A friend of mine spent last weekend choosing a medical aid for her family, and she did not open Google to do it. She is a homeowner in Johannesburg with two children and no particular interest in marketing. She opened ChatGPT, asked it in plain English which medical schemes were worth considering for a household her size, and received a confident, well-organised answer that named three schemes. The scheme she had been paying for years was not among the three.
Her medical aid had not lost a search ranking that weekend. It had lost the conversation.
This is the part of brand reputation that almost no South African marketing team is measuring yet, and it is the reason I started Cited Brands. When a customer asks an AI assistant what to buy, the assistant does not return ten blue links and let the customer decide between them. It decides. It names a few brands, describes each one briefly, and then moves on to the next question. If a brand is not among the few that get named, it was not outranked in any sense the marketing team would recognise. It was left out of the sentence, and the customer never saw that anything was missing.
To understand how often this happens to South African brands, we ran the first publicly pre-registered benchmark of how three frontier AI models cite their sources when answering questions about South African companies. The three models were OpenAI’s GPT-5, Anthropic’s Claude Sonnet 4.5, and Google’s Gemini 2.5 Pro. The study covered 100 brands across 10 consumer industries, and it produced 16,500 AI responses containing 197,156 individual citations, every one of them classified by source type. We registered the full protocol in public before collecting any data, at osf.io/w4az2 . The headline result is uncomfortable but straightforward: most South African brands are losing visibility in AI search without realising the channel exists. This guide explains what AI search reputation is, what our benchmark data shows about it, and what a South African brand can do in response.
What is AI search reputation, and how is it different from SEO?
AI search reputation is what large language models say about your brand, and which sources they trust to describe it, when a customer asks them a question instead of typing that question into a search engine.
The difference from traditional search optimisation is structural rather than cosmetic. Google hands the customer a page of results and lets them choose between the options. An AI assistant reads a slice of the web on the customer’s behalf, forms a view, and delivers a single composed answer. Two consequences follow from that change. The first is that there is no second page of results, so a brand is either inside the answer or it is invisible. The second is that the customer cannot see what the model chose to leave out, which means there is no moment of doubt that sends them looking further. The edited answer feels complete.
AI search reputation therefore has two distinct layers, and a brand can fail at either one of them independently. The first layer is visibility, meaning whether the model mentions you at all when a customer asks an open question in your category. The second layer is narrative, meaning that when the model does discuss you, the sources it reads and the story those sources tell are working in your favour. A brand can be highly visible and still have its entire story narrated by complaint sites. A brand can have a spotless reputation and still be completely invisible. Both situations are failures, and they require different remedies.
For South African brands there is a third layer, and it is the one most likely to be overlooked. That layer is locality. An AI model answering a South African question can draw on South African sources or it can reach for international ones, and when it reaches offshore your brand’s story gets retold inside a context that was built for other markets. As the benchmark data below demonstrates, that borrowed context is rarely arranged in your favour.
How much of what AI says about South African brands comes from South African sources?
A little over half of it. Across all 197,156 classified citations, 53% pointed to South African domains and the remainder pointed offshore. The three models sat inside a fairly tight band, with GPT-5 drawing 62% of its citations from South African sources, Gemini 56%, and Claude 51%.
That overall average conceals the finding that genuinely matters for strategy. The South African citation share is not distributed evenly across industries. It ranges from 72.0% at the top down to 29.8% at the bottom, a gap of 42 percentage points between the most locally grounded category and the least grounded one.
| Industry | SA-source citation share |
|---|---|
| Automotive | 72.0% |
| Short-term insurance | 71.8% |
| Telecom | 60.0% |
| Banking | 59.3% |
| Medical aid | 57.1% |
| Streaming | 53.0% |
| Retail | 51.3% |
| Real estate | 45.3% |
| E-commerce | 33.2% |
| Restaurants | 29.8% |
The most useful way to read that table is as a measure of difficulty. A short-term insurer is operating in a category where AI search already prefers South African sources, because local comparison sites dominate the topic and foreign competition for citation space is close to nonexistent. That preference is a competitive moat, and it belongs to the incumbents to either defend or waste. A restaurant group or an e-commerce brand faces the opposite condition. The models treat those categories as regional branches of a global one, reaching by default for international review platforms and worldwide benchmarks, which means a local brand has to overcome that gravitational pull before it even begins competing against the brand down the road.
