The first publicly pre-registered benchmark measuring how three frontier Large Language Models (OpenAI GPT-5, Anthropic Claude Sonnet 4.5, Google Gemini 2.5 Pro) cite source material when answering South Africa–specific queries.
Conducted by Cited Brands (citedbrands.co.za). 5,500 prompts across 10 South African consumer industries (banking, telecom, grocery retail, medical aid, short-term insurance, automotive/EV, e-commerce, restaurants, streaming, real estate). 100 brands. 1,100 unique prompts × 5 replications = 5,500 prompts × 3 LLMs = 16,500 API calls.
Pre-registered hypotheses (H1–H7) include: SA-domain citation share, model-specific citation patterns, Latin Square position bias, reputation polarity gap, multilingual coverage in Afrikaans and isiZulu, Reddit citation asymmetry, and search-budget standardization validity.
Methodology: blind prompts, Latin Square counterbalancing for comparison questions, Bradley-Terry pairwise MLE for brand strength, bootstrap confidence intervals at prompt level, Cohen's kappa for inter-rater URL classification reliability.
Public release: protocol, code, prompts, classification rubric, and aggregate citation dataset under CC-BY-4.0. Raw response text retained by Cited Brands for commercial use (disclosed).
Total study cost: ~USD 1,450 (~ZAR 26,500). API Cost only no development cost included Re-runnable by any researcher with API access for the same cost.