Move Citations; Keyword Stuffing Actively Hurts

Princeton's GEO KDD 2024 randomized controlled trial established that explicitly citing sources in web content produces a 30 to 40 percent increase in AI citations. The causal effect of direct quotations on citation likelihood is confirmed by the same study. Both findings point toward the same mechanism: AI systems appear to weight content that demonstrates epistemic grounding, and that grounding shows up in citation practice.

The compound intervention worth noting: combining fluency optimization with statistics addition produces the largest measurable effect, confirmed across four independent sources. Keyword stuffing, by contrast, produces zero or slightly negative AI citation effect — not neutral, actively counterproductive. On exact magnitudes for each intervention in isolation, the data is genuinely contested: trial data suggests fluency optimization may produce a 15 to 30 percent lift and statistics addition a 30 to 41 percent lift, but these figures are not settled and should be treated as working estimates rather than benchmarks.

Two specific claims from the Princeton study also require correction. The claim that lower-ranked sites at positions 5 through 10 gain up to 115 percent in AI citations by citing sources — attributed to that same RCT — is refuted. Claims that direct quotations alone produce a 28 to 40 percent boost in RCT conditions are similarly refuted; the causal effect of quotations is confirmed, but that specific magnitude from controlled trial conditions is not. Whether Q&A format is the optimal content structure for AI citation is genuinely contested and should not be stated as settled.

The Social Layer Has Grown Into a Structural Input

Reddit AI citations in commercial categories grew 73 percent from Q4 2025 through Q1 2026. Thirty-one percent of Perplexity citations come from social and forum sources. A content strategy that treats these channels as supplementary is optimizing for a fraction of the citation surface. The corroboration logic applies here too: community-validated claims on high-authority forum threads appear to function as a trustworthiness input independent of individual page domain authority.

The pattern across this data is consistent enough to support a simple thesis: AI citation is a trust problem, not a markup problem. Domain authority, multi-platform corroboration, evidenced prose, and community presence are the levers. Schema, keyword density, and content freshness as a standalone tactic are not — the claim that substantive updates within two to three months reliably double citation rates, attributed to an AirOps 2025 temporal audit, is refuted. The shortcuts that get the most airtime in practitioner discussions are, on the evidence, either wrong or unverified. The interventions that work are slower and harder to fake.

The receipts: every fact, adjudicated

The article above asserts nothing our truth engine did not independently verify. Here is the full ledger — each fact, its verdict, confidence, authority tier, and corroborating sources — so you can check our work. Verdicts are grounded in the fetched content of each source, not a model's priors.

Verified facts

Corroborated and adversarially adjudicated — the evidence bar was met.

Contested

Evidence is split. Shown with both sides visible; do not treat as settled.

Unverified

Insufficient evidence either way — monitored, not asserted.

Refuted

The truth engine refuted these — shown so they are not repeated as fact.

How we verified this

Of 56 atomic facts: 21 verified, 0 supported, 26 contested, 9 refuted. Each fact was judged by a 3-reviewer adversarial panel prompted to refute it. A fact reaches 'verified' only when corroborated by at least two independent sources or one high-authority source AND the panel agrees. A single low-authority source reaches 'supported' at most — never 'verified'. Sources are scored per-fact: the same article can be right about one fact and wrong about another. This report is regenerated when new research shifts a verdict; the date reflects the last adjudication.