Why South Africa's most famous estate agencies are invisible in AI

A South African buyer used to start with a Google search and a drive past the For Sale boards. Now a lot of them start by asking ChatGPT, Gemini or Perplexity who the best agents are in their suburb. The assistant gives back a shortlist. Three or four names. The buyer reads those, and only those.

So we went and read the shortlists back. We took 432 real property answers from the four assistants South Africans actually use, and traced every source each one cited. We expected the big national names to dominate, the way they dominate billboards and bus stops.

They often don't come up at all.

AIV Index research — How AI Decides Which South African Businesses to Recommend — 432 SA residential-property answers, analysed at the source-URL level.

The reasons are specific, and most of them are fixable. Start with a setting almost no agency knows is switched on, because it decides whether the engines can quote your site at all.

The block that hides one layer below your robots.txt

Each AI engine fetches pages for its answers with its own search crawler: OAI-SearchBot for ChatGPT, Claude-SearchBot for Claude, PerplexityBot for Perplexity. If your server turns that crawler away, the engine can't read or cite your site. It can still name you from what directories and review sites say, so a block costs you your own voice in the answer, not necessarily your place on the list.

In the property study, one of the four biggest agencies in the country was refusing OpenAI's GPTBot at the server, and almost certainly didn't know it. GPTBot is OpenAI's training crawler, not the one ChatGPT search uses, so that rule kept the agency out of OpenAI's training data rather than out of ChatGPT's answers. The same kind of rule catches search crawlers just as easily.

Here is the part that should worry every agency that has already "checked." Their robots.txt file — the public list of which bots are welcome — can say the AI crawlers are welcome while the block happens one layer deeper, at the web server or firewall, where the standard SEO audit never looks. It is almost always a stock security setting nobody chose, and it is usually a one-line fix. But you only fix it if you know to look.

The only way to catch it is to fetch your own site the way each AI crawler does, per bot, and see what comes back. Almost no one does this. You can run that check on your own domain in about 30 seconds, free: check whether AI can read your site.

AIV Index research — How AI Decides Which South African Businesses to Recommend (Pattern 2); the property study's crawler probe, June 2026. Crawler-volume context: Vercel / Merj, December 2024.

The scale of this is not small. On one large hosting network, GPTBot alone made hundreds of millions of requests in a single month — and AI crawlers as a group now run at roughly a fifth of Googlebot's volume.

569MGPTBot requests in a single month, on one hosting networkVercel / Merj, December 2024

Most of that is crawling for model training rather than search, but it is the scale of traffic a single server rule can turn away without anyone noticing.

In a crowded category, AI reads the directories first

For a "best estate agent" question in a well-covered market, the assistants mostly don't read agency websites. They read the middlemen. In the property study, ten third-party sites — Property24, PrivateProperty, a handful of review and comparison sites — carry nearly a third of every citation we counted.

Famous locally, invisible on the pages AI actually reads.

That is the trap for a household name. The two clearest losers in the property study had a presence on zero of those ten sites. Their billboards are everywhere. The pages the assistant reads to build its answer mention them nowhere. So they don't make the shortlist, and no one inside the agency ever sees the question that left them off.

Being listed and reviewed on the directories AI already trusts is not a marketing nice-to-have in a crowded category. It is the floor you have to clear to be in the conversation at all.

AIV Index research — How AI Decides Which South African Businesses to Recommend (Pattern 3).

Reputation in the room counts for nothing the machines can read

The hardest truth for an established brand is that AI has no idea who is respected. A celebrated Atlantic Seaboard luxury agency — a name that means something to every buyer on that coastline — barely registers when the assistant is asked about Atlantic Seaboard luxury property. The reputation is real. It just hasn't been written down on enough of the pages a machine reads.

AI doesn't know who's respected. It knows who's corroborated.

That single line is the whole study in nine words. Corroboration is the currency: how many sources the rest of the internet has produced that independently vouch for you. Decades of standing in a community is worth a great deal in the room and close to nothing in the index, unless the room has been written about somewhere the crawler can reach.

AIV Index research — How AI Decides Which South African Businesses to Recommend (Pattern 7).

The four engines are four different races

One more thing the study made unavoidable: checking ChatGPT and calling it done measures, at most, a quarter of your visibility. The four assistants behave like four different search engines, and the same agency can win one and vanish from another.

• ChatGPT comes with a caveat. We measured it through OpenAI's API, and that model often chose not to search: in the property study 57% of its answers cited a source, against 93–100% on the other engines, and it named fewer agencies per shortlist. That's how the API model behaved, not proof of what the ChatGPT app does for you. When it did search, it tended to point at the brand's own website.
• Claude and Gemini cast a wider net and lean on third-party sources: directories, comparison pages, 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 answers, four different reasons. A visibility plan that isn't engine-by-engine is guessing.

AIV Index research — How AI Decides Which South African Businesses to Recommend (Pattern 1).

What to actually do about it

The pattern is the playbook. There is no schema trick and no clever rewrite that shortcuts it.

1. Make sure the engines can read you. Fetch your own site as an AI crawler, per bot — not your robots.txt, the actual server response. It is the cheapest, highest-impact move on this list and the one almost nobody does. Check it now.
2. Get onto the sources AI already trusts. In a crowded category that means real listings and reviews on the directories and comparison sites the assistants read — being there is the floor, not the ceiling.
3. Earn corroboration, not just reputation. Get written about where machines can reach it. Offline fame doesn't transfer on its own.
4. Check all four engines. Your ChatGPT result is not your visibility.

Your Google rank is an input to this, not the scoreboard. The scoreboard is whether you are named, and most SA agencies haven't looked. The general shape of the game is in the full study. Where your agency stands — which engines already leave you off, which sources matter in your suburb — is specific to you, and you can measure it directly.

A sibling piece goes deeper on the lever everyone reaches for and shouldn't: schema markup won't get you cited by AI — here's what does.

Sources: AIV Index research — How AI Decides Which South African Businesses to Recommend (432 SA residential-property answers, dissected at the source-URL level; read it here). Crawler-volume figures: Vercel / Merj, December 2024. AI answers shift week to week; this is a point-in-time read of patterns we measured, not a guarantee.