The practical consequence is that there is no single AI search strategy that works for every South African brand. An insurer is defending an existing advantage. A restaurant group is fighting the model’s default behaviour. The same quantity of effort produces very different results depending on which side of that 42-point gap a brand happens to sit.
Which South African brands does AI actually recommend?
The distance between the most visible and least visible brands is far wider than most executives would predict. We measured visibility as the organic mention rate, defined as how often a brand surfaced in a blind category query such as “the best telecom provider in South Africa”, with no brand named anywhere in the prompt, averaged across all three models.
MTN sits at the very top of the ranking, appearing in 99% of blind telecom queries, with its closest rival Vodacom reaching 97%. Takealot captured 96% of e-commerce queries, Showmax took 95% of streaming queries, Discovery Health reached 93% of medical-aid queries, Shoprite and Checkers took 90% and 88% of retail queries respectively, and Capitec reached 85% of banking queries. For every one of these brands, AI search now functions as a dependable source of demand, because the model reaches for them first whenever a customer asks an open question.
The other end of the ranking is where the warning sits. Of the 100 brands we measured, a cluster of well-known and entirely real companies returned an organic mention rate of exactly zero. Not low, but zero. Bidvest Bank, a licensed South African bank, appeared in 0% of blind banking queries across all three models. The same zero result held for the connectivity providers Vox and MWeb, and it held for Toyota SA and Volkswagen SA, two of the largest vehicle brands in the country, in blind automotive queries.
Let me be precise about what a 0% organic mention rate means, because the number is easy to soften into something more comfortable. It means that when a South African customer asks an AI model an open question inside that brand’s own category, the model never arrives at that brand on its own. The brand plainly exists. It has customers, branches, staff and a marketing budget. The model has not been given enough reason, by the sources it reads, to count that brand as a candidate answer. Every blind query of that kind represents a customer at the start of a buying decision, and the invisible brand is not in the room while that decision is being shaped.
This is the cleanest argument I can make for why AI search reputation needs a named owner inside the business. A brand can be the third-largest competitor in its category and still be invisible in the channel that is quietly becoming the first stop for purchase research. No internal dashboard reports the absence, and so the existing reports continue to look reassuring.
Why does AI forget some brands entirely?
Because organic visibility and prompted visibility are two genuinely different things, and most brands only ever measure the flattering one.
There is a simple test that separates them. Ask a model a blind question, such as “which fibre providers are good in South Africa”, and record which brands it names without any prompting. Then ask it a named question, such as “is Vox a good fibre provider”, and record how it responds. The distance between those two answers is the single most revealing measurement in AI search reputation. We call it sycophancy uplift, because the models are obliging by design: place a brand’s name inside the question and the model will almost always locate something to say about it.
Vox is the clearest example in the entire dataset. Its blind visibility measured 0%, while its named recall measured 100%. The model never volunteered Vox on its own initiative, yet when it was asked about Vox directly it produced a full and fluent answer on every single attempt. The uplift is the whole 100 points of difference. The same pattern repeated across other brands in our sample, including a major property portal and two well-known e-commerce names.
This distinction matters because of how AI search reputation usually gets checked inside a company. Someone on the marketing team asks ChatGPT about their own brand by name, watches a competent and accurate paragraph come back, and concludes that the brand is performing well in AI. What that person has measured is named recall, which is the flattering number, and they have learned almost nothing about whether a real customer asking an open question would ever be shown the brand at all.
