How Medical Aid Companies Can Improve Their AI Search in South Africa (2026)

If you market a SA medical aid, this is a thing your CEO will ask you about in the next six months: when a member asks ChatGPT which scheme to pick, does our name come up? For most of the industry the honest answer is “sometimes, and we don’t know why.” This post is the playbook to fix that.

We’ve been running an AI-citation benchmark across GPT-5, Claude Sonnet 4.5 and Gemini 2.5 Pro for six months — 100 SA brands, 232,254 cited URLs classified per source. The medical-aid slice is one of the cleaner industries in the dataset. The patterns are visible, the levers are short, and the per-scheme baselines live in the gated Quarterly Index Report. Five moves, in the order that actually matters.

First: find out where you sit today.

None of the moves below mean anything without a baseline. The bench measured ten SA medical aids on two numbers per scheme — blind visibility (the share of open-ended “recommend me a medical aid” prompts your name appears in) and complaint-platform exposure (the share of cited sources about you that come from Hello Peter, Trustpilot or member-review sites).

Discovery is the standout — named in 93% of open-ended prompts, the only SA medical aid above 90 in our data. Below Discovery the gap to the next scheme is large; below the top three the grade gets noticeably steeper. The per-scheme breakdown of all nine schemes lives in the gated Index Report.

The free Brand Scorecard returns your own numbers across all three AI models in under a minute. That’s the baseline. Once you have it, the rest of this post is the playbook.

Move 1 — Fix Hello Peter first.

Hello Peter is among the most-cited SA sources in the whole dataset and the only complaint platform in the SA top set. When you bin AI’s descriptions by the scheme’s Hello Peter star average, you don’t see a smooth slope. You see a flat line that bends sharply at a specific threshold — the value sits inside the Index Report. We re-ran the prompts to make sure the kink wasn’t noise. It wasn’t.

Above the threshold you get described neutrally. Below it you get described with defensive caveats — “some members report”, “reviews are mixed” — even when the recent reviews on the page are good.

The work, in order: assign one person to Hello Peter response within 24 hours of every complaint, resolve in private, ask publicly-resolved reviewers to update their rating (the platform permits it), and stop responding to complaints with policy quotes — AI reads tone, not just text. Most schemes cross the threshold in about 90 days when this becomes one person’s job. Nothing else on this list moves the AI description as quickly.

Move 2 — Win four publishers, not forty.

AI’s medical-aid description is composed from a short list of SA publishers. businesstech.co.za is the most-cited business publisher across the whole bench — medical aid included. The next three publishers that actually move the AI description for SA medical-aid queries are named in the Index Report, alongside the per-publisher pitch templates that have landed coverage in the last six months.

Pitch the four. Pitch them with bench data, not press releases — SA business editors are inundated and the only thing that gets through is research they can cite. Skip the rest of the trade press for AI-citation purposes; they don’t move the number.

If your PR partner doesn’t already have relationships at all four, that’s the line item to renegotiate. If they do, ask them what the last six months of placements there have been. If the answer is “none”, you have your problem.

Move 3 — Audit your prompt-conditional positioning.

Add the word “affordable” to a SA medical-aid prompt and one scheme’s share of mentions jumps. Add “top-quality” and it falls. That’s consistent across all three AI models. Most schemes have accidentally inherited a price-or-quality narrative that helps on one prompt class and caps on the other — and most don’t know which version they have.

You find out by running a prompt sweep: identical question, twenty word-substitutions, see how the brand-mention rate shifts. If you find a positioning you don’t want, the fix is editorial: the publishers from Move 2 need to start covering you in the frames you do want. AI catches up within a quarter when the source mix changes.

This is the one move that needs your own Scorecard data to start — the public bench shows industry-level patterns; your prompt-conditional shape is brand-specific and lives in the personalised report.

Move 4 — Build a content engine that compounds.

The leading scheme’s AI visibility isn’t about ad spend. It’s about a non-product content programme that generates a constant stream of news coverage AI reads. The detailed teardown of that programme — what it produces, how the citation lift compounds, what a SA scheme would need to replicate the shape — sits inside the Index Report.

The general principle is straightforward enough to state in public: build something that produces non-product content at scale. A research programme, an annual category report, a podcast, a foundation — anything that generates coverage that mentions your brand without being an ad for your brand.

Move 5 — Measure quarterly.

AI changes. OpenAI altered retrieval in late 2025 and Reddit citations across ChatGPT dropped from majority to almost nothing in a single update. Google AI Overviews change every two days. What worked in Q1 may not hold by Q4. The moves above need to be re-checked each quarter against a fresh bench.

We refresh SA-AEO-Bench every three months and republish the medical-aid chapter of the Index Report. If your in-house tracking matches our bench movement, your work is hitting. If they diverge, one of the two is measuring something the other isn’t — usually a sign to ask why.

What about the differences between the three AI models?

The three models disagree visibly on banks and telcos. They mostly agree on medical aid — Discovery first in 100% of prompts, the next four in the same order 80%+ of the time. The probable reason is publisher concentration: medical-aid coverage in SA lives on a small high-overlap set, so all three models pull from the same sources and converge on the same answers.

For you that means you can optimise for one model and the others mostly come with you. The same isn’t true in banking — you genuinely have to run separate per-model strategies there. Medical aid is the easier industry to win in for that reason.

Frequently asked questions

How much of this can my existing SEO agency do?

Most of moves 1, 2 and 5. Hello Peter is a customer-service workflow more than an SEO one. Publisher pitching is the same skill as PR. The prompt-substitution audit (move 3) and the content-engine build (move 4) are newer disciplines and most SA agencies are still climbing the learning curve on them — ask yours for examples.

How fast does this stuff actually move the needle?

Hello Peter score lifts show up in AI descriptions within 30–60 days. Publisher coverage lifts show up within 60–90 days. The content-engine play (move 4) is a 12-month build. The bench reruns quarterly and you’ll see the trend lines move at that cadence.

What if my scheme is in the bottom half of the ranking?

The grade is steeper but the playbook is the same. Schemes below position 6 in our data have, on average, twice the complaint-platform exposure of the top three — which means move 1 (Hello Peter) is even more leveraged for you, not less.

How does this interact with broker-channel marketing?

Brokers operate on different incentives and this bench measures consumer-direct AI prompting only. If broker-channel AI tooling becomes more common in 2026 we’ll add it to the bench. For now, the moves above are consumer-funnel work.

Can we get a personalised version of this analysis?

Yes — the free Brand Scorecard returns your scheme’s baseline across all three models. For a deeper read (per-prompt breakdowns, named publisher gaps, competitive-set analysis) book a call from the scorecard page.

---

One thing that didn’t fit anywhere above and probably deserves its own post: the Council for Medical Schemes published its 2025 annual report in March and the formal-complaint rankings don’t correlate with AI’s complaint-platform exposure rankings at all. Schemes near the top of the Council’s list aren’t the schemes AI cites complaint platforms about most often. The public-complaint surface and the regulator-complaint surface seem to draw from different populations. Worth digging into.

Last updated 23 May 2026. SA-AEO-Bench v1.2 pre-registered at osf.io/w4az2. Quarterly refresh on the next bench cycle.