Publicações

Three Ways to Fail to Conclude: A Null-report on Large Language Model Citation Claims for Brazilian Brands (N = 7,052, 12 days)

Null-report SSRN · 7.052 respostas em 12 dias · 3 hipóteses populares do GEO brasileiro refutadas por mecanismos independentes

SSRN eLibrary · Behavioral Marketing + Artificial Intelligence (Elsevier) · posted 6 mai 2026

Generative Engine Optimization (GEO) has emerged as a practitioner discipline asserting a growing list of causal claims about how large language models (LLMs) cite brands. We collected 7,052 LLM responses over 12 consecutive days (2026-03-24 to 2026-04-22) probing 69 entities (61 real brands and 8 fictitious decoy entities designed for false-positive calibration) across 4 verticals (fintech, retail, health, technology) against 4 production LLMs (OpenAI gpt-4o-mini-2024-07-18, Anthropic claude-haiku-4-5-20251001, Google gemini-2.5-pro, Perplexity sonar). We formulate three hypotheses popularly asserted in the Brazilian GEO market and analyse them under Benjamini-Hochberg false discovery rate correction, BCa bootstrap confidence intervals (10,000 resamples), and cluster-robust standard errors by collection day. The aggregate citation rate is 77.62% (95% BCa CI [76.62%, 78.57%]). For all three focal hypotheses we fail to reject the null but by independent mechanisms. Vertical heterogeneity (Cramér V = 0.23, p < 1e-82) and Portuguese-English divergence (h = 0.136, p < 1e-8) survive correction and subsampling. We release preregistration, code and data for adversarial replication.