The honest measure of AI visibility is the blind number, because it captures what a brand earns when nobody has mentioned it first. If you take a single testing habit away from this guide, take this one: never assess your AI reputation by typing your own brand name. Ask the open question that a customer would genuinely ask, and observe whether you appear in the answer.
Who controls your brand’s story when AI answers a complaint?
For an uncomfortable number of South African brands, the honest answer is a website located in another country.
Positive and negative questions do not behave the same way inside an AI model. When a customer asks something neutral or favourable, such as what a brand is known for, the model tends to reach for South African sources. When the same customer asks about complaints, the model’s citation pattern shifts offshore, moving toward international complaint platforms such as Trustpilot, Pissed Consumer and Complaints Board. Across the full dataset, the South African citation share drops from 53% overall to 36% on complaints queries — a 17-point swing offshore the moment the question turns critical. That study-wide drop is the average, and the brand-level numbers are considerably sharper than the average suggests.
MTN drew 74% of its citations from South African sources on positive queries and only 29% on negative ones, a 45-point swing offshore that occurred the moment the question turned critical. BYD SA swung 51 points on the same comparison, and Mercedes-Benz SA swung 36. For brands in that position, the favourable story is told at home while the critical story is told abroad, on platforms where a South African communications team has very little standing to respond, correct the record, or add missing context.
There is an aggregate version of this same problem, and it shows up plainly in the study’s most-cited sources. The second most-cited domain across the entire benchmark, behind only BusinessTech, is Trustpilot, an international complaint platform, with 4,942 citations. A foreign review site ranking second, ahead of every South African news publication, is the negative-surface problem stated as a single number. When the models reach for a widely trusted account of a brand, a complaints platform is one of the first places they agree on.
I want to be careful at this point, because this is exactly where AI search reputation tends to get misread. A complaint-site citation is not evidence that a brand is bad. It is evidence that the brand has not given the models a better-sourced and more authoritative account of itself to read instead. The model is not delivering a verdict. It is summarising the best material it can locate. When the strongest available material on a critical question happens to be a complaint thread, the model will summarise that complaint thread, fairly and fluently, for a customer who asked the question in good faith. The remedy is not to argue with the model. The remedy is to make certain that something better exists for it to cite.
Which sources do you need to win in South Africa?
A shorter list than most brands expect. Because AI models read the web through citations, AI search reputation reduces in practice to a question of which publishers describe you, and in South Africa a small handful of publishers carry the overwhelming majority of the weight.
Across all 197,156 citations, one domain stood far above every other. BusinessTech was the single most-cited South African source by a wide margin, drawing close to seventy percent more citations than the next publication behind it. Beneath BusinessTech sat HelloPeter, MyBroadband, TechCentral and a cluster of comparison and rate sites. None of these publications are obscure, and a brand’s PR team almost certainly knows all of them already. What the benchmark data adds is a firm ranking, supplying proof of which mastheads the models genuinely trust when answering a question about a given category.
| Rank | South African source | Citations |
|---|---|---|
| 1 | businesstech.co.za | 7,984 |
| 2 | hellopeter.com | 4,708 |
| 3 | mybroadband.co.za | 2,521 |
| 4 | techcentral.co.za | 2,121 |
| 5 | rateweb.co.za | 1,950 |
| 6 | hippo.co.za | 1,897 |
Two features of that list deserve a second look. The first is HelloPeter occupying second place, because a customer-complaint platform being the second-most-cited South African source is the narrative problem from the previous section made entirely concrete. If a brand is not actively shaping what the editorial publishers say about it, the complaint platform fills the resulting vacuum by default. The second feature worth noticing is how short the list is. A brand does not need to appear everywhere on the internet. It needs to be accurately and recently described across the ten or fifteen publications that the models genuinely read.
