How AI Decides Which South African Businesses to Recommend

When a South African asks ChatGPT "who should I list my house with in Cape Town?", or asks Perplexity "what's the best business bank for my startup?", they get a shortlist. Three or four names. Your business is on it, or it isn't. There's no page two, no sponsored workaround, no SEO trick that puts you back in the running. What the AI doesn't surface, the buyer never sees.

That shortlist is becoming the front door. And it runs on a different set of signals from the ones your Google rank alone reflects.

So I went and measured them. Not opinions — data. My team and I took real buyer questions across three SA industries — property, banking, crypto — put them to the four assistants South Africans actually use (ChatGPT, Claude, Perplexity, Gemini), and dissected every source each one cited. Then we looked at what the brands that get named have in common, and what the ones that get ignored, often bigger ones, are missing.

Seven patterns came back. They held across all three industries. Here they are.

Pattern 1 — It's not one race. It's four.

The single biggest mistake SA brands make is checking "what ChatGPT says about us" and thinking they've measured their AI visibility. You've measured a quarter of it, at most.

The four assistants behave like four different search engines:

• ChatGPT needs a caveat. We measured it through OpenAI's API (gpt-4o-mini, with web search offered), and that model often chose not to search: in the property study only 57% of its answers cited a source, against 93–100% on the other three, and it named fewer agencies per shortlist (2.9, against 4.2 to 5.9). That's how the API model behaved, not proof of what the ChatGPT app does for you. When it did search, it pointed at the brand's own website.
• Claude and Gemini cast a wide net and lean on third-party sources — directories, comparison sites, guides.
• Perplexity is the social one. Nearly a fifth of what it cites in property is Facebook, YouTube, TikTok and Reddit — far more than any other engine.

Same question, four different answers, four different reasons. If your visibility plan isn't engine-by-engine, it's guessing.

Pattern 2 — Crawlable beats clever. And your robots.txt can mislead you.

An engine can only cite your own pages if its search crawler is allowed to read them: OAI-SearchBot for ChatGPT, Claude-SearchBot for Claude, PerplexityBot for Perplexity. Block it and that engine can't quote your site. It can still name you from what other sites say about you, so a block costs you your own voice in the answer, not necessarily your place on the list.

Here's the part that should worry every brand that's "checked": in the property study, one of the four biggest agencies in the country refused OpenAI's GPTBot at the server while its robots.txt file said the AI bots were welcome. The refusal was one layer deeper, at the web server or firewall, invisible to the standard SEO check everyone runs.

GPTBot happens to be OpenAI's training crawler, so that rule kept the agency out of OpenAI's training data, not out of ChatGPT search (when we checked its search crawler in September and October 2026, it was served normally). But the same kind of server rule catches search crawlers just as easily, and nothing in robots.txt will show it. The only way to catch it is to fetch the site as each crawler and see what comes back. Almost no one does. It's nearly always a stock security setting nobody chose, and it's a one-line fix — but only if you know to look.

Pattern 3 — In a crowded market, get onto the sources AI already trusts.

For "best X" questions in a well-covered industry, the AI doesn't mostly read brand websites — it reads the middlemen. In property, ten third-party sites (Property24, PrivateProperty, BestAgent, a handful of review and comparison sites) carry nearly a third of every citation. In banking, the single most-cited source for "best business bank" isn't a bank at all — it's a government-guide site and a US accounting blog. In crypto, it's tax calculators and comparison sites.

The pattern is iron: the two clearest losers in property had a presence on zero of these sites. Famous locally, invisible on the pages AI actually reads. If you're in a crowded category, being on those middleman sources isn't a marketing nice-to-have — it's the floor you have to clear to be in the conversation at all.

Pattern 4 — In a thin market, your own site is the source.

The opposite is just as true, and it's where smaller brands win. In young or thin SA categories — solar, immigration services, online schooling — there are no trusted middlemen for the AI to lean on. So it falls back on the best thing it can find: a well-built brand website. We watched a solar company's own site become the single most-cited source in its entire category — beating every directory and comparison site, because there weren't any worth citing.

The lesson: the right move depends on how crowded your category is. Crowded → get onto the trusted third parties. Thin → become the definitive source yourself. Most advice ignores this. Knowing which game you're in is half the battle.

Pattern 5 — Be the answer, not the address.

When AI cites a brand's own site, which page it picks tells you everything. The brands that win citations are cited for pages that answer the buyer's question — "what commission do agents charge", "how to choose an agent", "sole vs open mandate". The brands that lose are cited, if at all, for thin contact and office pages — an address, not an answer. One big agency's most-substantial pages were 11,000-word directories of staff names. High word count, zero answer. AI walked past them.

If your site doesn't answer the questions buyers actually ask, you're handing those citations to whoever does.

Pattern 6 — Schema and SEO aren't the secret. (Sorry.)

Here's the one that breaks the GEO-consultant playbook. The single most-recommended estate agency in South Africa has zero structured data — no schema markup at all — on its homepage. The brand with the most schema ranks mid-table. Across the board, the fashionable technical checklist — schema, JSON-LD, the works — did not separate the winners from the losers.

What did? Three things, in order: corroboration (are you on the sources AI trusts), authority (does your brand have a real organic search footprint — this correlated most strongly of any single number), and category fit. On that last one: one of SA's largest financial brands scores a flat zero for "best business bank" — because it's an insurer, not a bank. All the authority in the world won't earn you a recommendation in a category you don't credibly belong to.

Pattern 7 — Famous offline doesn't mean visible to AI.

The hardest truth for established brands. A celebrated Atlantic Seaboard luxury agency barely registers when AI is asked about Atlantic Seaboard luxury. A crypto exchange whose own website AI can barely read still gets recommended second in the country — because the rest of the internet vouches for it. Your reputation in the room counts for nothing if it hasn't been written down on the pages machines read. AI doesn't know who's respected. It knows who's corroborated.

What this means for you

Strip it back and the playbook is short, and honest:

1. Check all four engines, not one. Your ChatGPT result is not your visibility.
2. Fetch your own site as an AI crawler. Not your robots.txt — the actual server response, per bot. This is the cheapest, highest-impact thing on this list, and the one almost no one does.
3. Work out which game you're in. Crowded category → get onto the third-party sources AI cites. Thin category → make your own site the definitive answer.
4. Build pages that answer questions, not pages that list your offices.
5. Stop over-investing in schema and start investing in corroboration and genuine category authority.

Your Google rank is an input to this, not the scoreboard. The scoreboard is whether you are named — and most SA brands haven't looked.

Here's the uncomfortable part, though: everything above is the general shape of the game. It still can't tell you the one thing you actually want to know — which of the seven patterns you're losing on, which engines already leave you off the list, and which sources matter in your category. Those answers are specific to your brand; you can't read them off a framework, including this one. Knowing the patterns is step one. Knowing where you stand — and closing the gap — is the work.

An honest caveat, because we'd rather be trusted than impressive: AI answers shift week to week, this is a point-in-time read, and these are patterns we measured, not laws of physics. We're not promising that ticking these boxes guarantees a recommendation — we're showing you, with evidence, what the brands that do get recommended have in common. The only real proof is moving a brand from invisible to cited and watching it happen. That's the work.