Case study
Which bank does AI think is best in South Africa?
Five AI models answered "what is the best bank?" across eight research waves and ten LSM segments, from subsistence households to the private-banking tier. Capitec dominates first mentions; FNB leads the wider conversation; each challenger bank owns a distinct slice of the income ladder; and the ranking shifted between waves.
- AI models
- 5
- LSM segments
- 10
- research waves
- 8
- responses
- 3,763
The headline
Capitec is the reflex; FNB is the conversation
Ask the models to name the best bank and Capitec is the answer roughly seven times in ten: the clear top-of-mind leader all campaign long, against a share of voice around 21%. FNB is the opposite: the most-discussed bank in the corpus, but a distant second on first mentions. The same two-leader pattern appears in medical aid, sharpened here. Being discussed is a different contest from being recommended.
Capitec holds the recall lead wave after wave. FNB holds the edge on share of voice. The challenger pack reshuffles underneath as segment and wave change.
Bubble size is the third metric. Trails mark each bank’s position in earlier waves.
The LSM arc
Each challenger owns a precise slice of the income ladder
The ten LSM segments run from subsistence households to the private-banking tier. FNB holds 25 to 30% share of voice at every step. Capitec holds steady around 20%.
The challengers are where the arc gets interesting. TymeBank peaks in the emerging middle (14.0% at LSM 5) and fades at the top. Discovery Bank climbs from 2.5% at the base to around 10% in the upper bands. Investec is nearly invisible until LSM 7, then climbs to 8.1% at LSM 10. The models have indexed each challenger’s market position precisely.
Ribbon thickness is share of the segment’s conversation; vertical position is rank within the segment.
The wave trend
Capitec overtook FNB by Wave 8
Wave 1 was dominated by Standard Bank at 37.6% share of voice, a one-wave artefact of that wave’s model mix. From Wave 2 the field settled: FNB led, Capitec tracked it, and by Wave 8 Capitec had overtaken FNB, 23.2% to 19.6%.
Wave-over-wave movement is the metric a single snapshot cannot give you: the window where a challenger overtakes an incumbent has already closed once in this data.
Share of voice by segment. Click a scheme to show or hide its line.
Model dialect
Five models, one ranked list, different weightings
All five models produce the same top-three (FNB, Capitec, Standard Bank in some order), but the weightings diverge sharply. Claude gives Standard Bank 27.5% of the conversation; DeepSeek gives FNB 28.7%; GPT-5.4 Mini leans Capitec at 30.1%. Gemini is the outlier: it gives Discovery Bank 12.0%, roughly double any other model, and TymeBank 15.2%.
A bank tracking its AI presence on one model is reading one dialect. The Gemini and Discovery Bank alignment alone would mislead a strategy built on a single-model view.
Method
How this study was run
One question ran through five AI models (GPT-5.4 Mini, Claude, Gemini, Grok, and DeepSeek) across eight research waves, against a no-audience baseline and ten LSM segments spanning the full South African income spectrum. Brand variants were merged under parent names before metrics were calculated.
The same design runs on any category: segments, entity classes, and questions are all configurable per campaign.
Run this study on your market.
Book a demo to run the same research design on your category.