GEO-PTBR: a Brazilian-Portuguese replication of Generative Engine Optimization

Optimizing for one generative engine can penalize you on another. In this replication of the nine content-side GEO techniques (Aggarwal et al., arXiv:2311.09735) on a 525-query Brazilian-Portuguese benchmark across three engines from three families (gemini-3.5-flash-lite, claude-haiku-4-5, gpt-5.6-luna), the three engines agree on the direction of the effect in only 4 of 9 techniques — and all four are techniques that reduce visibility. The largest effect measured anywhere in the study is a −81.5% penalty, not a gain; the original English-language study's best techniques (+27% to +41%) do not reproduce in magnitude, with the largest positive effect at +2.6%. Five of the nine techniques induce an engine to cite a source that does not answer the question, including on the health third of the benchmark — a safety failure, not an optimization. Includes the benchmark, the measurement pipeline, the compiled paper (47 pp.), and the per-engine and cross-engine results with bootstrap confidence intervals and Holm correction over a family of 27 tests. Study landing page, with the full findings in Portuguese: https://www.aeobr.com.br/estudos/paper-geo-ptbr/ Code is MIT-licensed; data and paper are CC BY 4.0. Canonical dataset copy: https://huggingface.co/datasets/epicchi2103/geo-ptbr — source repository: https://github.com/epicchiotti2103/geo-ptbr Produced by Caracol AEO (AEO BR), Caracol Media, Brazil: https://www.aeobr.com.br