Case study
How AI recommends medical aid in South Africa
Seven AI models answered the same medical aid recommendation question 1,757 times, across a no-audience baseline and five socioeconomic segments. The answers map how South African schemes are recalled, ranked, and recommended in the AI layer. Two schemes lead, depending on which measure you read.
- AI models
- 7
- segments + baseline
- 5+1
- responses
- 1,757
- schemes tracked
- 9
The headline
The category has two leaders, depending on the measure
Discovery Health owns the reflex. When a model answers the recommendation question, Discovery is the name it reaches for first: 43.1% top-of-mind recall against just 11.0% share of voice. Bonitas owns the discussion: 29.9% share of voice, the highest in the set, with nearly three times Discovery’s share of scheme mentions once the response moves past its opening line.
Which scheme leads depends on whether you value being named first or being discussed most. Discovery sits far above the pack on first mentions; Bonitas far ahead on mention volume.
The attention gap
First mention is a different contest from the conversation
Every one of the seven models makes Discovery Health its most frequent first mention, from 52.4% of responses on Gemini 2.0 Flash down to 35.1% on Sonar Pro. But the runner-up splits by model: on GPT-5.4, Sonar Pro, and Kimi, Bonitas is the challenger; on the two Geminis, Grok, and Claude, Momentum Health takes that role.
A scheme tracking its AI position on one model is reading a single dialect and missing the rest of the language.
The SEM arc
The recommendation changes shape with household income
Across the socioeconomic arc from C1 Traditional to C5 Elite, the mid-tier reshuffles. Momentum Health falls from 14.6% share of voice in C1 to 9.7% in C5, while Bestmed nearly doubles from 8.4% to 16.3%. Profmed, a restricted scheme for degreed professionals, climbs from 0.5% to 4.7%: a nine-fold gain. Discovery holds within a point of 11% everywhere; its position is income-proof.
The baseline column shows the answer models give when no audience context is set. Against that, Momentum over-indexes in lower segments and Bestmed in upper ones; Discovery barely moves.
Ribbon thickness is share of the segment’s conversation; vertical position is rank within the segment.
The affordability frame
AI decided the default buyer is price-constrained
The question named no budget, yet "affordable" is the single largest attribute in the corpus: 17.5% of attribute mentions, appearing in 59.5% of responses. Hospital plans dominate the product-type conversation at 33.2%, and the single most-mentioned plan in the entire campaign is Momentum Ingwe, an entry-level income-scaled option.
For premium-positioned schemes, this is the finding that matters most: the frame of the conversation is set before any scheme name appears, and it is set to price.
The plan layer
Below the brands, a second contest: which plans AI quotes
When a model recommends, it usually names a specific plan: "start with Momentum Ingwe", "look at Bonitas BonCore". Across the responses, models named 431 distinct plan-level entities; the ten leaders carry the conversation, coloured by parent scheme, and seven of them are budget or network options. AI recommends medical aid from the bottom of the price list up.
Momentum Ingwe leads on every measure: 11.9% of plan mentions, present in 48.0% of plan-naming responses, and the highest sentiment of any plan. Practically all of Momentum’s down-market strength in the SEM arc is this one plan. Bonitas plays it differently, fielding four plans in the top ten for a combined 22.7% of the plan conversation: no single star, but coverage of every budget tier the models discuss.
Friction
The warnings AI attaches to the category
Alongside recommendations, the models volunteer warnings. Five pain points recur at scale, ranked by the share of responses that raise them. These are the sentences a prospective member reads immediately after the recommendation.
Network restrictions are the most-raised friction (37.6% of responses), and they land hardest on the cheap network plans the models push most: the advice and the warning travel together. Late-joiner penalties carry the harshest tone, with the lowest sentiment in the set; the models frame them as punitive and permanent, and use them to argue for joining now. Co-payments and waiting periods are the fine-print duo, surfacing as list items under recommended plans rather than as arguments.
None of the five is scheme-specific in the models’ telling. They are category-level warnings applied to whichever scheme is on the table, which teaches a prospective member that medical aid is a product of restrictions and penalties before they ever see a benefit table.
The source economy
Comparison sites feed the answers
Of the 1,757 responses, 1,336 cited web sources. The most-cited domain in the category is hippo.co.za, a comparison site, ahead of any scheme’s own website. Scheme sites earn citations roughly in proportion to their share of voice, which means the schemes being discussed most are also feeding the discussion.
For a scheme wanting to shift how AI describes it, this list is the target surface: these are the pages the models actually read.
All models combined. Filter by model to compare source mixes.
Method
How this study was run
One recommendation question ran through seven AI models (GPT-5.4, Claude, two Gemini versions, Grok, Kimi, and Sonar Pro) in the context of five SEM segments, 50 AI-generated audience members per segment, plus a no-audience baseline. Variant scheme names were merged under parent brands before share-of-voice, top-of-mind, and penetration 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.