One further detail is worth knowing, and it depends entirely on which assistant a brand’s customers prefer. Gemini cites a class of sources that GPT-5 and Claude ignore completely. Reddit is the clearest illustration, because Gemini drew 746 citations from Reddit across the study while GPT-5 and Claude drew exactly zero between them. If a brand’s customers rely on Google’s AI answers, community forums form part of that brand’s reputation surface. If those same customers use ChatGPT or Claude instead, the identical forums are close to irrelevant. The channel is not one channel. It is three of them, and the three do not read the same web.
What should a South African brand do about this?
Begin by measuring the right thing, and then repair narrative before chasing visibility. Here is the sequence I would hand to any South African brand, in priority order.
Measure blind visibility rather than named recall. Select the five open questions that a real customer would ask inside your category, with no brand named in any of them. Run those questions through ChatGPT, Claude and Gemini. Record whether your brand appears, how it is described when it does, and which sources are cited in support. That record becomes your baseline, and you should repeat the exercise every quarter. Resist the temptation to type your own brand name, because doing so measures the wrong layer and will always feel reassuring.
Establish which side of the locality gap you occupy. A brand in short-term insurance, automotive, telecom or banking is operating where AI search already leans on South African sources, so the priority is making certain those sources describe the brand well. A brand in e-commerce, restaurants or real estate faces a model whose instinct is to go global, so the priority becomes giving that model a reason to stay local instead.
Repair narrative before chasing additional visibility. If complaint sites are currently telling a brand’s story on critical questions, then additional visibility only delivers more customers toward that unfavourable story. The negative-query sources have to be dealt with first. For South African brands that means running a genuine response operation on HelloPeter, and it means securing authoritative, well-sourced editorial coverage on the publications the models trust, so that the model has something stronger to cite.
Treat owned content as raw material for machines, not only for people. The models read your website. Pages that state plain facts about the brand clearly, covering products, coverage areas, leadership and history, hand the model something accurate to cite. Vague and decorative marketing copy hands it nothing, which forces it back onto third-party sources. Make the factual pages clean, current and specific.
Verify your structured data and entity records. The models lean on structured signals to establish which entity a brand is. A correct Wikidata entry, accurate schema markup, and a consistent brand name across the web together reduce the chance that the model confuses one brand with a similarly named brand in another market. This work is unglamorous, and it matters more than it looks.
Give the whole problem one named owner, and review it every quarter. AI search reputation currently falls between SEO, PR and corporate communications, which means that inside most companies it belongs to nobody at all. Assign it an owner. Hand that owner the blind-visibility baseline and a standing quarterly review. The brands that win the next five years of AI search will be the ones that began measuring while their competitors were still debating whether the channel mattered.
Where this is going
AI assistants are still early in their development. The models will keep changing, the citation patterns will keep moving, and some of the specific numbers in this guide will look different a year from now. That is precisely why the measurement habit matters far more than any single result. A brand that checks its blind visibility every quarter will watch the channel shift and adjust to it in time. A brand that waits for AI search to settle down will discover, when it finally looks, that the answers about its category were composed years earlier and it was never written into them.
The medical scheme my friend had been paying is a real company with real customers and a real marketing budget. It had done nothing wrong in any traditional sense. It had merely failed to notice that the most important conversation about its category had moved somewhere the company was not watching. That is the entire risk, and it is also the entire opportunity. The channel is still new enough that the brands paying attention today are not catching up to anyone. They are arriving first.
If you want to see where your own brand currently stands, that is what we built Cited Brands to show you. The free Brand Scorecard returns your brand’s AI-visibility position in a few minutes, and the SA AI Citation Index sets that position against your industry and the competitors you are measured next to. We refresh both every quarter. Look yourself up. The number will tell you whether you are defending a lead or starting from behind, and either way, it is far better to know than to guess.
Cited Brands runs the SA-AEO-Bench, a publicly pre-registered study of how frontier AI models cite South African brands. Methodology and full data are available at osf.io/w4az2 . This guide is refreshed every quarter as new data lands.