How AI search engines cite South African enterprise brands.
188,877 citations classified. 14,826 successful AI responses. 100 brands across 10 industries and 3 frontier LLMs (GPT-5, Claude Sonnet 4.5, Gemini 2.5 Pro).
Pre-registered at osf.io/w4az2 before data collection — methodology, prompt set, and analysis code public. All seven pre-registered hypotheses (H1–H7) are confirmed or tracking in their predicted directions.
Five findings no published SA study has surfaced before.
- 01
MTN is the brand AI cannot stop talking about. Bidvest Bank is the brand it has forgotten.
Across blind organic queries, MTN surfaced in 99% of “best telecom” questions. At the other end, Bidvest Bank surfaced in 0%. Every percentage point of organic visibility is direct AI-channel demand capture.
- 02
A 42-percentage-point SA-citation gap separates the most-protected industry from the least.
Short-term insurance routes 72.7% of citations to SA sources — the most SA-protected category. Restaurants routes only 29.8%. SA brands in low-share industries fight AI’s preference to route to international competitors; brands in high-share industries inherit a moat.
- 03
Negative-framing queries route 64% to international sources.
When the same brand is asked about positively vs negatively, the SA-citation share drops from 44.4% to 36.4% — a −7.9pp shift toward international complaint platforms. Reputation surfaces sit on Trustpilot, Complaintsboard, and PissedConsumer for the queries that matter most for purchase decisions.
- 04
Gemini cites Reddit 698 times. GPT-5 and Claude cite Reddit zero times.
The single sharpest model-specific finding in the dataset. Pre-registered H6 set the threshold at 5×. Observed ratio: ∞ (literally infinity — the other two return zero). Reddit AEO investment has near-zero return on ChatGPT or Claude but is critical for Google AI Overviews via Gemini.
- 05
Brands with high ‘sycophancy uplift’ have AI visibility scores that are partly an artefact.
When the brand name is in the prompt, AI returns it confidently. When it is not, the brand may surface at 0%. Ten brands in the dataset have a +100pp blind-to-named gap. Any AEO measurement tool that does not compare blind vs named probes overstates these brands’ real organic AI standing.
Seven hypotheses, locked in before the data was collected.
Pre-registration on OSF (osf.io/w4az2). Hypotheses + analysis plan + prompt set + scoring code committed before any LLM was queried — eliminating the “p-hacking” degree of freedom.
Where AI’s SA-source preference is strongest, and where it’s weakest.
Share of citations routed to South African source domains by industry. Higher = AI prefers SA sources. Lower = international platforms dominate.
“What’s the best telecom in South Africa for a small business?”
Brands AI talks about without prompting.
- 01MTNtelecom99%
- 02Vodacomtelecom97%
- 03Takealotecommerce96%
Brands AI cannot recall when asked.
- 01Bidvest Bankbanking0%
- 02Afrihosttelecom0%
- 03Webafricatelecom0%
SA-domain citations across all 3 LLMs.
The operational target list for any SA brand competing for AI visibility. Editorial placement on these domains compounds across every engine measured.
The single sharpest model-specific finding in the dataset.
Gemini cites 698 Reddit pages for SA brand queries. GPT-5 and Claude cite Reddit zero times.
Most SA marketing teams have no idea this asymmetry exists. Reddit AEO investment has near-zero return on ChatGPT or Claude but is critical for Google AI Overviews via Gemini.
Framing changes which sources LLMs surface.
SA-share of citations by prompt framing. Negative-polarity queries route 63.6% to international sources — a structural gap of −7.9pp vs positive framing.
The three frontier engines cite SA at different rates.
Download SA-AEO-Bench v1 · 2026-05-19 snapshot.
Per-brand citation breakdowns. Per-industry SA-share. Full hypothesis test results. Top-25 SA + model-exclusive domain lists. The complete pre-completion deliverable as a PDF + raw JSON dataset.
Browse research documents →